GPT-6 Luna: API Pricing, Caching and Task Budgets

The general API model ID is gpt-6-luna. It accepts text and images and produces text, with a 1,050,000-token context and a 128,000-token maximum output. Image generation is available through a tool in Responses; the model’s native output is text.

OpenAI’s release announcement makes Luna available in ChatGPT Work and Codex to Plus, Pro, Business, Enterprise and Edu users, with Free and Go access in the desktop app. It says the new models are not available in Chat at that release. Availability in a subscription surface and metered API billing are separate choices: the token budgets below describe API calls, not a ChatGPT plan allowance.

Use Responses for built-in tools and function calling. Chat Completions supports function calling only at reasoning_effort none. Responses effort options are none, low, medium (the default), high, xhigh and max. An agent client supplies the repository, tool execution and permissions; a model choice alone does not run a workflow.

These are USD per million billed tokens on the Standard OpenAI API, checked October 8, 2026. For prompts above 272,000 input tokens, input and cache rates double and output rates increase by 50% for the entire request. The context ceiling and the base-price window are different limits.

Batch and Flex use 50% of applicable Standard rates; Fast uses twice the applicable rates. Regional processing adds 10% where available. Add separately billed tools and every retry to the token bill. Reasoning tokens count toward billed output even when they are not part of the visible answer.

An input token is charged as ordinary input, a cache read or a cache write. The write price replaces the ordinary input price for that token; it is not added on top. Read input_tokens_details.cached_tokens and cache_write_tokens from response usage, and subtract those from total input_tokens before pricing the ordinary remainder.

Luna uses the newer prompt cache controls: a minimum reusable prefix of 1,024 visible input tokens and a default minimum lifetime of 30 minutes after the latest write or reuse. In explicit mode, set prompt_cache_options.mode to explicit and place supported prompt_cache_breakpoint markers. Without markers, that mode creates no cache writes. Implicit mode chooses eligible boundaries automatically.

For five requests with an eligible 20K-token prefix, 1K new input and 500 billed output tokens per request, one prefix write costs $0.0025 and four full prefix reads cost $0.0008. Adding new input ($0.0005) and output ($0.00125) gives $0.00505 for the group. This assumes four actual hits in the base tier; changes to the prefix or cache misses change the result.

At Standard rates, assumed uncached usage of 20K input and 2K output costs $0.003. A 120K-input/10K-output request costs $0.017, while 300K input and 10K output costs $0.0675 after the long-context multiplier. Here uncached means tokens billed at the ordinary input rate, not an eligible prefix billed as a new cache write.

A run of 10,000 accepted extraction jobs at the first example’s usage would cost $30 in token charges if every job succeeds in one attempt. If 1,000 jobs each need one identical additional attempt, it becomes $33. Those are planning examples. Save all attempts and returned usage, then divide the total bill by accepted jobs; rejected answers still consume budget.

For tool workflows, retain tool charges and the cumulative conversation sent on later turns. A short final reply can follow substantial reasoning and repeated input. Choose effort against your acceptance checks rather than using visible answer length as a proxy for spend.

Set OPENAI_API_KEY on your server and send the example through Responses. Explicit cache mode without a breakpoint keeps this short trial out of cache-write billing. It uses medium effort and a billed output cap; these settings make the experiment easier to compare.

Check the extracted vendor and amount against the invoice. Keep the returned usage, incomplete status and answer together. Count an incomplete or incorrect result as an attempt needing review before expanding to a larger document collection.

Both start at $0.10 input and $0.50 output per million tokens. Haiku’s fivefold price step begins above 100K prompt tokens; Luna’s smaller increase begins above 272K. Equal billed usage of 120K input and 10K output costs $0.085 on Haiku and $0.017 on Luna. Identical text can tokenize differently, so recount prompts before migrating.

Plotly’s private 43-question analytics evaluation reports Luna at 38 correct and $0.13 for the full round, versus Haiku at 37 and $0.38. Haiku’s reported median is 14 seconds per question, versus about 18 for Luna. Effort and repeat counts are undisclosed; this small analytics set supports a task-specific tradeoff.

The existing Haiku comparison also covers BrenBuilds’s small TypeScript app: the same tasks and hidden checks used different Claude Code and Codex CLI harnesses. Its cost ranges bound incomplete logs, and the explanation timings exclude a failed Luna-medium sample. Read that method before treating the results as a prediction for your own agent.

For a fixed label, yes/no condition or ordered rubric, OpenAI Decisions serves Luna through a separate /v1/decisions endpoint. It returns structured probabilities and uses input-only pricing. Use general Responses when you need a written reply or tools; use Decisions when application code can act on a defined answer set.

Compare that endpoint with Perplexity Decider v1.1 on the same labeled cases and threshold policy. Include uncertain cases and refusals in the review path, and price each endpoint from its actual usage rather than borrowing a chat output-token estimate.

OpenAI’s October 14 deprecation notice schedules gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 for API shutdown on January 14, 2027. It recommends gpt-6-luna for those exact snapshots; their fine-tuned versions follow the same shutdown schedule.

Before changing an application, compare Luna’s endpoint, tool support, prompt/token counts and accepted outputs with the old integration. This is a retirement and migration notice, not a new Luna release or a change to its price table. Preserve an evaluation set and verify the replacement before the shutdown date.