GPT-6.2 Astra and Sol at a glance
| GPT-6.2 Astra | GPT-6.2 Sol | |
|---|---|---|
| Role (OpenAI) | Frontier model for your hardest projects | Efficient, versatile model for everyday work |
| API model ID | gpt-6.2-astra | gpt-6.2-sol |
| Released | November 17, 2026 | November 17, 2026 |
| Context window | 1,050,000 tokens (max input 922,000) | 1,050,000 tokens (max input 922,000) |
| Max output | 128,000 tokens | 128,000 tokens |
| Input / output | Text and image in, text out | Text and image in, text out |
| Reasoning effort | low, medium, high, xhigh, max | low, medium, high, xhigh, max |
| Knowledge cutoff | May 18, 2026 | June 1, 2026 |
| API price per 1M tokens | $8 input, $0.40 cached, $40 output | $2 input, $0.10 cached, $10 output |
| Compared by OpenAI with | GPT-6 Astra | GPT-6.1 Sol |
| Where available | ChatGPT Work and Codex (Plus, Pro, Enterprise, Edu), API, GitHub Copilot, Microsoft Foundry | ChatGPT Work and Codex (Plus, Pro, Enterprise, Edu), API, GitHub Copilot |
GPT-6.2 Astra vs Sol: which to use
The two models share a context window, output limit, tools and reasoning settings. They differ in capability and price. OpenAI's API changelog puts it this way: use Astra for demanding reasoning and coding tasks, and Sol for everyday coding and agent workflows.
- Price: Sol is $2 and $10 per million input and output tokens. Astra is $8 and $40, four times as much.
- Capability: on OpenAI's charts Astra scores higher than Sol on every percentage score in the table below, for example 77.1% vs 68.0% on Terminal-Bench-Science and 52.4% vs 41.2% on AutomationBench.
- Coding at max effort: on Cognition's FrontierCode 2.0, OpenAI says Sol came within half a point of Astra at about one fifth the cost. That figure is for this one benchmark, not a general cost ratio.
- Default: new ChatGPT Work chats start on Sol (Light effort on consumer plans, Medium in Enterprise workspaces), and OpenAI says many of its own employees use Sol as their default.
- Rule of thumb: run Sol for daily coding and agents, and move to Astra for hard bugs across services, large changes and work where Sol stalls.
GPT-6.2 benchmarks
These are OpenAI's own numbers from its launch charts, as reported by VentureBeat, which notes that each model's best plotted score is used, so effort settings may differ. Astra is compared with GPT-6 Astra and Sol with GPT-6.1 Sol. They are not independent tests.
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Sol |
|---|---|---|---|---|
| Agents' Last Exam | 78.3% | 73.1% | 73.0% | 72.5% |
| FrontierCode 2.0 (Cognition) | 44.2% | 42.8% | 43.8% | 39.0% |
| Terminal-Bench-Science | 77.1% | 68.1% | 68.0% | 57.0% |
| AutomationBench | 52.4% | 41.4% | 41.2% | 36.1% |
| BenchCAD 2.1 Core | 82.5% | 62.6% | 73.1% | 52.6% |
| Harvey LAB-AA | 15.0% | 8.6% | 11.1% | 6.9% |
| Vals Science Index | 68.4% | 65.1% | 58.5% | 59.6% |
| LifeVal | 86.1% | 79.8% | 83.8% | 79.5% |
| GDPval v2 | +6.2 points | - | +5.3 points | - |
What the numbers say, and what they leave out
- Sol vs GPT-6.1 Sol: Sol gains on most rows, most of all on BenchCAD (73.1% vs 52.6%) and Terminal-Bench-Science (68.0% vs 57.0%). It scores slightly lower on the Vals Science Index (58.5% vs 59.6%), and Agents' Last Exam barely moves.
- Efficiency: OpenAI says Astra's GDPval v2 gain came with 22% fewer output tokens than GPT-6 Astra, and both models finish demanding tasks with fewer output tokens.
- Against Claude Fable 5.2, OpenAI's post says Astra scores 5.5 points higher on Agents' Last Exam v1.1 at 5% of the cost. OpenAI notes the leaderboard's Fable 5.2 results exclude nine tasks that used a fallback model.
- On Terminal-Bench-Science, OpenAI's footnote says the Fable 5.2 figure in its chart is Artificial Analysis's run in mini-swe-agent. Anthropic reported 74.6% at max reasoning in Claude Code.
- Artificial Analysis's AA-3D (blinded human votes on 3D and browser builds): Astra 1,318 Elo, Claude Fable 5.2 1,271, Sol 1,259.
- Artificial Analysis's Coding Agent Index (in Codex, max effort), as reported by VentureBeat: Astra flat and Sol down 3 points, largely from a 14-point drop on SWE-Atlas-QnA-AA.
- No SWE-bench score was published in any source we read.
GPT-6.2 pricing
API prices per million tokens, from OpenAI's model pages on November 18, 2026. Astra's input and output are 20% below GPT-6 Astra and its cached input is 60% lower. Sol keeps GPT-6.1 Sol's prices. If a request has more than 272K input tokens, input and cache rates double and output is 1.5x for the whole request. Batch is 50% of standard. Fast mode is 2x standard. Ultrafast is 3x standard through February 28, 2027 and 6x from March 1, 2027, with US data residency only. Cache writes cost $10 (30-minute) or $20 (12-hour) on Astra and $2.50 or $5 on Sol.
| Model | Input | Cached input | Output | Over 272K input (input / output) | Batch (input / output) |
|---|---|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 | $16 / $60 | $4 / $20 |
| GPT-6.2 Astra, Fast (2x) | $16 | - | $80 | - | - |
| GPT-6.2 Astra, Ultrafast (3x until Feb 28, 2027) | $24 | - | $120 | - | - |
| GPT-6.2 Sol | $2 | $0.10 | $10 | $4 / $15 | $1 / $5 |
| GPT-6.2 Sol, Fast (2x) | $4 | - | $20 | - | - |
| GPT-6.2 Sol, Ultrafast (3x until Feb 28, 2027) | $6 | - | $30 | - | - |
| GPT-6 Astra | $10 | $1 | $50 | - | - |
| GPT-6.1 Sol | $2 | $0.10 | $10 | - | - |
Where you can use GPT-6.2
- ChatGPT Work and Codex: Plus, Pro, Enterprise and Edu, per OpenAI's launch post. Enterprise and Edu admins turn the models on first. In ChatGPT the models are in Work, not regular Chat. Codex and Work share usage. For what Work can do outside coding, see ChatGPT for marketing.
- Ultrafast in ChatGPT and Codex: Pro $500 gets it with either model, Pro $200 gets it with Sol, and Enterprise and Edu get it once an admin enables it.
- OpenAI API:
gpt-6.2-astraandgpt-6.2-solon the Responses, Chat Completions and Batch endpoints. Tool calling needs the Responses API. There is no Realtime, Assistants or fine-tuning support. US and EU data residency cover Standard, Fast and Batch. - GitHub Copilot: generally available since November 17 on Copilot Pro+, Max, Business and Enterprise, in VS Code, Visual Studio, Copilot CLI, the coding agent, github.com, JetBrains, Xcode, Eclipse and others. The rollout is gradual, and Business and Enterprise admins enable it through model policy.
- Microsoft Foundry:
gpt-6.2-astrais listed in the Foundry catalog as generally available, version 2026-11-17. The catalog page shows no prices or regions.
How to use GPT-6.2 in Codex
Update Codex and pick the model
In the Codex app, choose GPT-6.2 Sol or GPT-6.2 Astra in the model picker. In the CLI, update to 0.169.1 or later (released November 17) and run
/model. On a ChatGPT plan, it needs Plus, Pro, Enterprise or Edu.# Codex CLI 0.169.1 or later, then pick the model in the session /model # choose GPT-6.2 Sol or GPT-6.2 Astra # Codex app: choose the model in the model pickerStart on Sol, step up to Astra
OpenAI's Codex docs recommend Astra for demanding work and Sol for everyday work. If you don't set a model, Codex uses one it recommends.
Use the long context when you need it
Both models take up to 1 million tokens of context when configured in Codex.
Calling the API? Check the migration notes
Reasoning effort must be
lowor higher, sincenoneis rejected, and OpenAI suggests starting new integrations atmedium. Omittemperature,top_p,logprobsandtop_logprobs. Use the Responses API for tool calls. Keep prompt prefixes unchanged and append new instructions so the cache can be reused.# OpenAI API: change the model ID model = "gpt-6-astra" # before model = "gpt-6.2-astra" # after model = "gpt-6.1-sol" # before model = "gpt-6.2-sol" # after # reasoning effort: low | medium | high | xhigh | max # "none" is rejected; tool calling needs the Responses API
GPT-6.2 vs Claude Fable 5.2 and Opus 5.5
List prices and specs only. Each vendor publishes its own benchmarks with its own setup, so the scores aren't a head-to-head run. For the full comparison, see GPT-6.2 vs Claude.
- GPT-6.2 Astra vs Claude Fable 5.2: the same per-token price, $8 input and $40 output. Cached input is $0.40 on Astra and $0.20 on Fable 5.2.
- GPT-6.2 Astra vs Claude Opus 5.5: Opus 5.5 costs half as much per token ($4 and $20). Opus 5.5 is the default model in Claude Code on paid plans.
- GPT-6.2 Sol vs Opus 5.5: Sol costs half of Opus 5.5 ($2 and $10), the same as Claude Sonnet 5.5.
- Context: 1.05M tokens on both GPT-6.2 models and 1M on Fable 5.2 and Opus 5.5. All four have a 128K max output.
- Independent indexes, as reported by VentureBeat: Artificial Analysis puts Astra second on its Intelligence Index (62), behind Fable 5.2. Epoch AI scores Astra and Fable 5.2 the same, at 174.
- Tools: GPT-6.2 runs in Codex and Claude models in Claude Code; see Claude Code vs Codex. Microsoft AI announced another coding model the same day: MAI Code 2.
Limits and safety
- Cybersecurity: both models meet OpenAI's Critical threshold for cybersecurity, and their tighter safeguards also restrict some legitimate security research. OpenAI points approved defenders to GPT-6 Astra, GPT-6 Cyber or Daybreak.
- Blockers: OpenAI says that in two tests both models were more likely to leave out blockers they hit, such as a missing file. It added Codex instructions that substantially reduce this. If you run them outside Codex, ask for blockers to be reported explicitly.
- Other ratings: both are at High for biology and chemistry, and Astra is OpenAI's first externally released model at High for AI self-improvement.
- Reasoning items: the API drops them when conversation history changes mid-conversation, and OpenAI says answer quality can decline.
- Early reception: VentureBeat reported negative early reviews from some power users and mixed ones from others, so test both models on your own tasks before switching.
Put GPT-6.2 on a standing job
Crewmark turns Codex and Claude Code sessions into AI employees with a role, a goal, working hours and a heartbeat that take issues from a board. A Codex employee uses whatever model its Codex session runs, so you can put GPT-6.2 Sol on everyday issues and Astra on the hard ones.
Pick a role
Such as the AI software engineer for a backlog of coding issues.
Set the model in its session
Run
/modelin the employee's Codex session and choose GPT-6.2 Sol or GPT-6.2 Astra.Let it run
It wakes on its heartbeat and works through the board; see Codex long-running tasks.
FAQ
When was GPT-6.2 released?
On November 17, 2026. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on the same day in ChatGPT Work, Codex and the API, and GitHub made both available in Copilot the same day.
How much does GPT-6.2 cost?
On the API, GPT-6.2 Astra costs $8 per million input tokens, $0.40 cached and $40 output. GPT-6.2 Sol costs $2, $0.10 and $10. Over 272K input tokens, input doubles and output is 1.5x ($16 / $60 for Astra, $4 / $15 for Sol). Batch is half price, Fast mode 2x and Ultrafast 3x.
What's the difference between GPT-6.2 Astra and Sol?
They have the same context window, output limit and tools. Astra is the more capable frontier model for hard projects. Sol is the efficient everyday model at a quarter of the price, and it's the default in new ChatGPT Work chats.
Is GPT-6.2 free? Which plans have it?
OpenAI's launch post lists Plus, Pro, Enterprise and Edu for ChatGPT Work and Codex. Free and Go aren't listed. In ChatGPT the models are in Work, not regular Chat. GitHub Copilot offers them on Pro+, Max, Business and Enterprise.
What is the context window of GPT-6.2?
1,050,000 tokens for both models, with a max input of 922,000 and up to 128,000 output tokens. In Codex, both take up to 1 million tokens when configured.
Is there a GPT-6.2 Luna?
No GPT-6.2 Luna has been announced; the small model is GPT-6 Luna. OpenAI's Codex docs still recommend GPT-6 Luna for focused, repeatable tasks.
Is GPT-6.2 Astra better than Claude Fable 5.2?
Both cost $8 and $40 per million tokens. OpenAI's charts put Astra ahead on some benchmarks, with caveats it states itself, and Artificial Analysis puts Fable 5.2 ahead on its Intelligence Index. See GPT-6.2 vs Claude for the full comparison.
How do I use GPT-6.2 in Codex?
Choose it in the Codex app's model picker, or run /model in the Codex CLI (version 0.169.1 or later). On a ChatGPT plan, you need Plus, Pro, Enterprise or Edu.
Checked against OpenAI's launch post, OpenAI API docs, ChatGPT/Codex docs and GitHub's changelog on November 18, 2026.