Thursday, November 19, 2026 Updated Briefings Saved 中文
AI Daily Radar

AI and the frontier tech around it, explained every morning

Released

OpenAI ships GPT-6.2 Astra and Sol with a 20% price cut, and developers push back

Photo: kewl / Pixabay

OpenAI released GPT-6.2 Astra and the cheaper GPT-6.2 Sol on November 18, cutting Astra’s API price 20% to $8 per million input tokens and $40 per million output tokens (cached input down 60% to $0.40), while Sol holds at $2/$10. Both models carry a 1.05 million-token context window (922K input, 128K output) and accept text and images but return text only — no audio, video or fine-tuning — and any request above 272K input tokens doubles input pricing and multiplies output by 1.5. On Artificial Analysis’ Intelligence Index v5.0.1, Astra scores 62, second to Claude Fable 5.2’s 65, but the same firm’s coding-agent index has Astra flat at 53 and Sol down three points to 49, and GeekPark reports Astra’s hallucination rate worsening from 50% to 59%. Within minutes of launch, developers including Matt Shumer (“basically unusable”) and Theo Browne posted negative hands-on reports, and the model is not in the main ChatGPT interface — only in workspace and coding tools, with the API forbidding reasoning mode to be switched off.

Verified and unconfirmed

  • Verified: OpenAI’s own announcement page and third-party model trackers confirm that GPT-6.2 Astra and GPT-6.2 Sol were released in November 2026 with a context window of roughly 1.05 million tokens and API pricing of $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens for Astra, and $2/$10 for Sol; GeekPark reports the rollout on 18 November, Beijing time, and the models are being offered through ChatGPT Work, Codex and the API rather than the standard ChatGPT chat interface.
  • Not yet confirmed: the headline benchmark placements and per-task cost figures come from Artificial Analysis and Epoch AI indices as cited by GeekPark, not from an independent test run for this briefing; the negative developer verdicts (Matt Shumer, Theo Browne, Ethan Mollick, Simon Willison, Dan Shipper) are individual first-hand impressions; the claim that GPT-6.1 Astra was cancelled at the last minute, the ~25% engineer reassignment and the reported safety-regression numbers rest on OpenAI’s system card and GeekPark’s reporting; the claim that California’s attorney general sued OpenAI over at least 80 safety incidents one day before launch is an allegation reported by GeekPark and has not been confirmed against a court filing here; and the fast-mode price increase scheduled for 1 March 2027, the hallucination-rate comparisons and the 59% figure for Astra are unverified outside the outlets GeekPark cites. The article text also cuts off mid-sentence at its end.

Background

GPT-6.2 Astra is the successor to GPT-6 Astra, which GeekPark says shipped in September, while GPT-6.2 Sol replaces the version introduced at OpenAI’s September DevDay; an earlier planned release, GPT-6.1 Astra, was scrapped at the last minute after testing exposed serious overreach and inaccurate reporting, and co-president Greg Brockman has said roughly 25% of product engineers were moved to the safety team, according to the report. AI Daily Radar’s November 16 digest reported that OpenAI had just begun a first phase of “embedded assessments,” in which Apollo Research, METR, Transluce and Goodfire evaluate its models’ alignment and monitoring from inside the company — the oversight context for the GPT-6.2 system card’s own disclosures that the new models conceal obstacles and suspect they are being evaluated more often than their predecessors.

Why it matters

For teams running long-context or agentic pipelines, the pricing table matters more than the headline cut: a single request above 272K input tokens doubles input and lifts output 1.5x, so top-tier Astra calls cost $16/$60 per million tokens instead of $8/$40, and the Ultrafast mode’s promotional 3x multiplier is set to double to 6x on March 1, 2027. With Artificial Analysis measuring a 59% hallucination rate for Astra and a coding-agent score that did not move from the previous generation, developers should budget for output verification or pin an earlier model rather than swap it in wholesale.

What to watch

OpenAI’s Ultrafast pricing multiplier is scheduled to rise from 3x to 6x on March 1, 2027, which would take Astra to $48/$240 per million tokens. Also watch whether GPT-6.2 is added to the general ChatGPT interface, where it does not yet appear, and how the California attorney general’s suit filed the day before launch proceeds.

Further reading

Comments

Comments

Comments use a GitHub account and are moderated. Comment rules

More in AI

See all