GPT-6.2 Astra and Sol Explained: OpenAI’s Cheaper Frontier Model — and Why the Cheaper One Matters More
Table of contents
- What OpenAI actually launched
- The price card, in plain numbers
- What the benchmarks say — and what they don’t
- The safety rating nobody should skip
- Worked examples: what a real job costs on each model
- Who gets access, and where
- The reception gap: OpenAI’s numbers vs independent tests
- Should you switch?
- Frequently asked questions
- Sources
OpenAI shipped two new models on Tuesday, November 17, 2026 — GPT-6.2 Astra and GPT-6.2 Sol — and the interesting thing about the launch is not the flagship. Astra is the new top model: 20% cheaper than the GPT-6 Astra it replaces and better on OpenAI’s professional benchmarks. But Sol, the cheaper model, is the one that changes the maths for most people. On the coding benchmark OpenAI chose to showcase, Sol trails Astra by half a point while costing about a fifth as much.
If you use ChatGPT for work, here is the practical summary up front. Both models are available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API. They are not in regular Chat. Astra costs $8 per million input tokens and $40 per million output tokens; Sol costs $2 and $10 — the same as GPT-6.1 Sol. If you mostly write, code, or run agents inside Work or Codex, you can switch today. If you only use the free tier or regular Chat, nothing changes for you yet.
What OpenAI actually launched
GPT-6.2 Astra is the new flagship, aimed at the hardest professional and scientific work: long documents, spreadsheets, legal deliverables, scientific terminal work, and agentic coding. OpenAI calls it “the world’s best model for scientists,” according to The Decoder’s coverage of the launch, and it is positioned as the model for work where the output has to meet professional conventions — a contract, a financial model, a research analysis — without heavy reworking afterwards.
GPT-6.2 Sol replaces GPT-6.1 Sol one tier down, at exactly the same price. OpenAI describes it as the everyday model: smarter and more efficient on multistep agent workflows, clearer writing, and better collaboration. The pitch is that you get more done on the same paid plan, because Sol finishes demanding tasks with fewer output tokens than its predecessor.
Both models keep the same technical frame as the 6.1 generation: a 1.05 million-token context window and up to 128,000 output tokens. The difference is what they do inside that frame, and what they cost. OpenAI’s framing for both is cost per finished task, not cost per token — a theme it has pushed all year, and one our cost-per-finished-task comparison tracks across the frontier models.
The price card, in plain numbers
| Model | Input / 1M tokens | Cached input / 1M | Output / 1M tokens | vs predecessor |
|---|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 | 20% below GPT-6 Astra |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 | Same as GPT-6.1 Sol |
| GPT-6 Astra (old) | $10.00 | $1.00 | $50.00 | — |
| GPT-6.1 Sol (old) | $2.00 | $0.10 | $10.00 | — |
Two details in that table are worth more attention than the headline 20%. First, the cached-input discount deepened to 60%: Astra’s cached input fell from $1.00 to $0.40 per million tokens. Cached input is text the system has already processed — the repeated instructions, documents and context that long agent sessions keep re-sending — so the saving lands hardest on exactly the long, multi-step workflows Astra is built for. Second, Astra’s new list price matches Claude Fable 5.2 exactly: Anthropic cut Fable’s price by 20% a week earlier, and the two frontier models are now priced identically, dollar for dollar. Our Claude Fable 5.2 explainer has the other side of that comparison.
There is also a pricing catch that only bites on very long jobs. For requests above 272,000 input tokens, OpenAI reprices the whole request: input and cache rates double, and output multiplies by 1.5. Astra then costs $16 per million input tokens and $60 per million output; Sol costs $4 and $15. If your workflow regularly sends whole codebases or long document sets in one request, that threshold — not the headline rate — is the number to remember. The worked examples below show what it does to a real bill.
What the benchmarks say — and what they don’t
All of the numbers in this section come from OpenAI’s own announcement unless marked independent. Vendor benchmarks are best read as a statement of where a lab aimed its model, not as a neutral ranking — but they still tell you what changed.
Agents’ Last Exam v1.1, a test of long, multistage professional tasks, is OpenAI’s headline result: Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost per task, and 5.5 points above Claude Fable 5.2 at about 5% of Fable’s cost. That second comparison carries a caveat worth knowing: Fable’s leaderboard result excludes nine tasks that used a fallback model, so the gap is not quite apples to apples.
Terminal-Bench-Science 0.1, which tests research carried out inside a terminal, shows Astra 9 points above GPT-6 Astra at less than a third of the cost per task. On Harvey LAB-AA v1.1, a legal test that checks whether a model can find the relevant facts across source documents and produce a deliverable that meets every requirement, Astra scores 15.0% and Sol 11.1% — OpenAI says Astra more than doubles Fable 5.2’s pass rate at about a sixth of the cost. On GDPval v2, Astra gains 6.2 points over GPT-6 Astra while using 22% fewer output tokens, and Sol gains 5.3 points over GPT-6.1 Sol.
The Sol numbers are the ones to linger on. On FrontierCode 2.0, which grades code edits on real projects against the bar maintainers set for merging, Sol comes within 0.5 points of Astra at about one-fifth of the cost, with both models at maximum reasoning. On LifeVal, OpenAI’s test built around everyday chores such as picking kindergartens after a relocation or drawing up meal plans, Sol beats GPT-6 Astra at about one-seventh of the cost. If those two results hold up in real use, Sol is not a “lite” model in any meaningful sense for coding and daily work — it is the same job for a fifth of the money.
The safety rating nobody should skip
Both models ship with a safety label that is easy to miss and important to understand. Under OpenAI’s Preparedness Framework, both GPT-6.2 Astra and Sol are rated Critical for cybersecurity capability — the highest tier — and OpenAI launched them with tighter safeguards on higher-risk cybersecurity assistance. The system card puts it in plain terms: Astra jumps 22 points over GPT-6 Astra on ExploitBench Internal Port, reaching a 53.5% arbitrary code execution rate, with Sol at 50.5%, against 31.5% for GPT-6 Astra. In other words, these models are substantially better at finding and exploiting software vulnerabilities than their predecessors, and OpenAI has tightened what they will help with in response.
There is a second first here. Astra is the first externally released model OpenAI rates High for AI self-improvement — the capability to help improve AI systems themselves. Sol does not meet that threshold. Both models are rated High, but not Critical, for biological and chemical capabilities. For most users, the practical effect is narrower than the labels sound: if you ask for help with a security task that looks like an attack, expect more refusals or more guardrails than you got from GPT-6 Astra. That is deliberate, and it is the price of the capability jump.
Worked examples: what a real job costs on each model
Token prices are abstract. Finished jobs are not. Here are four, using the official November 17 rates.
Example 1 — A standard writing and research job: 100,000 input tokens, 20,000 output tokens. On Astra, the input costs $0.80 (100,000 × $8 per million) and the output costs $0.80 (20,000 × $40 per million): $1.60 total. On GPT-6 Astra the same job cost $1.00 + $1.00 = $2.00, so the 6.2 upgrade saves $0.40, or 20%. On Sol, the same job costs $0.20 + $0.20 = $0.40 — a quarter of Astra’s price.
Example 2 — A cached-heavy agent session: 1 million cached input tokens, 50,000 output tokens. On Astra, cached input costs $0.40 and output costs $2.00: $2.40 total. On GPT-6 Astra, cached input cost $1.00 and output cost $2.50: $3.50. The saving is $1.10, or about 31% — bigger than the headline 20%, because the cache discount did the heavy lifting. On Sol the same session costs $0.10 + $0.50 = $0.60.
Example 3 — A month of standard jobs: 200 jobs like Example 1. Astra: 200 × $1.60 = $320. GPT-6 Astra: 200 × $2.00 = $400. Sol: 200 × $0.40 = $80. The monthly gap between Astra and Sol for the same volume of work is $240 — which is why the FrontierCode result matters so much. If Sol really does the coding job within half a point of Astra, that $240 is pure saving for most teams.
Example 4 — The long-context catch: 300,000 input tokens, 20,000 output tokens. This request crosses the 272,000-token repricing threshold, so Astra’s rates double on input and rise 1.5× on output. Input costs $4.80 (300,000 × $16 per million) and output costs $1.20 (20,000 × $60 per million): $6.00 total. Without the reprice, the same job would cost $2.40 + $0.80 = $3.20. The threshold nearly doubles the bill. If you send whole repositories or long document sets in single requests, split them or keep them under the threshold — or budget for the higher rate.
Who gets access, and where
Both models are available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol. Ultrafast mode is available at launch for both models: Pro subscribers at $500 a month get it with both, and Pro subscribers at $200 get it starting with Sol.
What is not included matters as much as what is. The release does not add either model to regular Chat. Free users and Plus users chatting in the standard interface will not see Astra or Sol as options — you have to be in Work or Codex, the modes built for professional and coding tasks. That split is now the defining line in ChatGPT: regular Chat is the consumer front door, and Work and Codex are where the new models land first. Our guide to the shared ChatGPT and Codex usage allowance explains how the December allowance change interacts with this, and our ChatGPT Wallet explainer covers the other new capability rolling out to Pro users this month.
Developers should note the API IDs and the context limits: 1.05 million tokens of context, up to 128,000 tokens of output, and the 272,000-token repricing threshold above. If you are migrating from GPT-6 Astra, the price cut applies immediately.
The reception gap: OpenAI’s numbers vs independent tests
The launch did not land the way OpenAI’s numbers suggest it should have. Early coverage described a bumpy reception, with reviewers cooler on the models than the benchmark tables imply — and the first independent measurements explain why.
Artificial Analysis, the independent benchmarker, scores Astra at 62 on its Intelligence Index — behind Claude Fable 5.2 alone, at 65, and just ahead of Gemini 4 Carbon at 60. Sol scores 57. On the Coding Agent Index, Astra came out flat against GPT-6 Astra — no improvement at all — and Sol fell 3 points against GPT-6.1 Sol, driven in part by a 14-point drop on a question-answering test. Artificial Analysis also measured hallucination rates of 59% for Astra and 55% for Sol on its AA-Omniscience test at maximum effort — high numbers that sit awkwardly next to the “best model for scientists” billing.
That gap — strong vendor benchmarks, flat-to-mixed independent results — is the honest picture of this release. Astra is cheaper and, on OpenAI’s own professional tests, better. On the independent tests that exist so far, it is a small step, not a leap, and Sol actually went backwards on one coding measure. The price cut is real and immediate; the capability jump is, for now, a claim that needs more independent testing.
Should you switch?
A short decision guide, by who you are:
- You code or run agents in Codex or ChatGPT Work on a Plus plan. Try Sol first, not Astra. If your work matches the FrontierCode result, you get nearly the same output for a fifth of the cost. Move to Astra only for the jobs where Sol’s output falls short — long legal or scientific deliverables, or tasks where the extra reasoning depth shows.
- You build on the GPT-6 Astra API. Switch to 6.2 Astra: same or better results, 20% cheaper, and a 60% cheaper cache rate that compounds on long sessions. Watch the 272,000-token threshold if your requests are large.
- You use regular Chat, or the free tier. Nothing changes. This release is not for you yet, and there is no announced date for a wider rollout.
- You care about the safety labels. Expect tighter refusals on higher-risk security help. If your work involves legitimate security testing, test the new guardrails before you migrate a workflow that depends on them.
- You are choosing between Astra and Fable 5.2. They now cost exactly the same per token. The choice is no longer about price — it is about which model’s independent results and workflow fit yours better, and Fable still leads on the independent Intelligence Index.
Checked against primary sources on November 22, 2026. Benchmark figures are as reported by OpenAI unless marked independent; see Sources below.
Frequently asked questions
GPT-6.2 Astra is OpenAI’s new flagship AI model, launched on November 17, 2026, for the hardest work — long documents, spreadsheets, legal deliverables and agentic coding. It costs $8 per million input tokens and $40 per million output tokens, 20% less than the GPT-6 Astra it replaces, and it is available in ChatGPT Work, Codex and the API.
GPT-6.2 Sol is the cheaper model launched alongside Astra on November 17, 2026, for agent workflows. It costs $2 per million input tokens and $10 per million output tokens — the same as GPT-6.1 Sol, and a quarter of Astra’s price. On the FrontierCode 2.0 coding benchmark, Sol trails Astra by 0.5 points at one-fifth of the cost.
No. GPT-6.2 Astra and Sol are available only in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the OpenAI API. The November 17 release did not add either model to regular Chat, so free users and Plus users in the standard chat interface will not see them as options.
GPT-6.2 Astra is 20% cheaper on input and output: $8 instead of $10 per million input tokens, and $40 instead of $50 per million output tokens. Cached input is 60% cheaper, falling from $1.00 to $0.40 per million tokens. On a standard job of 100,000 input tokens and 20,000 output tokens, Astra costs $1.60 against $2.00 for GPT-6 Astra.
It depends on whose tests you trust. On OpenAI’s benchmarks, Astra beats Fable 5.2 — by 5.5 points on Agents’ Last Exam at 5% of the cost — though that comparison excludes nine Fable fallback tasks. On the independent Artificial Analysis Intelligence Index, Fable 5.2 leads at 65 against Astra’s 62. The two models now cost the same per token.
Under OpenAI’s Preparedness Framework, both GPT-6.2 models are rated Critical for cybersecurity capability — the highest tier — because they are better at finding and exploiting vulnerabilities than their predecessors. They launched with tighter safeguards on higher-risk cybersecurity help, so expect more refusals on security tasks that look like attacks. Astra is the first model rated High for AI self-improvement.
Sources
- OpenAI, “GPT-6.2 Astra and Sol are now available” (Developer Community, November 17, 2026) — https://community.openai.com/t/gpt-6-2-astra-and-sol-are-now-available/1417531/1
- ExplainX, “GPT-6.2 Astra and Sol: $8/$40 and $2/$10 (Nov 17)” — https://explainx.ai/blog/openai-gpt-6-2-astra-sol-launch-pricing-benchmarks-november-2026
- BetaNews, “OpenAI GPT-6.2 Astra cuts API prices 20%, adds Sol model” — https://betanews.com/article/gpt-6-2-astra-api-pricing/
- CASRAI, “OpenAI GPT-6.2 Astra and Sol: what researchers get” — https://casrai.org/news/openai-gpt-6-2-astra-sol-release-2026
- ForkLog, “OpenAI launches GPT-6.2 Astra and Sol” — https://forklog.com/en/openai-launches-gpt-6-2-astra-and-sol/
- Elite Era Trends, “Claude Fable 5.2 Explained” — https://eliteeratrends.com/claude-fable-5-2-explained/
- Elite Era Trends, “Gemini 4 Neon Explained” — https://eliteeratrends.com/gemini-4-neon-explained/

Rana is the editor of Elite Era Trends, where he explains Medicare, Social Security and AI news for US readers in plain language — costs, deadlines and plan choices, checked against official sources.