Claude Haiku 5.5 vs GPT-6 Luna: Price per Token and Cost per Task

As of October 2026, Claude Haiku 5.5 and GPT-6 Luna list the same base rates. Both charge $0.10 per million input tokens and $0.50 per million output tokens at list price.

Anthropic launched Claude Haiku 5.5 on October 7, 2026. It charges $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 above that. OpenAI lists GPT-6 Luna in its API, and it powers the free and Go ChatGPT tiers. OpenAI cut Luna's input price by 50% against GPT-5.6 Luna. Haiku 5.5's list price changes above 100,000 tokens and the sources do not state Luna's, so total cost depends on token use as well as token rates.

Current benchmark tables published by Anthropic place Claude Haiku 5.5 ahead of GPT-6 Luna on computer use, coding, and knowledge evaluations. However, the sources reviewed contain no independent head-to-head cost-per-task study.

Claude Haiku 5.5 vs. GPT-6 Luna: Side-by-Side

Are you one of these vendors? Update your listing

Is Haiku 5.5 Cheaper Than Luna? Price per Token Versus Price per Task

At list price, Claude Haiku 5.5 and GPT-6 Luna are tied for prompts up to 100,000 tokens at $0.10 per million input tokens and $0.50 per million output tokens. The cheaper model for any specific automation workflow is whichever finishes the job with fewer total tokens and fewer failed retries.

Anthropic's own benchmark table has Claude Haiku 5.5 ahead of GPT-6 Luna on every task where it shows a Luna score. Anthropic ran those evaluations itself, so the results are vendor-reported and are not independent.

The sources reviewed contain no independent head-to-head cost-per-task study. Three questions narrow the choice before you run a pilot.

• Does the workflow send prompts over 100,000 tokens? Haiku 5.5 then charges $0.50 input and $2.50 output per million tokens, and the sources do not state Luna's price at that size.
• Does the task involve computer use or agentic coding? Anthropic's figures put Haiku 5.5 ahead on both, though the sources do not show whether that lowers retries.
• Does the application repeat the same context? Haiku 5.5 cache reads cost $0.01 per million tokens up to 100,000 tokens, and the sources do not state Luna's cached input price.

→ Compare per-token pricing for every AI model

Token Pricing and Context Tier Comparison

Claude Haiku 5.5 and GPT-6 Luna list the same entry rates. Haiku 5.5 adds a higher tier above 100,000 tokens and cache prices, and the sources do not state either for Luna. For prompts containing up to 100,000 tokens, Claude Haiku 5.5 charges $0.10 per million input tokens and $0.50 per million output tokens. Prompts that exceed 100,000 tokens transition to a higher tier of $0.50 per million input tokens and $2.50 per million output tokens.

Anthropic says Claude Haiku 5.5 is especially good value for prompts up to 100,000 tokens. That range made up around 90% of requests to the previous Haiku model. Anthropic also offers prompt caching on Claude Haiku 5.5. For prompts up to 100,000 tokens, cache reads cost $0.01 per million tokens and cache writes cost $0.125. Prompts over 100,000 tokens incur cache reads of $0.05 and cache writes of $0.625 per million tokens. Claude Haiku 4.5 cost $1.00 input and $5.00 output. Anthropic says Claude Haiku 5.5 costs on average around 75% less to run than Haiku 4.5. Review detailed tier structures in the Claude Haiku 5.5 pricing guide and on the Anthropic pricing page.

GPT-6 Luna lists identical rates of $0.10 per million input tokens and $0.50 per million output tokens via the OpenAI API. OpenAI reports that this pricing reflects a 50% cut on input tokens and a 58.3% cut on output tokens compared to GPT-5.6 Luna. The sources reviewed do not state Luna's price above 100,000 tokens, its cached input price, or any batch discount. Check the OpenAI pricing page for current updates, or read the GPT-6 Luna pricing overview.

→ Compare per-token pricing for every AI model

Cost-per-Task Formula and a Hypothetical Worked Example

Total task expense is determined by cumulative token volume and retry attempts rather than nominal unit rates alone. Count the input tokens times the input rate, plus the output tokens times the output rate. That is the cost of one attempt. Multiply by the number of attempts needed to finish the task. When two models charge matching list rates, differences in cost come entirely from token usage and failure rates.

Consider an illustrative, hypothetical data extraction task where token counts are round numbers chosen to show the mechanics of the formula. These numbers are purely illustrative and are not measured performance data. Suppose Model A uses 2,000 input tokens and produces 500 output tokens on a single successful attempt. Its cost is 2,000 multiplied by $0.0000001 ($0.0002) plus 500 multiplied by $0.0000005 ($0.00025), which equals $0.00045 for that attempt.

Suppose Model B needs a 4,000-token prompt and produces 1,200 output tokens for the same job. Its cost is 4,000 multiplied by $0.0000001 ($0.0004) plus 1,200 multiplied by $0.0000005 ($0.0006), totaling $0.00100 for one attempt. If Model B fails on its initial try and requires a second attempt to return valid output, its total expense reaches $0.00200. In this hypothetical scenario with identical list rates, Model B costs more than four times as much as Model A due to prompt verbosity and a retry loop.

• Cost per finished task formula: (input tokens x input rate + output tokens x output rate) per attempt, multiplied by total attempts
• At identical token rates, extra tokens from verbose prompts or formatting overhead increase processing costs directly
• Failed attempts that trigger validation retries multiply total task cost by the number of attempts needed

→ Compare per-token pricing for every AI model

What Anthropic's Cost Charts Show and Do Not Show

On its launch page, Anthropic plots accuracy against cost per attempt and cost per task for Claude Haiku 5.5, Claude Haiku 4.5, Claude Sonnet 5.5 and GPT-6 Luna. The page text gives no exact dollar values for GPT-6 Luna. These plots evaluate models across adjustable effort levels: Low, Med, High, Xhigh, and Max.

The charts use logarithmic cost axes, plotting the OSWorld benchmark in cost per attempt and the GDPval-AA benchmark in cost per task. Claude Haiku 5.5 is Anthropic's first Haiku-class model with an adjustable effort setting, so users choose cost versus intelligence.

OpenAI published cost-per-task claims for GPT-6 Luna against its own previous model generation rather than Claude. OpenAI reported that GPT-6 Luna reduced cost per task by 58% at high reasoning effort compared to GPT-5.6 Luna on AutomationBench 1.0.6 across 47 business tools. That comparison shows efficiency improvements inside OpenAI's product line, but it does not establish cost parity or advantage against Claude Haiku 5.5.

→ Compare per-token pricing for every AI model

Benchmark Results Reported by Anthropic

Benchmark scores Anthropic published on October 7, 2026 show Claude Haiku 5.5 leading GPT-6 Luna on every test where Anthropic lists a Luna score. Anthropic conducted these evaluations internally and published them on its launch page, meaning the figures for GPT-6 Luna were generated by a competitor. Anthropic notes that detailed evaluation procedures are available in the Claude Haiku 5.5 System Card.

On GDPval-AA v2.1 (knowledge work, Elo), Claude Haiku 5.5 scored 1620. GPT-6 Luna scored 1437, Claude Haiku 4.5 scored 735 and Claude Sonnet 5.5 scored 1840. On AA-Briefcase v1.1, Claude Haiku 5.5 recorded 1578 Elo, while GPT-6 Luna reached 1336 Elo, Claude Haiku 4.5 scored 614, and Claude Sonnet 5.5 achieved 1824.

On Terminal-Bench 4.0 assessing agentic coding, Claude Haiku 5.5 scored 39.2% while GPT-6 Luna scored 16.4% and Claude Haiku 4.5 scored 0.0%. On the OSWorld 2.1 offline subset measuring computer use, Claude Haiku 5.5 recorded 72.4% compared to 48.9% for GPT-6 Luna and 15.7% for Claude Haiku 4.5. On FrontierCode 1.1 Main, Claude Haiku 5.5 scored 46.4% against 42.4% for GPT-6 Luna. On Chartography (visual reasoning, no tools) it scored 46.4% against 29.1%. On Humanity's Last Exam it scored 45.9% without tools and 57.4% with tools. Anthropic shows no Luna score for that test. For further analysis, see the Claude Haiku 5.5 benchmark report and the GPT-6 Luna benchmark guide.

Where GPT-6 Luna Leads and What Is Not Published

OpenAI cut GPT-6 Luna's prices against GPT-5.6 Luna, but several pricing details are not stated in the sources reviewed. In the sources reviewed, GPT-6 Luna leads Claude Haiku 5.5 on no benchmark Anthropic shows. Its documented points are the matching list price and its place in the free and Go ChatGPT tiers.

GPT-6 Luna powers the free and Go ChatGPT tiers, as covered in What Free ChatGPT Users Get With GPT-6. Its 58% cost-per-task efficiency improvement on AutomationBench 1.0.6 at high reasoning effort is OpenAI's own comparison against GPT-5.6 Luna. It is not a comparison against Claude Haiku 5.5.

The sources reviewed do not state GPT-6 Luna's rates for prompts above 100,000 tokens, its prompt caching read or write prices, or any batch API discount. Check the OpenAI pricing page before sizing a large-context or cached workload.

Claude Haiku 5.5 launched October 7, 2026, under model ID claude-haiku-5-5. It is available on all platforms including Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic says Claude Haiku 5.5's cybersecurity safeguards are more restrictive than Haiku 4.5's and still block penetration testing. Teams that run penetration testing should read the safeguards details on Anthropic's Claude Haiku 5.5 launch page before choosing it. GPT-6 Luna is available through OpenAI's API and powers the free and Go ChatGPT tiers.

Testing Cost per Task Across Fifty Production Workflows

Because published benchmark evaluations do not provide head-to-head dollar figures, teams must gather primary data on their own prompts before scaling up API calls.

A standard pilot begins by selecting 50 diverse inputs from production systems, including document extraction, agent tool calls, and classification requests. Run each input through both APIs and record total input tokens, cached tokens and output tokens.

Each output must then be tested against programmatic validation criteria. If a model generates invalid JSON, misses required entities, or fails an automated schema check, the retry must be executed and logged as an additional attempt. Multiplying token totals by published rates across all attempts establishes the true cost per task for each model.

• Step 1: Select 50 representative prompts from actual business automation workflows
• Step 2: Run all prompts through both model APIs with logging enabled for input, output, and cached tokens
• Step 3: Track automated validation passes and record necessary retry calls for each model
• Step 4: Apply the cost-per-task formula to determine which option yields lower production expenses

→ Compare per-token pricing for every AI model

Workload Selection Criteria for Haiku 5.5 and GPT-6 Luna

Match the workload to what Anthropic and OpenAI report. If your prompts are up to 100,000 tokens and repeat the same system instructions, Claude Haiku 5.5 lists cache reads at $0.01 per million tokens. The sources do not state Luna's cached input price.

For computer use, Claude Haiku 5.5 holds a lead in Anthropic's reported benchmark data. Anthropic reports 72.4% for Haiku 5.5 and 48.9% for GPT-6 Luna on the OSWorld 2.1 offline subset. The sources do not show whether that gap lowers retry costs.

GPT-6 Luna is available via the OpenAI API and powers the free and Go ChatGPT tiers. If your prompts often exceed 100,000 tokens, Claude Haiku 5.5 charges $0.50 input and $2.50 output per million tokens. Check Luna's long-prompt price on the OpenAI pricing page before deploying. For other routing options, see Claude Haiku 5.5 explained, GPT-6 Luna explained, GPT-6 Luna alternatives and Claude Haiku 5.5 vs GPT-6.1.

The Verdict

At list price the two models cost the same for prompts up to 100,000 tokens: $0.10 input and $0.50 output per million. The cheaper one is whichever uses fewer tokens and retries per finished task. Anthropic's reported scores on OSWorld 2.1 (72.4% versus 48.9%) and Terminal-Bench 4.0 (39.2% versus 16.4%) put Claude Haiku 5.5 ahead on Anthropic's own tests. Those figures are vendor-run and no independent study confirms lower retry costs.

OpenAI cut GPT-6 Luna's prices by 50% on input and 58.3% on output against GPT-5.6 Luna. The 58% cost-per-task drop OpenAI reports applies to GPT-6 Luna versus GPT-5.6 Luna on AutomationBench 1.0.6. It is not a comparison with Haiku 5.5.

Anthropic says Claude Haiku 5.5's cybersecurity safeguards are more restrictive than Haiku 4.5's and still block penetration testing. Teams that run penetration testing should read the safeguards details on Anthropic's Claude Haiku 5.5 launch page before choosing it. The sources do not say whether GPT-6 Luna has comparable limits.

The tie at list price would change if independent testing shows GPT-6 Luna reaching equal task completion with fewer output tokens. It would also change if OpenAI publishes a cached input price below $0.01 per million tokens. Similarly, if a workload relies on prompts above 100,000 tokens, Claude Haiku 5.5 charges $0.50 input and $2.50 output per million tokens. The sources do not state GPT-6 Luna's price above 100,000 tokens, so the comparison for long prompts cannot be made here.

Claude Haiku 5.5 is a risky pick for teams whose prompts regularly pass 100,000 tokens: its rates rise fivefold there, and the sources do not state Luna's price in that range, so long-prompt cost cannot be compared here. Neither model is the pick for anyone who needs an independent cost-per-task result before committing, since every benchmark here is Anthropic-run. GPT-6 Luna is not the pick for teams that need computer use or agentic coding first. Anthropic's table puts it behind Haiku 5.5 on OSWorld 2.1 (48.9% versus 72.4%) and Terminal-Bench 4.0 (16.4% versus 39.2%). It is also not the pick for teams that need a published cached-input price today. The sources reviewed state no dollar cost-per-task figure for either model.