GPT-6.2 Astra vs GPT-6.2 Sol: Price, Specs and Which to Use
OpenAI cut Astra's API price 20% and kept Sol flat. Here is the full price, benchmark and safety breakdown.

GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens on OpenAI's API — a 20% cut from GPT-6 Astra — while GPT-6.2 Sol holds steady at $2 input and $10 output, unchanged from GPT-6.1 Sol. Both models share a 1.05 million-token context window and a 128,000-token output limit. Astra is OpenAI's new frontier model "for your hardest projects"; Sol is the cheaper, faster model built for everyday work. Here's how the two actually compare on price, benchmarks and safety, and which one is worth using.
- GPT-6.2 Astra: $8/$40 per million input/output tokens (short context), 1.05M context window, 128K max output. OpenAI's frontier model, first to clear the "High" threshold for AI self-improvement under its Preparedness Framework.
- GPT-6.2 Sol: $2/$10 per million input/output tokens, same 1.05M context window and 128K output. Positioned as the efficient, everyday model; on Cognition's FrontierCode 2.0 it scores within 0.5 points of Astra at roughly a fifth of the cost.
- Availability: Both roll out immediately to ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers via the API as
gpt-6.2-astraandgpt-6.2-sol.
What is GPT-6.2 Astra?
GPT-6.2 Astra is OpenAI's new flagship model, announced November 17, 2026 as the successor to GPT-6 Astra. OpenAI positions it as the model for "your hardest projects" — complex coding, research and agentic tasks where raw capability matters more than cost. According to OpenAI's own launch post, Astra improved 6.2 percentage points on GDPval v2 (an internal evaluation of real-world professional tasks) over GPT-6 Astra, while using 22% fewer output tokens to get there. On Agents' Last Exam, OpenAI says Astra scores 5.2 points above GPT-6 Astra and 5.5 points above Anthropic's Claude Fable 5.2. On Terminal-Bench-Science 0.1, a benchmark for scientific command-line tasks, Astra is 9 points ahead of its predecessor.
Astra is also the first model OpenAI has released externally that clears the "High" threshold for AI self-improvement under the company's own Preparedness Framework, alongside the "Critical" threshold for cybersecurity capability it shares with Sol. OpenAI says that combination means both models ship with tighter safeguards for higher-risk cybersecurity assistance than earlier releases — a restriction OpenAI acknowledges can also get in the way of legitimate security research.
What is GPT-6.2 Sol?
GPT-6.2 Sol is the smaller, cheaper model in the same release, pitched by OpenAI as "efficient and versatile" for everyday work rather than frontier-level research. It improves 5.3 points over GPT-6.1 Sol on GDPval v2. The headline fact about Sol isn't its raw score, though — it's how close that score lands to Astra's for a fraction of the price. OpenAI cites Cognition's FrontierCode 2.0 benchmark, where Sol comes within 0.5 points of Astra while costing roughly one-fifth as much per task at matched settings. That's the core trade-off this generation: Astra for ceiling, Sol for value.
GPT-6.2 Astra vs Sol: price per million tokens
Both models use the same two-tier pricing structure OpenAI has used since GPT-6: a "short context" rate for requests up to 272,000 input tokens, and a higher "long context" rate above that. Cached input — tokens OpenAI has already processed in a recent request — is discounted on both tiers.
| Pricing (per 1M tokens) | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Input (≤272K tokens) | $8.00 | $2.00 |
| Cached input (≤272K tokens) | $0.40 | $0.10 |
| Output (≤272K tokens) | $40.00 | $10.00 |
| Input (>272K tokens) | $16.00 | $4.00 |
| Output (>272K tokens) | $60.00 | $15.00 |
| Context window | 1.05M tokens | 1.05M tokens |
| Max output | 128K tokens | 128K tokens |
Astra's price cut is new this generation: OpenAI dropped both its input and output rates by 20% compared with GPT-6 Astra, and widened the cached-input discount to 60% off the standard input rate. Sol didn't get a price change at all — it carries over GPT-6.1 Sol's $2/$10 rate exactly, which means the gap between the two models is still a flat 4x on input and output pricing, same as last generation.
How GPT-6.2 fits into OpenAI's model lineup
OpenAI has now shipped three ".2"-style refreshes of its GPT-6 family inside roughly a year: GPT-6 Astra and Sol launched first, then GPT-6.1 brought Terra into the lineup as a mid-tier option between the two, and now GPT-6.2 replaces the original Astra and Sol outright rather than adding a fourth tier. That matters for anyone with code or workflows hard-coded to a specific model ID: GPT-6 Astra and GPT-6.1 Sol still work today, but OpenAI's pricing and benchmark pages make clear the newer IDs are meant to be the default going forward, not a parallel option. For a closer look at what changed in the launch itself — the exact rollout timeline, who got access first and the context-window numbers before third-party verification — see our GPT-6.2 Astra and Sol launch rundown.
How Astra and Sol stack up against Claude Fable 5.2
OpenAI's launch post leans heavily on comparisons to Anthropic's Claude Fable 5.2, which is telling about who OpenAI sees as the benchmark to beat this cycle. On Agents' Last Exam, OpenAI says Astra scores 5.5 points above Fable 5.2. On Harvey's LAB-AA legal-reasoning benchmark, OpenAI claims Astra more than doubles Fable 5.2's pass rate. But the independent Artificial Analysis AA-3D human-preference ranking tells a tighter story: Astra leads at 1,318 Elo, but Fable 5.2 sits second at 1,271 — ahead of OpenAI's own Sol, which trails at 1,259. In other words, the model people say they prefer talking to doesn't perfectly track the benchmarks OpenAI chose to headline. If you're deciding between Anthropic's and OpenAI's current flagships rather than between OpenAI's own two models, our full Claude Opus 5.5 vs GPT-6.2 Astra and Sol comparison covers that match-up in price and spec detail.
Benchmark performance: where each model leads
OpenAI's own figures, cross-checked against independent evaluators, point to a consistent pattern: Astra leads on frontier reasoning and agentic tasks, while Sol trails by a small, usually single-digit margin at a much lower cost.
- GDPval v2 (OpenAI internal): Astra +6.2 points over GPT-6 Astra; Sol +5.3 points over GPT-6.1 Sol.
- Agents' Last Exam: Astra scores 5.2 points above GPT-6 Astra and 5.5 points above Claude Fable 5.2, per OpenAI.
- Harvey LAB-AA (legal reasoning): OpenAI says Astra more than doubles Claude Fable 5.2's pass rate.
- Artificial Analysis AA-3D (human preference): independent rankings place GPT-6.2 Astra first at 1,318 Elo, Claude Fable 5.2 second at 1,271, and GPT-6.2 Sol third at 1,259 — meaning Sol's human-preference score is closer to Astra's than its price would suggest.
- Cognition FrontierCode 2.0: Sol lands within 0.5 points of Astra at roughly a fifth of the cost, OpenAI says, which is the strongest evidence that Sol is the better default for routine coding work.
Independent coverage is less uniformly positive than OpenAI's launch post. Decrypt reports that in some third-party tests, Astra ranks second behind Claude Fable 5.2 rather than first, and that neither new model shows a meaningful jump in coding-agent performance specifically — a reminder that vendor benchmarks and independent agent evaluations don't always agree, and that OpenAI's own numbers above are the company's figures, not third-party verified results.
Safety thresholds: what changed under the hood
Both GPT-6.2 Astra and Sol meet the "Critical" threshold for cybersecurity capability under OpenAI's Preparedness Framework, the same threshold GPT-6 Astra met. What's new is that Astra is the first model OpenAI has released externally to also meet the "High" threshold for AI self-improvement — a capability tier the company tracks because a model that can meaningfully speed up its own development carries different risks than one that simply performs well on benchmarks. Both models also meet the "High" threshold for biological and chemical capabilities. OpenAI says the practical result is tighter safeguards around higher-risk cybersecurity assistance for both models, which it acknowledges can occasionally complicate legitimate security research workflows that rely on the API.
Availability: ChatGPT plans and API access
Both models are live now. In ChatGPT, Astra and Sol are available to Plus, Pro, Enterprise and Edu subscribers inside ChatGPT Work and Codex. OpenAI's fastest inference tier, Ultrafast, is available for both models at launch, though access depends on plan: Pro 500 subscribers get Ultrafast with either model, while Pro 200 subscribers get it only with Sol. Developers reach both models through the API using the model IDs gpt-6.2-astra and gpt-6.2-sol.
For teams already calling gpt-6-astra or gpt-6.1-sol in production, the switch is a model-ID swap rather than an API redesign — OpenAI hasn't changed the request or response schema between generations, and the pricing tiers follow the same short-context/long-context split that's been in place since GPT-6. The practical migration risk isn't compatibility, it's cost drift: Astra's prompts got 20% cheaper, so a workload that was previously tuned to stay under a budget on GPT-6 Astra now has more headroom, while Sol's price held flat, so nothing there needs re-budgeting at all.
Which one should you use?
If the job is a genuinely hard coding, research or agentic task where getting the answer right matters more than the bill — large codebase refactors, scientific analysis, multi-step agent workflows — Astra's frontier-level scores and self-improvement capability make it the safer pick, even at 4x Sol's price. If the job is everyday drafting, summarizing, routine coding or anything run at high volume, Sol's FrontierCode 2.0 score within half a point of Astra, at roughly a fifth of the cost, makes it the more rational default. Teams running agents at scale will likely want both models wired up and routed by task difficulty rather than picking one for everything, which is exactly the two-tier structure OpenAI built this release around.
- Pick Astra if: you're running agentic coding pipelines, scientific or legal research, or any workflow where OpenAI's safety-threshold language ("Critical" cybersecurity, "High" self-improvement) signals real capability gains worth paying 4x for.
- Pick Sol if: you're building a high-volume product feature, a customer-facing chatbot, or routine code generation where FrontierCode 2.0's half-point gap to Astra doesn't justify quadrupling your per-token cost.
- Pick both if: you're building an agent system with a router — use Sol as the default and escalate to Astra only for tasks that fail a confidence check, which is the pattern OpenAI's own pricing tiers are clearly designed to encourage.
One practical wrinkle worth budgeting for: the "long context" pricing tier kicks in automatically once a request crosses 272,000 input tokens, doubling Astra's input rate to $16 and Sol's to $4. For workflows that regularly feed large documents, codebases or multi-turn histories into the model, that threshold — not the headline $8/$2 rate — is often the number that actually determines your bill.
What next
Independent, third-party benchmark testing of both models is still rolling in beyond OpenAI's own launch figures, and the gap between vendor-reported and independently verified scores — already visible in Decrypt's reporting on Astra's ranking — is worth watching before committing a production workload to either model. For the authoritative, current specification and pricing tables, OpenAI's own GPT-6.2 Astra and Sol announcement and its API pricing page remain the sources to check before building against either model.
Frequently asked questions
How much does GPT-6.2 Astra cost compared to GPT-6.2 Sol?
Astra costs $8 per million input tokens and $40 per million output tokens (short context, up to 272K tokens). Sol costs $2 input and $10 output — a flat 4x cheaper than Astra on both ends.
Did GPT-6.2 Astra get cheaper than GPT-6 Astra?
Yes. OpenAI cut Astra's input and output prices by 20% versus GPT-6 Astra, and widened the cached-input discount to 60% off the standard rate.
What is the context window for GPT-6.2 Astra and Sol?
Both models share a 1.05 million-token context window and a 128,000-token maximum output, according to OpenAI's API model documentation.
Is GPT-6.2 Sol almost as good as Astra?
On Cognition's FrontierCode 2.0 benchmark, OpenAI says Sol scores within 0.5 points of Astra at roughly one-fifth the cost. On OpenAI's harder internal benchmarks like Agents' Last Exam, Astra leads by a larger margin.
What's new about GPT-6.2's safety classification?
GPT-6.2 Astra is the first model OpenAI has released externally to meet the 'High' threshold for AI self-improvement under its Preparedness Framework. Both Astra and Sol also meet the 'Critical' threshold for cybersecurity and the 'High' threshold for biological and chemical capabilities.
How do I access GPT-6.2 Astra and Sol?
In ChatGPT, both are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Developers can call them via the API using the model IDs gpt-6.2-astra and gpt-6.2-sol.
Sources
- OpenAI — Introducing GPT-6.2 Astra and Solopenai.com
- OpenAI API Pricingdevelopers.openai.com
- OpenAI API Models documentationdevelopers.openai.com
Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.


