GPT-6.2 Astra & Sol: $8/$40, $2/$10, Benchmarks, Migration
GPT-6.2 Astra and Sol, announced Nov 17, 2026: Astra costs $8/$40 per 1M tokens, Sol $2/$10. Specs, 1.05M context, benchmarks, Fable 5.2 comparison, migration
GPT-6.2 Astra and GPT-6.2 Sol are two new models OpenAI announced on November 17, 2026. Astra is the flagship for the hardest projects, priced at $8 input / $40 output per 1M tokens, 20% below GPT-6 Astra ($10/$50). Sol is the everyday model at $2 input / $10 output, keeping the price of GPT-6.1 Sol while improving performance. The API model IDs are gpt-6.2-astra and gpt-6.2-sol.
This article relies on OpenAI's API model pages and post-launch coverage (ITdaily, Tech Monitor, 9to5Mac). OpenAI's own announcement page could not be fetched, so launch details were checked through the OpenAI Developer Community post and press reports. Almost all benchmark figures are OpenAI-reported, not independently reproduced.
Where Astra and Sol fit
- GPT-6.2 Astra: the flagship for the hardest projects, with gains in professional work (slides, spreadsheets, documents), scientific research, coding, websites and 3D design, per OpenAI
- GPT-6.2 Sol: an efficient model for everyday work and multistep agentic workflows, described as smarter than GPT-6.1 Sol with more natural writing
- OpenAI says both use fewer output tokens on demanding tasks
- Astra can reportedly use a slide screenshot as a template and match its layout, colors and font size (9to5Mac)
The lineage runs from GPT-6 Astra (September) and GPT-6.1 Sol (DevDay) to this simultaneous 6.2 update. GPT-6.1 Terra ($1 input) is recommended for prose and fiction, according to a moderator reply in the OpenAI Developer Community.
Specs
The following is from OpenAI's API model pages. The two models share most specs.
| Item | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Model ID | gpt-6.2-astra | gpt-6.2-sol |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Max input | 922,000 tokens | 922,000 tokens |
| Max output | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | 2026-05-18 | 2026-06-01 |
| Reasoning effort | none is rejected | none is rejected |
| Endpoints | Chat Completions / Responses / Batch | Chat Completions / Responses / Batch |
| Not supported | Realtime, Assistants, fine-tuning, image, video, audio | Same |
The API rejects requests that set reasoning effort to none. The model pages I could access did not list the other accepted values, so check the current docs before implementing.
Pricing
Standard rates per 1M tokens (short context).
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
| GPT-6 Astra (previous) | $10 | — | $50 |
| GPT-6.1 Sol (previous) | $2 | $0.10 | $10 |
| Claude Fable 5.2 | $8 | — | $40 |
Astra's cached input is 95% off its input rate and, per press reports, 60% cheaper than GPT-6 Astra's. Cache writes cost $10 (30-minute) or $20 (12-hour) for Astra, and $2.50 or $5 for Sol.
- Long context: above 272K input tokens, input and cache rates double and output costs 1.5x for the whole request
- Batch: 50% of Standard
- Fast mode: 2x Standard
- Ultrafast: 3x Standard through February 28, 2027, then 6x from March 1, 2027 (Pacific time)
A simple estimate for 100M input and 20M output tokens a month (no caching or long-context surcharge) gives $1,600 for Astra and $400 for Sol. GPT-6 Astra would cost $2,000 on the same workload.
Benchmarks
All figures below are OpenAI-reported or quoted by the press unless noted.
| Evaluation | Result | Source type |
|---|---|---|
| Agents' Last Exam | Astra (xhigh) scores 5.2 points above GPT-6 Astra (max) at less than half the cost per task (press: under a third) | OpenAI |
| Agents' Last Exam vs Claude Fable 5.2 | Astra scores 5.5 points higher at 5% of the cost; OpenAI excluded nine tasks where a fallback model was used for Fable 5.2 | OpenAI (ITdaily) |
| Terminal-Bench-Science 0.1 | Astra is 9 points above GPT-6 Astra at under a third of the cost | OpenAI |
| FrontierCode 2.0 (Cognition) | Sol within 0.5 points of Astra at max reasoning, about one-fifth the cost | OpenAI |
| ExploitBench Internal Port | Up 22 points over GPT-6 Astra | OpenAI |
| LifeVal | Sol beats GPT-6 Astra at about one-seventh the cost | OpenAI |
| AA-AnalystAgent (Artificial Analysis) | Astra answers 62.5% of complex financial-report questions correctly across five attempts | Third party (via press) |
Sol's pitch shows clearly in FrontierCode 2.0: within half a point of Astra at roughly a fifth of the cost per task. Coverage also notes that independent testing is still needed to confirm the cost reductions.
How to use it
The API supports Chat Completions, Responses and Batch. Switching is a model ID change. Here is a minimal Responses API example.
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6.2-sol", # flagship: "gpt-6.2-astra"
reasoning={"effort": "medium"}, # "none" is rejected
input="Summarize the decisions in these minutes in three bullets.",
)
print(resp.output_text)In ChatGPT, both models are available in ChatGPT Work and Codex (Plus / Pro / Enterprise / Edu), not in regular Chat, according to a reply in the OpenAI Developer Community. Availability depends on plan, workspace permissions and rollout progress, and some users report not seeing them yet. Update the desktop app and the Codex CLI.
- Plus / Pro / Enterprise / Edu: ChatGPT Work and Codex
- Pro ($200): Ultrafast with Sol only (the first time Ultrafast reaches Pro 200)
- Pro ($500): Ultrafast with both models
- Enterprise / Edu: Ultrafast once an admin enables it
- Ultrafast speed: up to 4.5x the standard token rate (9to5Mac), using more of your allowance
Versus Claude Fable 5.2 and others
Anthropic's Claude Fable 5.2 is priced at $8/$40, the same as GPT-6.2 Astra. OpenAI claims Astra beats Fable 5.2 by 5.5 points on Agents' Last Exam at 5% of the cost, but that is OpenAI's own comparison, and since list prices match, the cost gap reflects token consumption per task. Validate on your own data.
| Aspect | GPT-6.2 Astra | GPT-6.2 Sol | Claude Fable 5.2 |
|---|---|---|---|
| Input/output (1M) | $8 / $40 | $2 / $10 | $8 / $40 |
| Positioning | Hardest tasks | Everyday and agentic | Anthropic's top model |
| Context window | 1,050,000 | 1,050,000 | Not verified here |
| Basis | OpenAI | OpenAI | Our earlier article (pricing only) |
Safety classification
According to Tech Monitor, OpenAI classifies both models as meeting the Critical threshold for cybersecurity and the High threshold for biological/chemical capability and AI self-improvement under its Preparedness Framework, with added safeguards (the cyber classification matches other coverage). Outputs are monitored and dual-use cases restricted. OpenAI says it is working to offer the models to approved defenders with fewer refusals through Daybreak, so security-related prompts may be refused more often than usual.
Migration notes from GPT-6.1
- Change the model ID: gpt-6.1-sol to gpt-6.2-sol; try gpt-6.2-astra only for work that needs the flagship
- Sol price unchanged: $2/$10, so cost estimates hold
- Reasoning effort: none is rejected; change any request that sets it
- Above 272K tokens: measure how often you cross the long-context surcharge boundary
- Output tokens: OpenAI says they fall; measure on your workload
- Ultrafast pricing: moves from 3x to 6x on March 1, 2027
- Deprecations: none found in the sources I could access
As in the GPT-6.1 Sol migration guide, run old and new side by side and compare accuracy and token usage on your own data.
Choosing between them
| Use case | Candidate | Why |
|---|---|---|
| Everyday coding, documents, agents | GPT-6.2 Sol | $2/$10 and within 0.5 points of Astra on FrontierCode 2.0 |
| Professional work (decks, legal, finance, science) | GPT-6.2 Astra | Positioned by OpenAI for the hardest work |
| High-volume routine jobs | Add Batch (50%) | Half of Standard |
| Latency-critical | Ultrafast | Faster, at 3x to 6x the rate |
FAQ
How much do GPT-6.2 Astra and Sol cost in the API?
Per 1M tokens, per OpenAI's model pages: Astra is $8 input, $0.40 cached input and $40 output; Sol is $2, $0.10 and $10. Astra is 20% cheaper than GPT-6 Astra; Sol matches GPT-6.1 Sol. Above 272K input tokens, input rates double and output is 1.5x.
What is the context window?
Both models list a 1,050,000-token context window, 922,000 max input and 128,000 max output on OpenAI's model pages, with the long-context surcharge beyond 272K input tokens.
How do I choose between Astra and Sol?
Astra targets the hardest professional, scientific and coding work; Sol targets everyday and agentic work. On FrontierCode 2.0 Sol comes within 0.5 points of Astra at about a fifth of the cost, so start with Sol and move up only where it falls short.
Can I use them in regular ChatGPT?
A reply in the OpenAI Developer Community says they are not added to regular Chat; they are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, subject to plan, permissions and rollout.
What should I check when migrating from GPT-6.1 Sol?
Change the model ID, remove any reasoning effort of none (it is rejected), review long-context requests over 272K tokens, and note Ultrafast moving to 6x pricing on March 1, 2027. I found no deprecation notice in the sources I could access.
Can I trust the benchmark numbers?
Most are OpenAI's own. Artificial Analysis' AA-AnalystAgent result (62.5% for Astra) is third-party, but coverage says independent verification of the cost claims is still needed. Test on your own data.
株式会社オブライト
Oflight Inc. is a Tokyo-based IT company providing software development, digital signage, web development, network systems, and IoT/AR/VR technology services.
Related free tools (no sign-up, instant results)
Feel free to contact us
Contact Us