GPT-6 Sol and Luna are released: pricing, differences and choices for Codex/API

When did GPT-6 Sol and Luna launch, and why are they no longer rumors?

OpenAI’s ChatGPT Release Notes record the launch of GPT-6 Sol and Luna in ChatGPT Work and Codex on September 22, 2026. Official API model pages list gpt-6-sol and gpt-6-luna respectively. Primary sources now establish the release date and model names; earlier assessments based on the absence of official entries before September 19 no longer apply.

This article retains its original URL and first-publication date while updating release facts, pricing definitions and selection guidance. The word rumor in the old URL reflects the article’s original tracking of rumors, not the current release status. Ordinary ChatGPT conversations, ChatGPT Work, Codex and the API are not interchangeable entry points.

Choosing Sol or Luna: group tasks first, then assess total cost

Official positioning places Sol in complex coding and agent work, and Luna in focused, high-volume tasks. Editorial suggestion: include Sol in comparisons for cross-file changes, repeated tool use or complex debugging; include Luna for clearly defined classification, extraction and small rewrites. This is a starting point for selection, not a success guarantee.

Moving every task to Luna because of its low unit price can let rework and extra review consume the savings. Start with tasks whose correctness your team can judge. Fix inputs, tool permissions and acceptance criteria, then record completion rate, human revision count, duration and actual token usage before deciding which tasks to migrate. This article does not claim to have benchmarked both models.

GPT-6 Sol and Luna API pricing: distinguish Standard from other modes

As of September 27, 2026, official Standard short-context prices per million tokens are: GPT-6 Sol input $2, cache read $0.20, cache write $2.50 and output $10; GPT-6 Luna input $0.10, cache read $0.01, cache write $0.125 and output $0.50.

Batch and Flex are 50% of the corresponding Standard prices; Fast is 2 times the price. When input exceeds 272,000 tokens, input and cache prices for the entire request become 2 times the base rates, and output becomes 1.5 times the base rate. Eligible regional processing adds 10%; EU data residency for Sol/Luna supports Standard only. Do not treat discounted-mode pricing as the price of ordinary requests.

Budgets should itemize input, cache read, cache write, output and tool charges, with mode and context tier recorded. For example, a short-context Standard request with no caching, tools or regional surcharge and 10,000 input plus 2,000 output tokens costs $0.04 in model tokens for Sol and $0.002 for Luna. This is a unit-price calculation, not an actual invoice or speed test.

Where to check ChatGPT Work, Codex and API access

For ChatGPT Work or Codex, first inspect the current workspace’s model selector and settings. Official release notes explicitly say availability varies by plan and workspace settings, so “released” does not prove availability at every entry point of your accounts.

For API use, follow gpt-6-sol or gpt-6-luna on the official model pages and check permissions and actual responses in an existing authorized test project. Do not apply old gpt-5.6-sol behavior or configuration to the new models without validation. API unit prices also cannot be converted into a fixed number of subscription tasks.

When you see allowance or reset notices, check the usage page for the relevant product first. Continue with the Codex usage-reset verification checklist to distinguish natural recovery, automatic resets and reset counts that can be saved.

Developer migration checklist: validate more than the model name

Save the current model, processing mode, prompts, tool definitions and acceptance cases first, then compare new models in an environment without production side effects. As needed, validate text output, structured results, streaming responses, tool parameters and long-input boundaries. Use test data and human confirmation first for tools that write.

On September 25, OpenAI said it had fixed an image-encoding issue in Sol and Luna that reduced image-understanding quality, affecting visual tasks and computer use in the API and Codex. Teams using image inputs should rerun pre-fix evaluations and affected workflows; old results should not be treated as current model performance.

Official documentation says the Responses API supports built-in tools and function calling; Chat Completions supports function calling only when reasoning_effort is none. Both models’ reasoning levels include none, low, medium, high, xhigh and max, with medium as the default. Check this combination before reusing old interface configuration; an interface constraint is not model unavailability.

If output fields are unstable, tool actions are wrong or costs exceed budget, stop expanding use and return to the previously validated configuration. Keep request identifiers, mode and usage evidence rather than replacing all tasks based on one successful example.

Missing models or failed calls: separate permissions, allowances and connections

If a model entry point is missing, first check the product, account, workspace and official model name. Handle permission or allowance notices according to their meaning. Changing IP addresses does not alter workspace permissions or restore exhausted allowances.

Once model identity and eligibility are confirmed, use the proxy connection troubleshooting checklist if DNS, TLS, proxy 407 or connection timeout errors remain. Server incidents, permission denials and connection timeouts require different evidence.

Current conclusion and what to recheck next

GPT-6 Sol and Luna now have official release and API documentation and merit comparison on real tasks. Measure migration benefits through completion quality, cost and rework. This article provides no untested performance ranking and does not guarantee that a particular subscription account will see the entry point.

Prices, entry points, the September 25 image-encoding fix and interface restrictions are summarized from official material directly readable on September 27, 2026. Future checks should focus on plan scope, processing modes, long-context charges and tool combinations, updating this page when new facts emerge.