Jev Pricing: What It Costs per Million Tokens

Jev pricing is $0.042 per million input tokens with no charge for output tokens.

This rate applies to the scoring model developed by TypeSafe AI.

The model entered early access on September 15, 2026, under founder Diogo Almeida.

It functions as a specialized classification engine rather than a general conversation generator.

At Layer3Labs, we help businesses evaluate and design automated classification systems, where choosing between dedicated decision models and general reasoning models determines ongoing infrastructure spend.

When evaluated alongside conversational alternatives, its pricing structure shifts how development teams estimate high-volume classification budgets.

The Published Rate Card and Free Output

TypeSafe AI charges a flat fee of $0.042 per million input tokens for Jev, while generated output carries zero per-token cost.

Output is free because Jev does not generate conversational prose or text tokens.

The model returns typed decisions with calibrated numerical probabilities.

Queries use one of three question types: Choice, Score, or Noul.

Choice selects from up to 255 items.

Score places inputs on defined rubrics.

Noul returns a single probability float between zero and one for yes or no questions.

TypeSafe AI sets a request budget of roughly 32,000 tokens, which equals about 150,000 English characters.

Early access requires joining a waitlist.

TypeSafe AI documents no free tier and provides no complimentary trial credits.

• Rate: $0.042 per million input tokens.
• Output cost: Free for all responses.
• Context limit: Roughly 32,000 tokens per request (about 150,000 characters).
• Question formats: Choice (up to 255 options), Score (ordered rubric tiers), and Noul (yes or no probability).
• Account access: Early-access waitlist with no documented free tier or free credits.

Jev Pricing and Volume Calculations for Real Workloads

At $0.042 per million input tokens, Jev pricing scales predictably from a few cents on small queues to modest dollar amounts on enterprise feeds.

A batch of 10,000 customer tickets averaging 500 input tokens each consumes 5 million tokens.

That run costs exactly $0.21.

Sizing up to 100,000 requests of 1,000 tokens each yields 100 million input tokens.

That enterprise queue totals $4.20.

At one million requests of 1,000 tokens each, the bill reaches $42.00.

Community implementations illustrate this arithmetic in practice.

The open-source audit tool JevSEO scores pages across Search Engine Optimization (SEO), Answer Engine Optimization (AEO), and Generative Engine Optimization (GEO).

The project repository documents an average crawl cost of about $0.0001 per page across 3,969 tokens, completing a ten-page review for roughly $0.002.

• 10,000 requests (500 tokens each): 5 million tokens cost $0.21.
• 50,000 requests (1,000 tokens each): 50 million tokens cost $2.10.
• 100,000 requests (1,000 tokens each): 100 million tokens cost $4.20.
• 1,000,000 requests (1,000 tokens each): 1 billion tokens cost $42.00.
• JevSEO audit crawl: The maintainer reports roughly $0.0001 per page for 3,969 tokens.

Comparing Jev API Costs to General Language Models

Jev lowers raw API spend on classification tasks by removing output token charges that inflate standard language model bills.

Standard Large Language Models (LLMs) charge separately for incoming context and generated answers.

Even concise JSON replies require paying for output tokens.

Consider an identical classification queue of 100,000 requests with 1,000 input tokens each.

On Jev, 100 million input tokens cost $4.20 with zero output charge.

Running that workload on OpenAI GPT-5.6 Luna costs $0.20 per million input tokens and $1.20 per million output tokens.

Assuming 50 output tokens per classification decision, GPT-5.6 Luna incurs $20.00 in input charges and $6.00 in output fees for a $26.00 total.

Jev lowers that bill by 84%.

On Anthropic Claude Haiku 4.5, input costs $1.00 and output costs $5.00 per million tokens, generating a $125.00 bill.

Jev reduces that cost by 97%.

• Jev total cost: $4.20 ($0.042 input rate, $0 output rate).
• GPT-5.6 Luna total cost: $26.00 ($20.00 input plus $6.00 for 5M output tokens).
• Claude Haiku 4.5 total cost: $125.00 ($100.00 input plus $25.00 for 5M output tokens).
• Luna cost reduction: Jev cuts API fees by 84% on this 50-token output benchmark.
• Haiku 4.5 cost reduction: Jev cuts API fees by 97% on this 50-token output benchmark.

Vendor Performance Claims and Accuracy Limits

TypeSafe AI claims that Jev processes queries with 70 to 500 millisecond latency, but internal benchmarks indicate lower raw accuracy than general models.

Diogo Almeida states that Jev operates 20 to 200 times faster and 40 to 400 times cheaper than conversational platforms.

These metrics reflect vendor-reported figures rather than independent measurements.

TypeSafe AI published several creator demonstrations to illustrate these claims.

These examples include analyzing 724 advertisements in 40 seconds for $0.09 and triaging 384 headlines in 24.9 seconds for $0.19.

Demonstrations also highlighted a browser-automation flight search completed in 7 seconds for $0.0039 and a trading bot executing decisions every 300 milliseconds.

Accuracy data tells a more cautious story.

On TypeSafe AI's internal workflow dashboard, Jev aggregates at 67.8% accuracy compared to 74.1% for an unidentified best comparator.

Task scores range from 61.7% to 76.0%.

The model trades some categorical accuracy for speed and lower token expense.

Reinforcement Learning for Calibrated Decisions (RLCD) trains the model to align probabilities with real outcomes.

Calibration is a training objective rather than an accuracy guarantee.

A confidence score indicates distribution shape, not verified correctness on individual inputs.

No independent or third-party benchmark exists as of late September 2026.

Early enterprise interest grew quickly, with Forbes reporting on September 19, 2026, that Vercel and Cloudflare moved rapidly to add Jev integrations.

• Latency: TypeSafe AI reports 70 to 500 millisecond end-to-end response times.
• Vendor efficiency claim: Diogo Almeida claims 20 to 200 times faster execution and 40 to 400 times higher cost efficiency.
• Vendor triage demos: 724 ads analyzed for $0.09; 384 headlines triaged for $0.19; flight search for $0.0039.
• Internal accuracy: TypeSafe AI documents 67.8% average accuracy versus 74.1% for an unidentified comparator.
• Calibration objective: RLCD targets probability calibration, but does not guarantee accuracy on specific tasks.
• Third-party benchmarks: No independent evaluation data exists as of September 2026.

Integration Expenses and Capability Constraints

The financial advantages of Jev narrow when development teams must write custom code to handle operations the model cannot execute.

Jev cannot generate text, summarize documents, or write code.

It cannot perform arithmetic, count records, evaluate dates, or follow indirect chains of logic.

TypeSafe AI advises developers to keep exact calculations in program code and present questions directly.

Inputs are text only, with no multimodal support for image or audio analysis.

Responses do not stream, and Choice questions accept a maximum of 255 discrete options.

The Application Programming Interface (API) uses an OpenAI-compatible endpoint with the model alias jev-latest.

Developers can also call the model through Requesty using typesafe/jev-1.13.0 or typesafe/jev-latest. The integration is flagged experimental, so check TypeSafe AI's own docs before building on it.

That Requesty integration is flagged as experimental and may alter specifications without advance deprecation notice.

Engineering teams must budget development hours to construct deterministic validation wrappers around these API calls.

• Generative restriction: No text generation, explanation writing, or code production.
• Logic boundaries: Cannot compute arithmetic, count items, compare calendar dates, or resolve indirect reasoning.
• Input constraints: Text-only ingestion up to 32,000 tokens, with no audio or image support.
• Operational limits: No streaming output; Choice questions cap at 255 options.
• Integration status: Primary API uses OpenAI compatibility, while Requesty endpoints remain experimental.

Workload Fit and Cost Decision Factors

Jev delivers optimal cost returns for high-throughput routing pipelines where deterministic decisions matter more than explanatory text.

Ideal workloads include content tagging, spam moderation, customer intent routing, and initial lead scoring.

These tasks run millions of times per month and require structured labels rather than generative reasoning.

Jev is not suitable for organizations requiring customer-facing chatbot responses, document synthesis, or complex problem-solving.

Teams with those requirements should deploy GPT-5.6 Luna or Claude Haiku 4.5.

Those models charge for output tokens but supply the required explanatory text.

Several operational developments would alter our financial evaluation of this architecture.

A shift in TypeSafe AI's billing model to introduce per-request fees or output surcharges would compress the current price gap.

Similarly, if competing frontier labs lower structured classification rates below $0.05 per million tokens, the necessity for a dedicated decision engine declines.

Before migrating existing classification tasks, calculate your total input-token spend against potential refactoring costs, and verify current rates on the TypeSafe AI website.

• Recommended use cases: High-volume triage, automated tagging, intent routing, and binary content filtering.
• Workloads to avoid: Document drafting, conversational agents, mathematical analysis, and multi-step reasoning.
• Trigger conditions: Per-request API charges, added output fees, or rival structured-output rate reductions.
• Evaluation rule: Weigh the 84% to 97% token savings against the engineering time required to build deterministic application wrappers.