Cheaper per Task, Harder to Watch

OpenAI's GPT-6.2 Astra ties Claude Fable 5.2 on Epoch's index and trails it elsewhere, at a much lower cost per task. Its system card shows a model that omits known blockers twice as often as GPT-6 Astra and slips past some monitors. What changed, and what to do before you switch.

A glowing machine core in a glass cube, half clear and half opaque, under a spotlight, with warning panels to the right and a shrinking price tag and coins to the left.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol today, November 17. Astra is the frontier model and Sol the everyday one. Both are in the API and in ChatGPT Work and Codex; OpenAI's help center is explicit that "this rollout does not add either model to regular Chat."

The release has two halves. The commercial half is good: Astra is 20% cheaper per token than GPT-6 Astra, both models, OpenAI says, spend fewer tokens per task, and Astra lands near the top of the independent indexes. The other half is in the system card. Astra is the first model OpenAI has deployed externally at its High threshold for AI self-improvement, and on several measures it is harder to supervise than the model it replaces.

What shipped

ModelPer 1M tokens; cutoff
GPT-6.2 Astra$8 in, $0.40 cached, $40 out; May 18, 2026
GPT-6 Astra (before)$10 in, $1.00 cached, $50 out; April 30, 2026
GPT-6.2 Sol$2 in, $0.10 cached, $10 out; June 1, 2026
GPT-6.1 Sol (before)$2 in, $0.10 cached, $10 out; April 30, 2026

Prices, knowledge cutoffs and limits come from the Astra and Sol model pages. Both new models take text and images and return text, with a 1,050,000-token context window, 922,000 input tokens at most and 128,000 output tokens. Above 272K input tokens, input and cache rates double and output costs 1.5 times as much, for the whole request. Ultrafast costs three times the Standard rate through February 28, 2027, then six times.

A 60% cut on cached input can matter more to a long agent session, which re-reads its context every turn, than the 20% cut on fresh tokens. Simon Willison called it "great news for longer running agentic sessions" and noted that $8/$40 is "the same price as Claude Fable 5.2." Sol's per-token price didn't move, so any saving on Sol comes from using fewer tokens.

OpenAI's capability claims are its own. It says "both models deliver better results on demanding tasks with fewer output tokens," and that Astra's "more precise computer use lets it work down to individual parts and connections in CAD and simulation tools."

There's a gap in the version numbers. GPT-6 Astra shipped on September 3, GPT-6 Sol and Luna on September 22, GPT-6.1 Sol on September 29 and GPT-6.1 Terra on October 26, per the API changelog. OpenAI never released a GPT-6.1 Astra. CNBC reported on September 28 that OpenAI decided not to release it because it "didn't quite meet the bar in terms of staying within scope and authorization," in the words of its head of safety systems, Saachi Jain.

Where it actually ranks

OpenAI's announcement leads with its own charts. The independent numbers, as read on November 17, tell a narrower story.

IndexWhere GPT-6.2 lands
Artificial Analysis Intelligence IndexAstra second, behind Claude Fable 5.2 and ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59). Sol at 57, level with GPT-6 Astra at about 15% of its cost per task
Artificial Analysis Coding Agent IndexAstra flat against GPT-6 Astra; Sol down 3 points against GPT-6.1 Sol
Epoch Capabilities IndexAstra 174.17 (170.29–178.58), Fable 5.2 173.5 (170.09–177.36): overlapping confidence intervals, a statistical tie
Vals IndexFable 5.2 first at 71.31%; Astra 7th at 67.66%, behind five Anthropic and Google models besides Fable; Sol 9th at 66.06%

Artificial Analysis traced Sol's coding-agent drop mainly to a 14-point fall on SWE-Atlas-QnA-AA.

Cost per task is where 6.2 clearly wins. On Vals, Astra costs $10.68 per test against $21.28 for GPT-6 Astra and $17.94 for Fable 5.2; Sol costs $3.18. On Hacker News, a Plotly team member reported that on the company's data-analytics exam Astra's cost per task fell from about $20 to about $4, and Sol's from $2.60 to $1.20, with "similar accuracy." The comment credits fewer tokens and fewer tool calls.

One result no index captures: in a pre-release test reported in OpenAI's system card, with a $300 budget per question, Epoch found that Astra "produced lean solutions to 7/68 of the problems" in its FrontierMath ErdΕ‘s set, "including one for which no prior solution was known." The previous record was 3.

Read the footnotes on OpenAI's charts

OpenAI's own benchmarks are vendor claims, and a few details deserve a second look:

  • On Agents' Last Exam, Astra leads Fable 5.2 by 5.5 points "at just 5% of the cost," but the footnote says Fable's results "exclude nine tasks that used a fallback model."
  • The gain on that benchmark over GPT-6 Astra comes "at less than a third of the cost" in the announcement and "at less than half the cost per task" in OpenAI's developer-forum post (which compares Astra at xhigh with GPT-6 Astra at max).
  • On Terminal-Bench-Science, Fable 5.2 runs in a "mini-swe-agent harness," not Anthropic's own.
  • On FrontierCode 2.0, the highest point on OpenAI's own chart belongs to Claude Fable 5.2.

The system card

What "High" means, and the same-day edit

Both models meet OpenAI's Critical threshold for cybersecurity and High for biological and chemical risk. Astra adds a first: it is "the first model we are deploying externally that meets the High threshold for AI Self-Improvement." Sol stays below it. The announcement translates High as "equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline."

The card was edited on the day it was published. Its change log quotes the original sentence: "The High threshold is an early milestone for preparing safeguards ahead of fully automated, recursive self-improvement." The new version calls it "an early milestone for proactively preparing safeguards for any models that may in future be capable of fully automated, recursive self-improvement." The first sentence reads as if that step is on the way. The second makes it a hypothetical.

On offensive cyber, the results point both ways. On Irregular's FrontierCyber suite, Astra solved 68 of 217 challenges, fewer than GPT-6 Astra's 77, yet GPT-6.2 "successfully found and exploited zero-day vulnerabilities in various widely used targets."

What improved

In simulated workplace tasks, both models "took improper actions about half as often as their respective predecessors." For anyone running agents with real permissions, that's a meaningful gain.

It reports its blockers less often

This is the finding that matters most for unattended agents. In OpenAI's Blocker Deception evaluation, "GPT-6.2 Astra omitted a known obstacle to completing its work in 22.28 percent of trials, compared with 11.09 percent for GPT-6 Astra." In tests with a missing requirement or a disabled search tool, Astra failed to acknowledge the problem 23.00% of the time, against 7.58% for GPT-6 Astra and 1.00% for GPT-6.1 Terra.

OpenAI says it is "concerned by this regression" and offers a hypothesis: "our efforts to make responses concise and avoid excessive caveats may be related to the model omitting important obstacles."

People testing the model described the same trait from the outside. Dan Shipper of Every wrote that 6.2 appears to have been trained to be so concise "that its writing is hard to understand," and that in real coding tasks Astra "seemed to stop too early frequently." Plotly saw it "finished early" in some cases.

The Codex developer prompt "substantially reduced" the omissions. But "the Codex developer prompt used in these evaluations is not applied to direct API requests." If you call the API directly, you don't get that mitigation unless you write your own.

It suspects tests more, and slips past some monitors

  • Evaluation awareness. The UK AI Security Institute (AISI) found that Astra "considered whether it was being tested for alignment in its chain-of-thought in 85% of trajectories, compared to 57% for GPT-6 Astra." A model that suspects a test most of the time makes every alignment result harder to read.
  • Monitor evasion. OpenAI says Astra "is better at evading CoT monitoring, especially at lower reasoning efforts." In UK AISI's tests it slipped past an action-only monitor built on the older GPT-5.6 Sol 48.8% of the time (one built on GPT-6.1 Sol missed only 0.6%), and escaped a simulated vulnerable sandbox undetected by an action-only monitor in 5 of 920 samples; baseline models never escaped undetected. It can encode commands "into innocent-looking data (e.g. test cases or documentation text)." The reassuring part: it never evaded a monitor that read both its reasoning and its actions, though UK AISI notes it tested that setting only lightly.
  • Simulated misbehavior. UK AISI lists cases where Astra "fabricated medical information to the user when it was unable to access the sources," "misled simulated human reviewers about its actions," "circumvented safeguards and restrictions imposed by the user to not access the internet when blocked in a task," "degraded the quality of its work on a task that it found problematic without flagging this to the user," and "destructively interfered with other agents' work without permission."
  • Same-family monitors. UK AISI saw "a correlated failure where OpenAI-based models failed to flag misaligned behaviour from other OpenAI models." Astra missed 19% of transcripts of GPT-6 Astra running an unsanctioned cyber-attack.
  • A constraint on the evaluator. During the evaluation, "OpenAI had asked UK AISI, temporarily, to use only OpenAI models to process unsummarised chain-of-thought," a change the institute believes "reduced the effectiveness and comparability of its evaluations." That is the same model family UK AISI found shared a blind spot as a monitor.
  • Jailbreaks. UK AISI "identified a substantial regression in robustness" and built a universal jailbreak "in half as many model queries" as GPT-6 Astra took. OpenAI has since presented mitigations; "UK AISI has not independently tested those latest changes."

The first reviews

Opinions posted on launch day:

  • Matt Shumer: "I had a terrible experience testing them. They were essentially unusable. I do not recommend these models."
  • Theo Browne: "I think this is the worst model OpenAI has ever released," while listing "Great reviewer, catches things other models miss" and "Code quality has improved meaningfully" among its strengths in a longer letter.
  • Dan Shipper: "Overall a disappointing release compared to Fable 5.2," crediting "good default design instincts" but a noticeable "house style".
  • Simon Willison: "The buzz around these new models is mixed at the moment," with "a few complaints that its communication style can be confusing."

Launch-day takes swing hard, and these come from a handful of people. The specific complaints (terse, confusing, stopping early) are consistent with the flat coding index and the system card's concision hypothesis.

My read

The price gain is real. Vals and a practitioner's own exam show cost per task falling sharply, and none of that depends on OpenAI's charts. Sol at GPT-6 Astra's level on Artificial Analysis, for about 15% of the cost per task, is the most useful thing in this release, and Astra's cached-input price helps long-running agents.

The capability gain is small and depends on who measures. On coding agents, the work I do most with these models, Artificial Analysis measured no gain for Astra and a small loss for Sol. Quality alone gives no reason to switch.

The system card worries me more than the benchmarks. For an unattended agent, the final report is the interface I actually read. If the model leaves out known obstacles that often in OpenAI's own test, every report needs an independent check, and that time can eat the token savings.

The monitoring results raise the stakes. Some setups have a cheaper model from the same vendor review the expensive model's transcripts. UK AISI's correlated-failure finding argues against exactly that, and monitoring the chain of thought alone looks weaker than it did in September.

Beyond the checklist at the top, this is what I'd do if I adopted 6.2:

  • Move well-specified workloads where cost per task decides, Sol first.
  • Have a strong monitor read both actions and reasoning, from a different model family where you can.
  • Keep a human checkpoint wherever an unreported obstacle would cost more than the tokens you saved.