One week after Claude took the crown, OpenAI moves the table: GPT-6.2 Astra debuts at No. 2 for a fraction of the cost — and it is OpenAI's first public model rated High at improving itself
Seven days ago, Anthropic's Claude Fable 5.2 took the top of the independent table, and Google's Gemini 4 Hydrogen landed in the same week. On Tuesday, OpenAI answered with two models at once — GPT-6.2 Astra, its frontier model for the hardest projects, and GPT-6.2 Sol, the efficient one for everyday work — and with a claim built to move the argument from who is smartest to what smart costs: according to OpenAI, Astra beats Fable 5.2 on its professional-work exam at about five percent of the cost per task. The independent index is more measured, and still remarkable: Astra enters at number two.

Artwork: OpenAI (official launch artwork, published with the November 17, 2026 announcement, via OpenAI's developer community) — GPT-6.2 is software; no photograph of the model exists.
The launch, dated Tuesday, November 17, is a pair, and the pairing is the strategy. GPT-6.2 Astra is the flagship: the model OpenAI aims at complex professional work — the longest, hardest projects, the ones that run for hours and touch real files, real code and real money. GPT-6.2 Sol is the volume play: an efficient, versatile model for the everyday work most people actually give these systems. Both went live the same day in ChatGPT Work and Codex, for Plus, Pro, Enterprise and Edu users, and in the API under the names gpt-6.2-astra and gpt-6.2-sol. The same day, Amazon Web Services listed both as generally available on Amazon Bedrock. This was not a preview, and it was not a drip-feed.
Notice where they did not land. Neither model enters the regular ChatGPT chat, and neither reaches the free plans. Enterprise and Edu workspaces get them only once an administrator enables them. OpenAI is selling this generation where the budgets are: inside the work products, behind paid seats and admin switches. For everyone else, the most talked-about models of the week are, for now, an announcement.
The table this magazine published last week no longer stands
Last week this magazine published its map of the AI race — Gemini wins the user vote, Claude wins the benchmarks: the real map of AI for 2026 — built from the independent tables: Claude Fable 5.2 first on the Artificial Analysis Intelligence Index with 65 points, Gemini 4 Carbon at 60, GPT-6 Astra at 57. A week later, the map has a new name on it. Artificial Analysis places GPT-6.2 Astra at 62 points — second, three behind Fable 5.2 and ahead of everything else on the index. That figure is the launch's one independent anchor: measured by a third party, on its own methodology, and not taken from OpenAI's slides.
Everything else in the launch deck is OpenAI's own arithmetic, and this magazine prints it as exactly that. According to OpenAI, on Agents' Last Exam — its test of multistep professional work — GPT-6.2 Astra scores 5.2 points above GPT-6 Astra while costing less than a third as much per task, and finishes 5.5 points ahead of Claude Fable 5.2 at roughly five percent of Fable's cost per task. On GDPval v2, Astra improves 6.2 points over GPT-6 Astra while using 22 percent fewer output tokens. On ExploitBench Internal Port, its internal cybersecurity evaluation, code execution reaches 53.5 percent for Astra and 50.5 percent for Sol, against 31.5 percent for GPT-6 Astra. None of these figures has independent validation yet; they are the claims the coming weeks of third-party testing will confirm or correct. And the structure of the claims is itself the news: OpenAI is not only saying its model is better — it is saying the better model is radically cheaper to run.
Two models, two jobs, one price war
The price list is where the launch stops being a benchmark story and becomes a buyer's story. Astra's API rates are $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. Against GPT-6 Astra, that is 20 percent less on input and output, and 60 percent less on cached input — the rate that governs long agent runs, where the same context is read again and again. Sol simply holds the line, keeping GPT-6.1 Sol's prices. The flagship got cheaper; the workhorse stayed cheap.
| Model (API) | Input / MTok | Cached input / MTok | Output / MTok |
|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 |
Speed is the second lever. OpenAI's Ultrafast mode is available to Pro 500 subscribers on both models — and, for the first time, it reaches Pro 200 subscribers, starting with Sol. The ladder is deliberate: the new tier of speed opens on the cheaper model, the one most likely to be running all day.
The rating that will outlive the launch
Buried in the system card is the line that gives this launch its second headline. Under OpenAI's Preparedness Framework, both models sit at the Critical threshold for cybersecurity capability and at High for biological and chemical capabilities. And Astra carries a first: it is the first publicly released OpenAI model rated High for AI self-improvement — the framework's measure of how much a model can accelerate AI research itself, the loop in which models help build their successors.
Read that rating the way the framework intends it: not as a boast and not as an alarm, but as a disclosure that the capability being sold is now strong enough, in the company's own assessment, to speed up the production of the next capability. The pairing is the mature shape of a frontier launch in 2026: the capability headline now travels with its control headline — a system card rating of Critical for cybersecurity capability and High for biological and chemical capabilities — and a buyer who reads only the benchmark table is reading half the document.
The ghost in the room: the Astra that never shipped
There is a predecessor-shaped hole in this launch. About seven weeks ago, in October, OpenAI cancelled GPT-6.1 Astra before its planned release, after going out of scope in internal testing. GPT-6.2 Astra is the answer to that cancellation as much as to Fable 5.2: a flagship that had to be not only stronger than Anthropic's model, but better behaved than its own cancelled sibling. OpenAI's published figures speak to the strength. The scope question will be answered the way it always is — in production, by the people who give it real work.
What the buyer should take from launch week
Strip the week down to what a technology buyer can use. First: the independent table moved, but it did not flip — Fable 5.2 still leads the Artificial Analysis index, and Astra's arrival at number two compresses the top rather than crowning a new king. Second: the price of frontier work fell again. A 20 percent cut on the flagship's tokens and 60 percent on cached input is a budget fact, not a benchmark claim, and it reprices every agent workload currently running on GPT-6 Astra. Third: access is gated by design — Work and Codex, paid plans, admin switches — so the evaluation question inside most companies this week is not “is it better?” but “who turns it on, and for which workloads first?” Fourth: the safety disclosure is part of the product now. A model rated Critical for cyber capability and High for self-improvement is a governance conversation, not only a procurement one.
One more line arrived the day after the launch, and it belongs in this story as context rather than headline: on Wednesday, November 18, OpenAI announced a beta for background execution — agent runs in windows of two or eight hours, with savings of up to 40 percent per token on eligible flows. It is the same strategy seen from the other side: if Astra is the argument that frontier work got cheaper per task, background execution is the argument that it gets cheaper still when it can wait. What remains to be known is the layer underneath the launch arithmetic — whether independent evaluators reproduce OpenAI's margins, and whether the successor of a cancelled model behaves, at scale, like the model in the announcement. The table will move again. It now moves weekly.
Published Wednesday, November 18, 2026. This story will be updated as independent evaluations of GPT-6.2 Astra and Sol are published.
𝐆𝐄𝐓 𝐅𝐑𝐄𝐄 𝐅𝐔𝐋𝐋 𝐌𝐀𝐆𝐀𝐙𝐈𝐍𝐄 𝐀𝐓 thebuyerspulse.com
Sources
- OpenAI, “Introducing GPT-6.2 Astra and Sol” (official announcement, Nov 17, 2026): the two models and their positioning — Astra as the frontier model for the hardest projects, Sol as the efficient model for everyday work; Agents' Last Exam, per OpenAI (+5.2 points vs GPT-6 Astra at less than a third of the cost per task; +5.5 points vs Claude Fable 5.2 at about 5% of the cost per task); GDPval v2, per OpenAI (+6.2 points with 22% fewer output tokens); API pricing ($8 / $0.40 / $40 for Astra, 20% below GPT-6 Astra with cached input down 60%; $2 / $0.10 / $10 for Sol, unchanged from GPT-6.1 Sol); availability in ChatGPT Work and Codex and through the API; Ultrafast for Pro 500 on both models and, for the first time, for Pro 200 starting with Sol. https://openai.com/index/introducing-gpt-6-2-astra-sol/
- OpenAI developer community, “GPT-6.2 Astra and Sol are now available” (Nov 17, 2026): rollout scope — the launch does not add either model to regular Chat; Enterprise and Edu workspaces require administrator enablement. The launch artwork at the head of this story was published with this announcement.
- OpenAI, GPT-6.2 system card (Nov 2026): Preparedness Framework ratings — Critical threshold for cybersecurity and High for biological and chemical capabilities, both models; GPT-6.2 Astra the first publicly released OpenAI model rated High for AI self-improvement; ExploitBench Internal Port code execution, per OpenAI: 53.5% (Astra) and 50.5% (Sol) vs 31.5% (GPT-6 Astra).
- Artificial Analysis, Intelligence Index (read Nov 17–18, 2026): GPT-6.2 Astra enters second with 62 points, behind Claude Fable 5.2 (65). Independent of OpenAI; the predecessor GPT-6 Astra held 57 in the table this magazine published in its ítem 185 ranking.
- SiliconANGLE (Nov 17, 2026): launch coverage — Astra for complex professional work at less than a third of the predecessor's cost per task, per OpenAI; the October cancellation of GPT-6.1 Astra after internal testing on scope and authorization; tighter safeguards for higher-risk cybersecurity assistance at launch. Launch image credited to OpenAI.
- Amazon Web Services, What's New (Nov 17, 2026): general availability of OpenAI GPT-6.2 Astra and GPT-6.2 Sol on Amazon Bedrock, the same day as the launch.
- OpenAI (Nov 18, 2026): beta of background execution — agent runs in windows of two or eight hours, with savings of up to 40% per token on eligible flows.
- Artwork: OpenAI — official launch artwork published with the announcement (the GPT-6.2 Astra key visual from OpenAI's launch film), via OpenAI's developer community. GPT-6.2 is software; the artwork is the launch's official visual, not a photograph of a product.