3-Line TL;DR

  • OpenAI launched GPT-6.2 Astra (AA 62) and Sol (57) on November 17, cutting Astra's API rates 20% and boosting output speed 45–73%
  • Astra retakes #2 on Artificial Analysis ahead of Gemini 4 Carbon (60), while Claude Fable 5.2 stays #1 at 65
  • Performance and speed improved, yet the release feels slightly underwhelming and leaves Claude firmly on top (author's view)

Faster and Smarter, Yet Slightly Underwhelming

  • Task performance and speed clearly moved up, though overall the release still feels slightly short of expectations (author's view)
  • Released November 17, gpt-6.2-astra and gpt-6.2-sol support a 1,050,000-token context, 128,000 max output, and reject none reasoning (OpenAI changelog, Astra docs)
  • On AA Intelligence Index v5 (max), Astra rises 5 points to 62 at 64 tok/s (+45%), and Sol gains 2 points to 57 at 85 tok/s (+73%) (AA Astra, AA Sol)

Bar chart of AA Index v5 and output speed: Fable 5.2 (65, 112 tok/s), GPT-6.2 Astra (62, 64 tok/s), Carbon (60, 62 tok/s), Sol (57, 85 tok/s)

Made by HSL — AA Leaderboard, AA Astra, AA Sol, AA Fable 5.2, AA Carbon, checked November 18, 2026

  • OpenAI's self-reported benchmarks (not independent tests) put Astra 5.2 points above GPT-6 Astra on Agents' Last Exam and 6.2 percentage points higher on GDPval v2 (BigGo Finance)
  • On Cognition's FrontierCode 2.0 (also OpenAI-reported, not independent), Sol max trails Astra by 0.5 points at roughly one-fifth the cost (BigGo Finance)
  • On the independent AA Coding Agent Index (Codex harness, max), Astra stays flat and Sol drops 3 points behind GPT-6.1 Sol (AA report)
GPT-6.2 Astra and GPT-6.2 Sol—Models built for the real world
▶ GPT-6.2 Astra and GPT-6.2 Sol—Models built for the real world

Official launch overview uploaded by OpenAI

Price Cuts Do Not Shake Claude's #1 Spot

  • As things stand, Claude Fable 5.2 (AA 65, 112 tok/s) looks set to hold #1 firmly (author's view)
  • GPT-6.2 Astra cuts API rates 20% to $8 input and $40 output per 1M tokens, with cached input down 60% to $0.40 (Astra docs)

OpenAI GPT-6.2 Astra API docs capture showing $8.00 input, $0.40 cached input, $10.00 cache writes, and $40.00 output with a red box

Source: OpenAI API Docs — GPT-6.2 Astra pricing cards, checked November 18, 2026; red box added by HSL

  • GPT-6.2 Sol keeps its $2/$0.10/$10 rates, while prompts over 272K input tokens trigger 2x input/cache and 1.5x output pricing on both models (Sol docs)
  • Fable 5.2 costs $26.54 and 40.7 minutes per AA index task versus Astra's $4.00 and 10.7 minutes, yet still leads by 3 points (AA Leaderboard)

Welcome for Codex Users, Though Many Already Left

  • Existing Codex users should be happy with the update, though many developers have already switched to Claude (author's view; no official migration data)
  • Both models are live in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu, with $200/month Pro getting Ultrafast for the first time starting with Sol only (BigGo Finance)
  • $500/month Pro gets Ultrafast for both models (Enterprise and Edu when admin-enabled), while API Ultrafast costs 3x Standard through February 28, 2027 (BigGo Finance, Astra docs)
  • OpenAI's System Card notes improper workplace actions halved, yet model-only tests regressed on omitting blocking obstacles unless mitigated by the Codex developer prompt (System Card)

Two Months Late: Safety Reset vs. Modest Gains

  • Reporting points to that safety reset for the nearly two-month delay, while underwhelming capability gains may also have held it back (author's view)
  • GPT-6.2 arrives 75 days after GPT-6 Astra (September 3) and 49 days after GPT-6.1 Sol (September 29) (OpenAI changelog)

Timeline from Sep 3 GPT-6 Astra to late Sep 6.1 Astra shelved and Sep 29 6.1 Sol shipped, Oct 28-Nov 9 Carbon and Fable 5.2, and Nov 17 GPT-6.2

Made by HSL — OpenAI Changelog and BigGo Finance, checked November 18, 2026

  • In late September, OpenAI shelved GPT-6.1 Astra after internal tests found it fell short on staying within scope and reporting its work (BigGo Finance)
  • GPT-6.2 Astra is also OpenAI's first external release to meet the High threshold for AI Self-Improvement alongside Critical for Cybersecurity (System Card)

Back to #2 Over Google, Depending on the Test

  • Even so, retaking #2 from Google with 6.2 Sol and Astra is worth noting (author's view)
  • Google had taken #2 in our Carbon post, yet on AA v5 Astra (62) passes Carbon (60) and Claude Opus 5.5 (59) (AA Leaderboard)
  • Astra is also cheaper and faster per AA task at $4.00 and 10.7 minutes versus Carbon's $5.48 and 33.2 minutes (AA Leaderboard)
  • On Vals Index (finance, coding, legal, and tax), however, Carbon still leads Astra 70.45% to 67.66% behind Fable 5.2's 71.31% (Vals Index)

Comparison table of Claude Fable 5.2, GPT-6.2 Astra, Gemini 4 Carbon, GPT-6.2 Sol, and GPT-6.1 Sol showing API rates, AA task cost, and time per task

Made by HSL — AA Leaderboard and model pricing pages, checked November 18, 2026; Gemini 4 Carbon $2/$10 rate runs through November 30 before rising to $4/$20

Workload Start with Switch when
Everyday coding in Codex GPT-6.2 Sol ($2/$10) Key facts get omitted on Q&A or edits
Hard research or long-context agents GPT-6.2 Astra ($8/$40) Task needs top-end accuracy over budget
Max accuracy regardless of bill Claude Fable 5.2 ($8/$40) $26.54/task cost or 40.7-min wait hurts

The Catch: Fable Is Still 5.2, Not 5.5

  • The bigger worry for OpenAI is that Anthropic's top model is still labeled Fable 5.2 rather than 5.5 (author's view)
  • As documented in our prior Fable post, Opus, Sonnet, and Haiku are already at 5.5 while Anthropic's system card lists an unreleased internal model above Fable 5.2
  • No 5.5 name or release date is announced, yet if Anthropic holds a higher tier in reserve, the gap to #1 could widen again (conditional outlook)

Community Reactions

  • HN user HyperL0gi says their company allows one subscription per employee and they see no reason to leave Claude after switching on Opus 5.5, citing Fable's quality and Astra's earlier token burn (individual experience, no measurements) (Hacker News)
  • User jotato notes on OpenAI's FrontierCode 2.0 chart that 6.2 Sol at High looks like a modest bump over 6.1 Sol (38.3% at $0.68 vs 34.9% at $0.65) (commenter reading OpenAI's chart) (Hacker News)
  • User guybedo says Astra and Sol look strong on benchmarks but still over-engineer simple tasks in real work, pushing them back to Claude to clean up (individual impression, no measurements) (Hacker News)
  • User jrflo highlights the System Card's finding that both models cut improper workplace actions yet omitted blocking obstacles more often in model-only tests (quotes System Card) (Hacker News)

Q&A (Field Notes)

  • Q. Did GPT-6.2 Astra actually take #2 from Gemini 4 Carbon?
    • On Artificial Analysis v5, yes: Astra scores 62 against Carbon's 60 behind Fable 5.2's 65
    • On Vals Index (finance, coding, legal, and tax), Carbon still leads Astra 70.45% to 67.66%
  • Q. Is GPT-6.2 Sol a clear upgrade over GPT-6.1 Sol?
    • Output speed jumped 73% (49 to 85 tok/s) at the same $2/$10 API price and AA index rose 2 points to 57
    • On AA's Coding Agent Index, however, Sol dropped 3 points due to omitting key facts on SWE-Atlas-QnA-AA (-14 points)
  • Q. Who gets Ultrafast mode in ChatGPT and Codex?
    • Plus, Pro, Enterprise, and Edu users get both models, while Ultrafast opens on $200/month Pro (first time, Sol only), $500/month Pro (both), and admin-enabled Enterprise/Edu
  • Q. Why does Fable being 5.2 instead of 5.5 matter here?
    • Opus, Sonnet, and Haiku are already at 5.5 while Fable sits at 5.2, and Anthropic's system card lists a stronger unreleased internal model
    • If a higher Fable tier ships later, GPT-6.2's 3-point gap to #1 could widen again