GPT-6 Sol is OpenAI’s new mid-tier model for coding and agentic work, and GPT-6 Luna is its cheap, fast sibling for high-volume tasks. Both launched on 22 September 2026, sit below GPT-6 Astra, and cost half what their GPT-5.6 predecessors did in the API: $2/$10 per million tokens for Sol and $0.10/$0.50 for Luna.
The catch is where you can use them. In ChatGPT, Sol and Luna live in Work mode and Codex, not the normal Chat box, which is why many Plus subscribers think they are missing. Independent testing also shows the capability jump over GPT-5.6 is small. Details below are correct as of 24 September 2026.
What are GPT-6 Sol and GPT-6 Luna?
GPT-6 Sol and GPT-6 Luna are two smaller GPT-6 models OpenAI released on 22 September 2026. Sol is built for complex coding and agentic workflows. Luna is OpenAI’s most efficient model for high-volume work such as summarisation, extraction and quick answers. Both accept text and images, return text, and share a 1.05M-token context window.
OpenAI says both were “trained with similar methods” to GPT-6 Astra, and that Sol brings “stronger factual reliability”. There is no GPT-6 Terra: the middle tier from the GPT-5.6 generation was not carried over.
Where do Sol and Luna sit next to GPT-6 Astra?
Astra is the flagship. It arrived in limited preview on 3 September 2026 and costs $10/$50 per million tokens, five times Sol’s rate (see our GPT-6 Astra guide). Sol is the everyday workhorse, and Luna costs one-twentieth of Sol. Early reviewers suggest Astra plans, Sol does the work and Luna takes cheap subtasks. For which tier suits which job, read GPT-6 Sol vs Luna vs Astra.
When did GPT-6 Sol launch, and why the same day as Claude Opus 5.5?
GPT-6 Sol and Luna launched on 22 September 2026, with rollout continuing through the day. Anthropic released Claude Opus 5.5 first, and OpenAI’s announcement followed about an hour later. That timing matters for benchmarks: OpenAI’s comparison charts use Claude Opus 5 and Fable, not Opus 5.5, because neither lab could include the other’s new model.
Opus 5.5 costs $4/$20 per million tokens, double Sol, but scores higher on independent tests. Our Claude Opus 5.5 explainer covers Anthropic’s side, and the GPT-6 Sol vs Claude Opus 5.5 comparison puts the two head to head.
What are the GPT-6 Sol API model IDs, context window and effort levels?
The API model IDs are gpt-6-sol and gpt-6-luna, with no dated snapshots listed. Both offer a 1,050,000-token context window, 128,000 maximum output tokens and six reasoning effort levels: none, low, medium (the default), high, xhigh and max. Sol’s knowledge cutoff is 20 April 2026, and Luna’s is 18 May 2026.
According to OpenAI’s GPT-6 Sol model documentation, the Responses API supports web search, file search, code interpreter, computer use, MCP, function calling and structured outputs. Fine-tuning is not supported. Artificial Analysis lists Sol’s context as 872K, so treat OpenAI’s 1.05M figure as a ceiling.
How much does GPT-6 Sol cost? Full GPT-6 Sol pricing
GPT-6 Sol pricing is $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 input, $0.01 cached and $0.50 output. OpenAI says that is half the price of GPT-5.6, measured against GPT-5.6’s promotional pricing rather than its original list price.
| Model (per 1M tokens) | Input | Cached input | Output | Over 272K input: input / output |
|---|---|---|---|---|
| GPT-6 Sol | $2.00 | $0.20 | $10.00 | $4.00 / $15.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | $0.20 / $0.75 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $20.00 / $75.00 |
| GPT-5.6 Sol (promo) | $4.00 | $0.40 | $20.00 | 2x input / 1.5x output |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | 2x input / 1.5x output |
Standard rates as of 24 September 2026. Batch and Flex cost half. Fast mode (formerly priority processing) costs double, so $4/$20 for Sol. Regional data residency adds 10%. Luna’s output cut is actually 58%, from $1.20 to $0.50.
What is the 272K long-context surcharge?
Once a request passes 272K input tokens, the whole request, not just the overflow, is billed at 2x the input and cache rates and 1.5x the output rate. A Sol call with 300K input and 10K output costs about $1.35. At standard rates the same tokens would cost $0.70. The rule applies to Luna too.
Split Prompts Before They Cross 272K
Because the surcharge reprices the entire request, one 280K-token Sol prompt costs about $1.12 in input alone, while two 140K-token calls cost $0.56 combined. For long document or codebase jobs, chunk inputs to stay under 272K per call and keep a stable shared prefix so the $0.20 cached rate applies to the repeated part.
How good is GPT-6 Sol? OpenAI benchmarks vs independent scores
OpenAI-reported results show GPT-6 Sol beating Claude Opus 5 on business automation and landing close to Fable 5 on bug fixing, at a fraction of the cost. Independent scores are more sober. On the Artificial Analysis Intelligence Index, Sol (max) scores 48, one point above GPT-5.6 Sol, and Luna is flat at 37.
Headline OpenAI-reported numbers from the GPT-6 Sol and Luna announcement: Sol (xhigh) scored 33.2% on AutomationBench against Claude Opus 5 at 26.9%. On DeepSWE v1.1, Sol (max) reached 68.8% against Fable 5 at 69.9%, at roughly 80% lower cost per task, and Luna (max) hit 66.6%. OpenAI also says Sol “makes about half as many mistakes as its predecessor, approaching Astra-level reliability”.
The Artificial Analysis review of Sol and Luna says the index scores “remain level with GPT-5.6”. Its measured changes:
- Coding Agent Index: Sol up from 55 to 57, Luna down from 43 to 41.
- Hallucination rate on AA-Omniscience: Sol falls from 92% to 60%, Luna from 93% to 77%.
- GDPval-AA knowledge work: Sol drops about 100 Elo, Luna about 75.
- Cost per index task: Sol $1.06 versus $1.99 for GPT-5.6 Sol.
For context, Claude Opus 5.5 (max) scores 58 and GPT-6 Astra 53 on the same index. Sol is a cost story, not a new frontier.
Why is GPT-6 Sol not showing in ChatGPT?
GPT-6 Sol is not showing in ChatGPT because it is not a Chat model. The OpenAI Help Center article on ChatGPT Work and Codex says Sol and Luna “are not available in regular ChatGPT conversations.” Switch to Work using the toggle at the top of the screen, or open Codex, and the models appear.
Chat keeps older models: GPT-5.6 Sol for Plus, plus the Astra-powered GPT-6 Pro on Pro, Business and Enterprise. If Work itself fails to load, check whether ChatGPT is down on the OpenAI status page. The same logic applies to the flagship, covered in our guide on how to access Astra in ChatGPT.
ChatGPT Work vs Chat: what is the difference?
Chat is for “fast, conversational assistance and everyday questions”. Work is “an agent designed for longer, multi-step work and finished deliverables”. Graham Barlow at TechRadar’s explainer on ChatGPT Chat vs Work sums it up: “Chat feels collaborative; Work feels much more like delegation.”
Usage and credit rules in Work and Codex are separate from Chat. TechRadar reports that Plus subscribers see GPT-6 Luna High, Sol Light, Sol Medium, Astra Light and Astra Medium in Work.
Which ChatGPT plans get GPT-6 Sol and Luna?
| Plan | Chat mode | Work and Codex |
|---|---|---|
| Free and Go | GPT-5.6 Luna | GPT-6 Luna in the desktop app only |
| Plus | GPT-5.6 Sol | Sol, Luna and limited Astra |
| Pro ($100 / $200) | GPT-6 Pro, GPT-5.6 Sol | Sol, Luna, Astra |
| Business | GPT-6 Pro | Sol, Luna, Astra |
| Enterprise | GPT-6 Pro | Sol and Luna off until an admin enables them |
| Edu | Not announced | Sol and Luna |
As of 24 September 2026. OpenAI’s workspace documentation says Sol and Luna “are off by default in Enterprise workspaces at launch” and that an administrator must enable each model before members can select it. Work itself requires Enterprise Key Management (EKM) on eligible Enterprise and Edu workspaces.
Is GPT-6 Luna free?
Yes, with limits. Free and Go users can try GPT-6 Luna in the ChatGPT desktop app. It is not in Chat on the web or mobile, which stays on GPT-5.6 Luna. OpenAI has not published a usage cap for free desktop access. In the API, Luna is paid but costs just $0.10 per million input tokens.
How do you use GPT-6 Sol in Codex, GitHub Copilot, Foundry and OpenRouter?
GPT-6 Sol and Luna were available on day one in the OpenAI API, Codex, GitHub Copilot, Microsoft Foundry and OpenRouter. In Codex, pick Sol with the /model command, pass –model gpt-6-sol on the command line, or set it as the default in config.toml. Use codex exec -m gpt-6-sol for CI runs.
The GitHub Copilot changelog for GPT-6 Sol and Luna lists Sol on Copilot Pro+, Max, Business and Enterprise, and Luna on those plans plus Pro. Both are on by default unless an admin turns them off.
Microsoft Foundry made both generally available in 28 Global regions at OpenAI’s prices, with Data Zone deployments costing extra. OpenRouter listed both at the same prices on launch day. For the wider toolkit, see our roundup of AI tools for developers.
If Sol Is Missing From Codex CLI
A GitHub issue opened on 23 September reported that stable Codex CLI 0.154.0 did not list Sol, which only appeared in an alpha build and the desktop app. Check your CLI version before assuming a plan problem. Once it works, start at medium effort: Artificial Analysis measured a 142-second time to first token for Sol at max.
What are reviewers and developers saying about GPT-6 Sol?
Reaction to GPT-6 Sol is mixed to positive, and almost all of it centres on price. Reviewers praise Luna’s cost and Sol’s value for agent work. Critics point out that the capability gain over GPT-5.6 is small, the Chat and Work split is confusing, and some Codex runs have become slower.
Simon Willison, in his launch-day write-up of Opus 5.5, Sol and Luna, called Luna “one of the cheapest models OpenAI have ever released” and moved his Datasette Agent demo to it.
The criticism is just as consistent. Matthias Bastian at The Decoder’s analysis of the Sol and Luna launch wrote that the models “cut prices in half but barely move the needle on performance”, and noted that OpenAI left GDPval and Terminal-Bench 4.0 out of its presentation.
Developers have flagged speed: a GitHub issue on the Codex repository reported Sol tasks taking over 40 minutes where GPT-5.6 Sol took about 20. Naming fatigue is a common complaint too.
Should you switch from GPT-5.6 Sol to GPT-6 Sol?
Switch to GPT-6 Sol for most API coding and agent workloads. Independent tests show it matching or slightly beating GPT-5.6 Sol at about half the cost per task, with far fewer hallucinations. Hold back on knowledge-work deliverables such as reports, where GDPval-AA scores fell, on latency-sensitive apps, and on prompts above 272K tokens.
For Luna, the maths still favours switching. It uses more tokens per task than GPT-5.6 Luna (51K versus 41K), yet still costs $0.07 per index task against $0.18. Its Coding Agent score dipped slightly, so test coding subtasks first.
There is no rush. As of 24 September 2026, OpenAI has posted no deprecation notice for GPT-5.6 Sol or Luna, and GPT-5.6 Sol’s promotional price holds at least through 21 November 2026. That gives you time to run both side by side.
Conclusion
GPT-6 Sol and Luna are a price cut with a modest capability bump attached. For API teams running agents or coding pipelines on GPT-5.6, set up a side-by-side eval this week, keep requests under 272K tokens, and move traffic once your own results match the published ones.
For ChatGPT users, the fix is simpler: switch to Work or open Codex, and ask your Enterprise admin to enable the models if they are still hidden. If you are weighing these models for course design, content drafting or learner support, read our guide to GPT-6 Sol and Claude Opus 5.5 for L&D teams next.
FAQ
Q1. Is GPT-6 Sol free to use?
No. GPT-6 Sol is only available on paid ChatGPT plans (Plus, Pro, Business, Enterprise and Edu), inside Work mode and Codex, or through the paid API at $2 per million input tokens and $10 per million output tokens. Free and Go users get GPT-6 Luna, the cheaper sibling, in the ChatGPT desktop app only, as of 24 September 2026.
Q2. Does ChatGPT Plus include GPT-6 Sol?
Yes, but not in the Chat box. Plus subscribers can use GPT-6 Sol, GPT-6 Luna and a limited amount of GPT-6 Astra in ChatGPT Work and Codex. Regular Chat on Plus still runs GPT-5.6 Sol. Use the Chat and Work toggle at the top of the screen to switch, and note that Work usage limits are separate from Chat.
Q3. How do I enable GPT-6 Sol in a ChatGPT Enterprise workspace?
An owner or admin has to turn it on. OpenAI says GPT-6 Sol and GPT-6 Luna are off by default in Enterprise workspaces at launch, and each model must be enabled before members can select it. Workspace settings override local configuration. Work itself is available to eligible Enterprise and Edu workspaces that have Enterprise Key Management (EKM) enabled.
Q4. What is the cached input price for GPT-6 Sol?
Cached input for GPT-6 Sol costs $0.20 per million tokens, a tenth of the $2 standard input rate. Cache writes cost $2.50 per million. Above 272K input tokens the cached rate doubles to $0.40 along with the rest of the request. GPT-6 Luna’s cached input is $0.01 per million. OpenAI says you can change reasoning effort without losing the cache.
Q5. Is GPT-6 Sol better than Claude Opus 5.5?
Not on raw capability. Artificial Analysis scores Claude Opus 5.5 (max) at 58 on its Intelligence Index against 48 for GPT-6 Sol (max). Sol is much cheaper, though: $2/$10 per million tokens against $4/$20, and about $1.06 per index task against roughly $5.98 for Opus 5.5 at max effort. Sol suits high-volume agent work where cost matters most.
Q6. Can I use GPT-6 Sol in GitHub Copilot or Microsoft Foundry?
Yes, both had it on launch day. GitHub Copilot offers Sol on Pro+, Max, Business and Enterprise, and Luna on those plans plus Pro, turned on by default unless an admin disables them. Microsoft Foundry made Sol and Luna generally available in 28 Global regions plus US and EU Data Zones, at prices matching OpenAI’s own API.
Q7. Why is there no GPT-6 Terra model?
OpenAI did not carry the Terra tier over from the GPT-5.6 generation. The GPT-6 family launched with three models: Astra as the flagship, Sol as the mid-tier model for coding and agentic workflows, and Luna as the low-cost model for high-volume tasks. GPT-5.6 Terra remains listed in the API with no deprecation notice as of 24 September 2026.
