Original URL: https://getagenttools.com/blog/opus-5-5-gpt-6-sol-luna/ # Opus 5.5, GPT-6 Sol and Luna: what launched and what changes for agents 22 Sep 2026Marek. - ai-models - agent-tools - claude - codex Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna launched on September 22, 2026. GPT-6 Astra arrived on September 3. For an agent builder, the useful differences are the exact model ID, the product that exposes it, the billing tier, and the request settings that still work. This is a reading of the manufacturers' release notes and documentation, checked on September 22. We have not benchmarked these releases. - **Three launches today.** Anthropic shipped Opus 5.5; OpenAI shipped Sol and Luna. - **Separate access paths.** OpenAI's application rollout covers Work and Codex, with plan and workspace conditions. - **Explicit API names.** Use `claude-opus-5-5`, `gpt-6-sol` or `gpt-6-luna` with the matching provider. - **Prices need a denominator.** The rates below are dollars per million tokens, with the applicable tier stated. - **Migration has contracts.** Opus 5.5 changes thinking settings and forced tool selection. Check those before switching an agent. For the request and response loop behind those changes, start with [what function calling means](https://getagenttools.com/blog/what-is-function-calling/). For the September 29 update, see [GPT-6.1 Sol, dots and Codex plans](https://getagenttools.com/blog/gpt-6-1-sol-dots-codex-plans/). ## Which models launched on September 22? | Display name | Direct API model ID | Release date | Where this announcement applies | | --- | --- | --- | --- | | Claude Opus 5.5 | `claude-opus-5-5` | September 22, 2026 | Anthropic's new Opus release and supported provider platforms | | GPT-6 Sol | `gpt-6-sol` | September 22, 2026 | OpenAI API; rollout in Codex and ChatGPT Work | | GPT-6 Luna | `gpt-6-luna` | September 22, 2026 | OpenAI API; rollout in Codex and ChatGPT Work | | GPT-6 Astra | `gpt-6-astra` | September 3, 2026 | Earlier release, included here to distinguish the dates | The dates come from [Anthropic's Opus 5.5 announcement](https://www.anthropic.com/claude-opus-5-5) and the [OpenAI API changelog](https://developers.openai.com/api/docs/changelog). Sol and Luna accept text and images and return text through Responses and Chat Completions. An API endpoint appearing in a release note does not establish that every tool or request parameter works identically across endpoints. These are direct API IDs. If an application routes through another provider, look up that provider's identifier instead of copying the display name into configuration. ## Where can you use them? OpenAI's [application changelog](https://learn.chatgpt.com/docs/changelog) says Sol and Luna are rolling out to Plus, Pro, Business, Enterprise and Edu users in Codex and ChatGPT Work. Free and Go users get Luna in the desktop app. Availability depends on rollout and workspace settings, and Enterprise administrators must enable the models. Within ChatGPT, the announcement explicitly covers Work and Codex; Chat is excluded. The documented CLI choices are `/model`, `codex --model gpt-6-sol` and `codex --model gpt-6-luna`. Check the picker on the account you actually use. A launch date alone does not prove that a particular workspace has access. Anthropic's [platform release notes](https://platform.claude.com/docs/en/release-notes/overview) list Opus 5.5 on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Provider availability does not establish identical limits or configuration syntax. This article uses the direct Claude API settings for the migration discussion below. ## What do the API tokens cost? All amounts below are **USD per 1 million tokens**. Cached input means a cache read; cache writes are separate. | Model | Input | Cached input | Output | Scope of these rates | | --- | --- | --- | --- | --- | | Claude Opus 5.5 | $4 | $0.20 | $20 | Direct Claude API base rates | | GPT-6 Sol | $2 | $0.20 | $10 | Standard processing, prompts up to 272K input tokens | | GPT-6 Luna | $0.10 | $0.01 | $0.50 | Standard processing, prompts up to 272K input tokens | Sources: [Opus 5.5 overview](https://platform.claude.com/docs/en/models/opus-5-5/overview) and OpenAI's [September 22 API release](https://developers.openai.com/api/docs/changelog). The Opus overview documents a 1M context window and 128K maximum output. Its five-minute cache writes cost $5 per million tokens; one-hour writes cost $8. Batch and faster processing have separate pricing. These are API rates, not ChatGPT subscription prices or Work credits. The table also does not estimate a completed task. A useful task budget needs input, output, cache writes, cache hits, retries and tool charges. Lower token prices cannot tell you how many attempts a workflow will need. ## What breaks when moving to Opus 5.5? Anthropic's [what's new guide](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5) documents two request changes that deserve an explicit configuration check: 1. **Thinking stays on.** Omit `thinking` or use adaptive thinking. Requests that disable it, or use manual thinking with `budget_tokens`, return a 400 error. The default effort is now medium; Opus 5 used high. 2. **Forced tool choice is gone.** `tool_choice` with `any` or a named `tool` returns a 400 error. `auto` and `none` remain supported. `strict: true` constrains a tool call's arguments when a call occurs; it does not guarantee the model makes one. Two response and integration changes are covered in the [migration guide](https://platform.claude.com/docs/en/models/opus-5-5/migration-guide). Text between tool calls now arrives in thinking blocks. With the default omitted thinking display, those blocks can be empty, so a UI that expected visible progress text may lose it. Select content blocks by type rather than assuming the first block contains the answer. Computer use also needs a provider check. On the Claude API and Google Cloud, `computer_20251124` is unsupported; migrate to `computer_toolset_20260801`. Bedrock retains support for the older computer tool. Follow the guide for conversation history and thinking-block handling before resuming an existing thread on a different model. ## A proposed check before changing an agent The following is our suggested evaluation procedure, not a report of tests we have run: 1. **Record the complete configuration.** Save the provider, model ID, endpoint, effort, tool definitions and limits alongside the old version. Keep credentials out of the record. 2. **Choose a small set of representative tasks.** Include a successful tool call, a request that needs no tool, an invalid argument, and a failed tool result. Add a long conversation if your agent resumes work across turns. 3. **Define success before running.** For a coding agent, this could be passing the existing checks with no changes outside the allowed files. For extraction, require valid fields and a traceable source for each value. 4. **Hold the environment steady.** Use the same inputs and permitted tools. Record retries, human corrections, elapsed time and total cost. A cheaper successful call can still sit inside a more expensive task. 5. **Keep a way back.** Change a bounded workflow first and retain its previous configuration until you can explain any difference in behavior. A model change is also a good point to inspect [the instructions and hidden context that make up a system prompt](https://getagenttools.com/blog/what-is-a-system-prompt/). Record those inputs with the evaluation; otherwise, a prompt change can look like a model improvement. The release notes establish availability, prices and documented interfaces. Choosing a default for your agent still needs evidence from that agent's tasks. Keep the model ID and configuration with every result so the next decision has something concrete to compare. ## Frequently asked questions ### Did GPT-6 Astra launch alongside Sol and Luna? No. Astra launched on September 3, 2026. The release table above separates the dates. ### Can I use Sol and Luna in ChatGPT Chat? The announced rollout covers Work and Codex, with Chat excluded. See [availability and workspace conditions](https://getagenttools.com/blog/opus-5-5-gpt-6-sol-luna/#where-can-you-use-them) before choosing a model. ### Do these API prices tell me what an agent task will cost? No. They price tokens, while a task can involve retries, cache writes and tool charges. Use the [rates and billing scope](https://getagenttools.com/blog/opus-5-5-gpt-6-sol-luna/#what-do-the-api-tokens-cost) alongside your own usage records. ### Can Opus 5.5 still be forced to call a tool? No. The documented forced-tool options return a 400 error. The migration section above explains supported settings and why strict arguments do not guarantee a call. ### Which of these models should be my agent's default? This article does not establish a winner. Use the [proposed evaluation procedure](https://getagenttools.com/blog/opus-5-5-gpt-6-sol-luna/#a-proposed-check-before-changing-an-agent) on your own tasks before changing the default. Put this into practice [See the tool list](https://getagenttools.com/tools/)[Train your team](https://getagenttools.com/#training)