What are the AI Business Benchmarks?

The AI Business Benchmarks score AI models and tools on 63 real business jobs, such as writing a newsletter, deciding a refund, typing up invoices or abstracting a lease. Each job is scored on one question: could you send the result without fixing it? The benchmarks are published by Layer3Labs.

How are these different from SWE-bench, MMLU or LMArena?

Those benchmarks measure coding puzzles, exam questions and which chat answer people prefer. None of them measure whether AI can do ordinary business work to a standard you would send to a customer. The AI Business Benchmarks test the jobs a business actually hands off, graded the way a business owner judges work.

What does "Done" mean?

The Done rate is how many times out of 100 the work comes back needing no edits at all. Every job on this page states exactly what Done means for that job, so the standard is the same for every AI tested.

What does "Burned you" mean?

The Burned-you rate is how often the work comes back wrong in a way you would not catch before it went out: an invented figure, a policy your business does not have, a case or citation that does not exist. It is the number that decides whether you can hand a job off without checking every result.

How is cost calculated?

Cost comes from each vendor's published token prices, tracked on our AI model pricing page, applied to the size of each job. Cost per finished piece of work also counts the attempts you would throw away, so a cheap model that fails often does not look cheaper than it is.

Which AI models and tools are tested?

12 AI models, including Claude, ChatGPT, Gemini, Grok, DeepSeek, Kimi, Qwen, GLM, Mistral, MiniMax, Llama and Amazon Nova, plus 10 specialist tools such as ElevenLabs, Veo, Sora, HeyGen, Harvey and Abridge. Specialist tools are tested only on the jobs they are built for.

How often is each job run?

Every job is run five times on every AI. That shows whether a result is repeatable, which is what decides whether you can stop checking the work, not whether it got lucky once.

Can an AI lab submit a model?

Yes. AI labs can submit a model for scoring ahead of launch and receive an official score they can cite on release day. Email partners@layer3labs.io with the model name and launch window.

Is the clinic benchmark giving medical advice?

No. The clinic jobs cover paperwork only: visit notes, coding, prior authorizations, credentialing and appeals. No job asks an AI to diagnose, triage or recommend treatment.

Reviewed by Jonathan West · Updated Oct 3, 2026

AI Business Benchmarks: Which AI Actually Does the Job

The benchmark built for businesses, not researchers. We hand real jobs to every major AI and score one thing: could you send the result as-is?

Reviewed by Jonathan West · Updated Oct 3, 2026

Find a job

Task

Marketing

4 jobs

AI Business Benchmarks — Marketing: the best AI for every job, how often the result was Done, what it cost, how often it Burned you, and whether to double-check it
Job	Use	Done	Cost	Burned you	DoubleCheck It?
Make my writing not sound like AIA stiff draft that reads like a robot wrote it	GrokGrok 4.5 · xAI	82/100	~$0.06	3/100	No
Done means Reads like a person wrote it, and every fact and number in the original survives unchanged Burned you looks like A changed figure, a dropped qualifier, or a claim the original never made Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Undetectable.aihumanizerVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write my blog postA topic, an outline and your sources	ChatGPTGPT-5.6 Sol · OpenAI	97/100	~$0.05	0/100	Skim
Done means On the brief, in your voice, every claim traceable to the sources you gave Burned you looks like A statistic, quote or study that is not in your sources Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Undetectable.aihumanizerVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write my email newsletterThis month's news and links	ClaudeFable 5.1 · Anthropic	88/100	~$0.21	2/100	Skim
Done means Subject line, body and one clear ask, ready to schedule Burned you looks like A date, offer or link that was not in your notes Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Undetectable.aihumanizerVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Turn one post into social postsOne blog post and the platforms you post on	ClaudeFable 5.1 · Anthropic	87/100	~$0.18	3/100	No
Done means Written for each platform, no copy repeated, nothing added Burned you looks like A number or result that the original post never claimed Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Undetectable.aihumanizerVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

Ads & creative

4 jobs

AI Business Benchmarks — Ads & creative: the best AI for every job, how often the result was Done, what it cost, how often it Burned you, and whether to double-check it
Job	Use	Done	Cost	Burned you	DoubleCheck It?
Make an ad videoA script and a photo of the product	DeepSeekV4 Pro	85/100	~$0.04	3/100	No
Done means Watchable, on-brand, the product looks like your product, on-screen text is legible Burned you looks like A product that looks right at a glance but is the wrong shape, color or label Who is being tested ElevenLabsvoiceVendor pricing VeoGoogle videoVendor pricing SoraOpenAI videoVendor pricing MidjourneyimagesVendor pricing HeyGenpresenter videoVendor pricing
Put my product in a new photoA real photo of your product	ElevenLabsv3 · voice	97/100	~$0.10	0/100	No
Done means The same product in a new setting, not a lookalike Burned you looks like A lookalike product, wrong logo, or wrong details on the packaging Who is being tested ElevenLabsvoiceVendor pricing VeoGoogle videoVendor pricing SoraOpenAI videoVendor pricing MidjourneyimagesVendor pricing HeyGenpresenter videoVendor pricing
Record a voiceoverA script and the voice you want	ClaudeFable 5.1 · Anthropic	95/100	~$0.32	1/100	No
Done means Natural, and your brand and product names are pronounced right Burned you looks like A mispronounced name or a skipped line nobody notices until it airs Who is being tested ElevenLabsvoiceVendor pricing VeoGoogle videoVendor pricing SoraOpenAI videoVendor pricing MidjourneyimagesVendor pricing HeyGenpresenter videoVendor pricing
Make a talking-head videoA script and a presenter	ElevenLabsv3 · voice	93/100	~$0.21	2/100	No
Done means Lip-sync holds, same face in every take, nothing uncanny Burned you looks like Words the presenter says that are not in your script Who is being tested ElevenLabsvoiceVendor pricing VeoGoogle videoVendor pricing SoraOpenAI videoVendor pricing MidjourneyimagesVendor pricing HeyGenpresenter videoVendor pricing

Sales

4 jobs

AI Business Benchmarks — Sales: the best AI for every job, how often the result was Done, what it cost, how often it Burned you, and whether to double-check it
Job	Use	Done	Cost	Burned you	DoubleCheck It?
Write a cold email sequenceYour offer and the people you are writing to	ClaudeFable 5.1 · Anthropic	95/100	~$0.25	1/100	No
Done means Five emails you would send, no fake personalization Burned you looks like A detail about the prospect that is not true Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.
Look up a prospect before a callA company name	Gemini3.6 Flash · Google	94/100	~$0.02	1/100	Skim
Done means What they do, who runs it and what is new, every fact sourced Burned you looks like An invented funding round, headcount or executive Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.
Write a proposalThe client's brief and your pricing	ClaudeFable 5.1 · Anthropic	80/100	~$0.16	3/100	No
Done means Every requirement answered, your pricing exactly as given Burned you looks like A price, deliverable or guarantee you never offered Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Cost per try is estimated from each vendor's published token price for a job this size.
Follow up on a quoteA quote that has gone quiet for ten days	Gemini3.6 Flash · Google	94/100	~$0.03	1/100	Skim
Done means A nudge you would send, right tone, right ask Burned you looks like A discount or deadline that does not exist Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google<$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek<$0.01 per try GLMGLM 5.2 · Zhipu AI<$0.01 per try GrokGrok 4.5 · xAI<$0.01 per try NovaNova Premier · Amazon$0.01 per try KimiKimi K3 · Moonshot AI$0.02 per try ChatGPTGPT-5.6 Sol · OpenAI$0.02 per try ClaudeClaude Fable 5.1 · Anthropic$0.06 per try Cost per try is estimated from each vendor's published token price for a job this size.

Customer service

4 jobs

AI Business Benchmarks — Customer service: the best AI for every job, how often the result was Done, what it cost, how often it Burned you, and whether to double-check it
Job	Use	Done	Cost	Burned you	DoubleCheck It?
Answer a customer emailThe email and your help docs	ChatGPTGPT-5.6 Sol · OpenAI	97/100	~$0.11	0/100	Skim
Done means The correct answer, in your tone, ready to send Burned you looks like A promise, policy or feature your business does not have Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google<$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek<$0.01 per try GLMGLM 5.2 · Zhipu AI<$0.01 per try GrokGrok 4.5 · xAI<$0.01 per try NovaNova Premier · Amazon$0.01 per try KimiKimi K3 · Moonshot AI$0.02 per try ChatGPTGPT-5.6 Sol · OpenAI$0.02 per try ClaudeClaude Fable 5.1 · Anthropic$0.06 per try ElevenLabsvoiceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Decide whether to give a refundThe complaint and your refund policy	ElevenLabsv3 · voice	97/100	~$0.10	1/100	Skim
Done means The right call under your policy, and it escalates when it should Burned you looks like A refund your policy does not allow, explained with confidence Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google<$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek<$0.01 per try GLMGLM 5.2 · Zhipu AI<$0.01 per try GrokGrok 4.5 · xAI<$0.01 per try NovaNova Premier · Amazon$0.01 per try KimiKimi K3 · Moonshot AI$0.02 per try ChatGPTGPT-5.6 Sol · OpenAI$0.02 per try ClaudeClaude Fable 5.1 · Anthropic$0.06 per try ElevenLabsvoiceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write a help articleFive tickets you've already answered	ClaudeFable 5.1 · Anthropic	95/100	~$0.17	1/100	Skim
Done means An article that stops the sixth ticket Burned you looks like A step or setting that does not exist in your product Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try ElevenLabsvoiceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Answer the phoneYour script and a caller who interrupts	ClaudeFable 5.1 · Anthropic	97/100	~$0.12	1/100	Skim
Done means Handles the interruption, stays on script, gets the outcome Burned you looks like A booking, price or promise made up on the call Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try ElevenLabsvoiceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

Bookkeeping & admin

4 jobs

AI Business Benchmarks — Bookkeeping & admin: the best AI for every job, how often the result was Done, what it cost, how often it Burned you, and whether to double-check it
Job	Use	Done	Cost	Burned you	DoubleCheck It?
Fix a broken spreadsheetA workbook with broken formulas	ChatGPTGPT-5.6 Sol · OpenAI	96/100	~$0.09	1/100	Skim
Done means Formulas repaired, totals tie out, nothing else touched Burned you looks like A total that looks right and is not Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.
Type up a pile of invoicesTwenty messy supplier invoices	ChatGPTGPT-5.6 Sol · OpenAI	97/100	~$0.12	0/100	No
Done means Every field and total right, unreadable ones flagged rather than guessed Burned you looks like A guessed amount or invoice number entered as if it were read Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Match the books to the bankYour books and the bank export	Gemini3.6 Flash · Google	83/100	~$0.02	2/100	Skim
Done means Every difference found and explained Burned you looks like A difference smoothed over instead of explained Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Build the monthly reportRaw numbers and last month's report	Gemini3.6 Flash · Google	92/100	~$0.06	1/100	Skim
Done means Same format, right numbers, changes explained Burned you looks like A number or trend that is not in the data Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Cost per try is estimated from each vendor's published token price for a job this size.

Hiring & staff

4 jobs

AI Business Benchmarks — Hiring & staff: the best AI for every job, how often the result was Done, what it cost, how often it Burned you, and whether to double-check it
Job	Use	Done	Cost	Burned you	DoubleCheck It?
Write a job adThe role and the pay range	ClaudeFable 5.1 · Anthropic	85/100	~$0.21	3/100	No
Done means Ready to post, legal, nothing made up Burned you looks like A benefit or perk you do not offer Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.
Sort through a pile of resumesForty resumes and your criteria	Gemini3.6 Flash · Google	93/100	~$0.04	1/100	Skim
Done means A shortlist you agree with, and the same answer for two identical candidates with different names Burned you looks like A different ranking for identical resumes Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Write a performance reviewA year of your notes	ChatGPTGPT-5.6 Sol · OpenAI	91/100	~$0.09	2/100	Skim
Done means Fair, specific, tied to the evidence Burned you looks like An incident or achievement that is not in your notes Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Cost per try is estimated from each vendor's published token price for a job this size.
Plan a new hire's first monthThe role and the tools you use	ChatGPTGPT-5.6 Sol · OpenAI	97/100	~$0.03	0/100	No
Done means A day-by-day plan they can follow alone Burned you looks like A tool, team or process your business does not have Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.

Scores updated Sep 30, 2026 · 24 jobs scored · 18 AIs tested · prices from the Layer3Labs model price ledger

Industry

Law firm

5 jobs

AI Business Benchmarks — Law firm: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Review a contract against our playbookThe contract and your negotiation playbook	12 AIs + Harvey	$0.01 – $0.58
Done means Flags everything the playbook says to flag, misses nothing material Burned you looks like A clause reported as present or absent when it is not Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Harveylegal assistantVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Abstract a leaseA commercial lease	12 AIs + Harvey	$0.01 – $0.58
Done means Every date, term and option pulled out correctly Burned you looks like A wrong renewal date or rent figure Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Harveylegal assistantVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Find the red flags in a data roomA data room	12 AIs + Harvey	$0.01 – $0.58
Done means Real issues, ranked, each pointing to the document it came from Burned you looks like A citation to a document or case that does not exist Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Harveylegal assistantVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Run a title chainThe deeds and recorded documents	12 AIs + Harvey	$0.01 – $0.58
Done means Every gap and encumbrance listed Burned you looks like A chain reported clean when it is not Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Harveylegal assistantVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Redline a contractThe contract and your position	12 AIs + Harvey	$0.01 – $0.58
Done means Changes a partner would actually send, with the reasoning Burned you looks like A redline that quietly changes a term you did not ask about Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Harveylegal assistantVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

Accounting & tax

5 jobs

AI Business Benchmarks — Accounting & tax: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Reconcile two ledgersTwo ledgers that do not agree	12 AIs	$0.01 – $0.58
Done means Every difference found and explained Burned you looks like A plug entry presented as a match Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Tie out the financial statementsThe statements and the trial balance	12 AIs	$0.01 – $0.58
Done means Every number traces to its source Burned you looks like A figure marked as tied out that does not trace Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Research a tax positionThe facts and the question	12 AIs	<$0.01 – $0.37
Done means The right answer, with real authorities cited properly Burned you looks like A cited ruling, code section or case that does not exist Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Cost per try is estimated from each vendor's published token price for a job this size.
Run payrollHours, pay rates and the state	12 AIs	<$0.01 – $0.16
Done means Right state, right rates, right withholding Burned you looks like A withholding rate that is out of date or wrong for the state Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.
Find unusual expensesA year of expenses and your threshold	12 AIs	$0.01 – $0.58
Done means Everything above your threshold, nothing below it Burned you looks like An expense flagged or cleared on a number it misread Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.

Consulting

5 jobs

AI Business Benchmarks — Consulting: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Size a marketThe market and your sources	12 AIs + Hebbia	<$0.01 – $0.37
Done means Defensible, sourced, and it shows its working Burned you looks like A market figure with no source behind it Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write up a competitorA competitor name	12 AIs + Hebbia	<$0.01 – $0.37
Done means Current facts only, every one sourced Burned you looks like An invented revenue figure or customer Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Recommend a priceYour costs, market and goals	12 AIs + Hebbia	<$0.01 – $0.16
Done means A recommendation, with what changes if you are wrong Burned you looks like A benchmark price that is not in your data Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write the board reportRaw numbers and your format	12 AIs + Hebbia	<$0.01 – $0.37
Done means Right numbers, your format, ready for the board Burned you looks like A trend or figure that is not in the numbers Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Build the pitch storyThe data room	12 AIs + Hebbia	$0.01 – $0.58
Done means Nothing in the deck that isn't in the data Burned you looks like A traction claim that is not in the data room Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

Compliance & risk

5 jobs

AI Business Benchmarks — Compliance & risk: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Run a gap assessmentYour controls and the framework: GDPR, HIPAA or SOC 2	12 AIs + Vanta	$0.01 – $0.58
Done means Every real gap found against the actual control set Burned you looks like A control marked as met that is not Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try VantacomplianceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write a policy to a frameworkThe framework and how you work	12 AIs + Vanta	<$0.01 – $0.37
Done means Maps to the actual clauses Burned you looks like A clause number or requirement that does not exist Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try VantacomplianceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Work out what a rule change means for usThe rule change and your obligations	12 AIs + Vanta	<$0.01 – $0.37
Done means Specific to what you actually have to do Burned you looks like A deadline or obligation the rule does not contain Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try VantacomplianceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Screen a customer for AML/KYCCustomer details and your screening rules	12 AIs + Vanta	<$0.01 – $0.16
Done means Real matches caught, good customers not flagged Burned you looks like A clean customer cleared on a misread record Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try VantacomplianceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Assemble the ESG reportThe underlying data	12 AIs + Vanta	$0.01 – $0.58
Done means Built from the data, gaps reported as gaps Burned you looks like A missing figure filled in with an estimate Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try VantacomplianceVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

Clinic (paperwork only)

5 jobs

AI Business Benchmarks — Clinic (paperwork only): every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Write the visit noteThe visit transcript	12 AIs + Abridge	<$0.01 – $0.16
Done means Complete, accurate, nothing added Burned you looks like A symptom, dose or history item the patient never mentioned Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Abridgeclinical scribeVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Code the visitThe visit note	12 AIs + Abridge	<$0.01 – $0.16
Done means Correct ICD and CPT codes, at the right specificity Burned you looks like A code that does not match what happened Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Abridgeclinical scribeVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Put together a prior-auth packetThe chart and the payer's criteria	12 AIs + Abridge	$0.01 – $0.58
Done means Every criterion addressed from the chart Burned you looks like A criterion marked as met with no support in the chart Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Abridgeclinical scribeVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Check a credentialing fileA provider's credentialing file	12 AIs + Abridge	$0.01 – $0.58
Done means Every gap and expiration caught Burned you looks like An expired license reported as current Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Abridgeclinical scribeVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Write a denial appealThe denial and the chart	12 AIs + Abridge	<$0.01 – $0.37
Done means Cites the criteria that were actually met Burned you looks like A policy citation or chart detail that does not exist Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Abridgeclinical scribeVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

Insurance agency

4 jobs

AI Business Benchmarks — Insurance agency: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Assess a submissionThe submission and your guidelines	12 AIs	$0.01 – $0.58
Done means Right call against your appetite and guidelines Burned you looks like A risk detail misread from the submission Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Summarize a claim fileThe claim file	12 AIs	$0.01 – $0.58
Done means Facts straight, coverage position clear Burned you looks like A date, amount or party that is not in the file Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Compare two policiesTwo policy documents	12 AIs	$0.01 – $0.58
Done means The real differences, not a feature list Burned you looks like A coverage or exclusion one policy does not have Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Draft a coverage letterThe policy and the claim	12 AIs	<$0.01 – $0.37
Done means Accurate on what is and isn't covered Burned you looks like Coverage promised that the policy does not give Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Cost per try is estimated from each vendor's published token price for a job this size.

Construction

5 jobs

AI Business Benchmarks — Construction: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Take off quantities from drawingsThe drawings	12 AIs	$0.01 – $0.58
Done means Counts and measures correct Burned you looks like A quantity that does not match the drawings Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Check a submittal against the specThe submittal and the spec	12 AIs	$0.01 – $0.58
Done means Every deviation caught Burned you looks like A deviation reported as compliant Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.
Put together a permit packageThe project and the jurisdiction	12 AIs	<$0.01 – $0.37
Done means Complete for that jurisdiction Burned you looks like A form or requirement the jurisdiction does not have Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Cost per try is estimated from each vendor's published token price for a job this size.
Write a change orderThe change and your rates	12 AIs	<$0.01 – $0.16
Done means Priced from your rates, scope described accurately Burned you looks like A rate or line item you did not give it Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral<$0.01 per try QwenQwen3.6 Plus · Alibaba<$0.01 per try GeminiGemini 3.6 Flash · Google$0.01 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.02 per try GLMGLM 5.2 · Zhipu AI$0.02 per try GrokGrok 4.5 · xAI$0.02 per try NovaNova Premier · Amazon$0.04 per try KimiKimi K3 · Moonshot AI$0.05 per try ChatGPTGPT-5.6 Sol · OpenAI$0.06 per try ClaudeClaude Fable 5.1 · Anthropic$0.16 per try Cost per try is estimated from each vendor's published token price for a job this size.
Check plans against the safety codeThe plans and the code	12 AIs	$0.01 – $0.58
Done means Real violations, each cited Burned you looks like A code section that does not exist Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Cost per try is estimated from each vendor's published token price for a job this size.

Bids & procurement

5 jobs

AI Business Benchmarks — Bids & procurement: every job, which AIs are tested, and the estimated cost of one attempt from published token prices
Job	Tested on	Cost per try
Answer an RFPThe RFP and your capability library	12 AIs + Hebbia	$0.01 – $0.58
Done means Every requirement addressed from your library Burned you looks like A capability or certification you do not have Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Build a compliance matrixThe RFP requirements	12 AIs + Hebbia	$0.01 – $0.58
Done means Requirement by requirement, honest about gaps Burned you looks like A requirement marked as met that is not Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Score the vendorsVendor responses and your weighting	12 AIs + Hebbia	$0.01 – $0.58
Done means Your weighting applied, arithmetic correct Burned you looks like A score that does not add up Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Review a procurement contractThe contract and your policy	12 AIs + Hebbia	$0.01 – $0.58
Done means Flags what your policy says to flag Burned you looks like A term reported that is not in the contract Who is being tested LlamaLlama 4 Maverick · Meta$0.01 per try MiniMaxMiniMax M3 · MiniMax$0.02 per try MistralMistral Large 3 · Mistral$0.03 per try QwenQwen3.6 Plus · Alibaba$0.03 per try GeminiGemini 3.6 Flash · Google$0.04 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.07 per try GLMGLM 5.2 · Zhipu AI$0.07 per try GrokGrok 4.5 · xAI$0.10 per try NovaNova Premier · Amazon$0.14 per try KimiKimi K3 · Moonshot AI$0.17 per try ChatGPTGPT-5.6 Sol · OpenAI$0.23 per try ClaudeClaude Fable 5.1 · Anthropic$0.58 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.
Decide whether to bidThe opportunity and your criteria	12 AIs + Hebbia	<$0.01 – $0.37
Done means Reasoning you'd defend internally Burned you looks like A requirement or fact about the bid that is not in it Who is being tested LlamaLlama 4 Maverick · Meta<$0.01 per try MiniMaxMiniMax M3 · MiniMax<$0.01 per try MistralMistral Large 3 · Mistral$0.01 per try QwenQwen3.6 Plus · Alibaba$0.02 per try GeminiGemini 3.6 Flash · Google$0.03 per try DeepSeekDeepSeek V4 Pro · DeepSeek$0.04 per try GLMGLM 5.2 · Zhipu AI$0.04 per try GrokGrok 4.5 · xAI$0.05 per try NovaNova Premier · Amazon$0.09 per try KimiKimi K3 · Moonshot AI$0.11 per try ChatGPTGPT-5.6 Sol · OpenAI$0.15 per try ClaudeClaude Fable 5.1 · Anthropic$0.37 per try Hebbiadocument analysisVendor pricing Cost per try is estimated from each vendor's published token price for a job this size.

The Layer3Labs AI Business Benchmarks test 12 AI models and 10 specialist tools on 63 business jobs: 24 everyday jobs every business has, and 39 specialist jobs from law firms, accounting practices, clinics, insurance agencies, construction firms and compliance teams. No existing AI benchmark measures this work.

How Every Job Is Graded

Every result lands on one of four grades, whether the job is a social post or a prior-authorization packet. It is the same test a business owner applies to work that comes back to their desk.

The four grades used in the AI Business Benchmarks
Grade	What it means
Done	Send it. No edits.
Fix	Usable after five or ten minutes of edits.
Redo	Wrong. Do it yourself.
Burned you	Wrong in a way you would not have caught before it went out.

What We Measure

Done rate: how many times out of 100 the work comes back needing no edits.
Burned-you rate: how often the work comes back wrong in a way you would never catch: an invented figure, a policy you don't have, a case that doesn't exist. Every other benchmark scores what AI gets right. This one also scores what it gets wrong without telling you.
Cost per finished job: what one usable piece of work costs at published prices, counting the attempts you would throw away.
DoubleCheck It? Yes, skim or no. Each job runs five times in a row; if the AI doesn't slip and doesn't burn you, you can stop checking.

For AI labs: submit a model for scoring before launch and publish an official AI Business Benchmarks score on release day. Email partners@layer3labs.io with the model name and launch window.

How We Test

Every job comes with the inputs a real business would hand over: the messy invoices, the refund policy, the lease, the data room. Each AI gets the same inputs and the same instructions, five times. Jobs with a single right answer, like a spreadsheet total or a refund decision under a written policy, are graded mechanically. Every result is checked against the facts in its inputs, so anything invented is caught.

Prices come from each vendor's published rates, tracked on our AI model pricing page. When a price changes there, the cost on every job here changes with it.

Frequently Asked Questions

The AI Business Benchmarks score AI models and tools on 63 real business jobs, such as writing a newsletter, deciding a refund, typing up invoices or abstracting a lease. Each job is scored on one question: could you send the result without fixing it? The benchmarks are published by Layer3Labs.
Those benchmarks measure coding puzzles, exam questions and which chat answer people prefer. None of them measure whether AI can do ordinary business work to a standard you would send to a customer. The AI Business Benchmarks test the jobs a business actually hands off, graded the way a business owner judges work.
The Done rate is how many times out of 100 the work comes back needing no edits at all. Every job on this page states exactly what Done means for that job, so the standard is the same for every AI tested.
The Burned-you rate is how often the work comes back wrong in a way you would not catch before it went out: an invented figure, a policy your business does not have, a case or citation that does not exist. It is the number that decides whether you can hand a job off without checking every result.
Cost comes from each vendor's published token prices, tracked on our AI model pricing page, applied to the size of each job. Cost per finished piece of work also counts the attempts you would throw away, so a cheap model that fails often does not look cheaper than it is.
12 AI models, including Claude, ChatGPT, Gemini, Grok, DeepSeek, Kimi, Qwen, GLM, Mistral, MiniMax, Llama and Amazon Nova, plus 10 specialist tools such as ElevenLabs, Veo, Sora, HeyGen, Harvey and Abridge. Specialist tools are tested only on the jobs they are built for.
Every job is run five times on every AI. That shows whether a result is repeatable, which is what decides whether you can stop checking the work, not whether it got lucky once.
Yes. AI labs can submit a model for scoring ahead of launch and receive an official score they can cite on release day. Email partners@layer3labs.io with the model name and launch window.
No. The clinic jobs cover paperwork only: visit notes, coding, prior authorizations, credentialing and appeals. No job asks an AI to diagnose, triage or recommend treatment.

Want to Know Which Jobs AI Can Take Off Your Plate?

We map the jobs in your business that AI can do to a send-it-as-is standard, and the ones it can't yet. Start with a free AI workflow audit.

Book a Consultation

Related Resources

Resource

AI Business Benchmarks: Which AI Actually Does the Job

Marketing

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Ads & creative

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Sales

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Customer service

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Bookkeeping & admin

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Hiring & staff

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Law firm

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Accounting & tax

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Consulting

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Compliance & risk

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Clinic (paperwork only)

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Insurance agency

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Construction

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Bids & procurement

Who is being tested

Who is being tested

Who is being tested

Who is being tested

Who is being tested

How Every Job Is Graded

What We Measure