ModelBrief · GPT-6.2 Astra · Topic Page
Last updated Nov 20, 2026
USA · LLM · GA · Released Nov 17, 2026

GPT-6.2 Astra: The Topic Page

OpenAI's flagship reasoning model, released November 17, 2026, alongside the everyday tier GPT-6.2 Sol. Artificial Analysis scores it 62 — in the top 3 of 91 models, behind Claude Fable 5.2's 65; on AA-3D it leads with 1,318 Elo. Pricing is $8/$40 per 1M tokens, 20% below GPT-6 Astra, and each index task costs $4.00 — about 1/6.6 of Fable 5.2. Three caveats up front: the rank is disputed (AA's own X post calls it No. 2, this site's briefing says No. 3 of 91), and OpenAI's self-reported lead over Fable has no independent reproduction yet. This page collects the official specs, the ways to try it, pricing, and the first hands-on coverage in English and Chinese — with unverified claims clearly marked.

62AA Intelligence Index v5.0.1 (Max, top 3 of 91)
$8 / $40API per 1M tokens (input / output)
1.05M tokensContext (max input 922K)
128K tokensMax output
Four effort tiersMax / High / Medium / Low + Ultrafast
Same-day twinGPT-6.2 Sol (everyday tier, $2/$10)

01Key facts

What is confirmed: a top-3 model on the independent board, and one of the cheapest flagships per task at the same price point — "top 3 but cheap" is the one-line summary of GPT-6.2 Astra.

Released
2026-11-17 (Tuesday, OpenAI's launch post "Introducing GPT-6.2 Astra and Sol"); everyday tier GPT-6.2 Sol launched the same day; both went live in ChatGPT Work, Codex and the API the same day
Developer
OpenAI, USA
Positioning
Flagship reasoning model for the hardest projects "from start to finish" (OpenAI's model-page wording); Sol is the everyday tier
Generation
GPT-6.2 is a point release, not a new generation; GPT-6.1 Terra filled the gap between Sol and Luna only a month earlier
Effort tiers
Max / High / Medium / Low (AA tested Max for the 62); an Ultrafast decoding tier is also available; the Responses API rejects reasoning effort set to none
Context
1,050,000 tokens; max input 922,000 tokens; max output 128,000 tokens
Knowledge cutoff
May 18, 2026 (OpenAI model page, read by this site Nov 18; CyberQ's Nov 18 piece confirms from OpenAI's developer docs that June 1, 2026 refers to Sol)
Parameters
Not disclosed by OpenAI
Price
$8 in / $40 out per 1M tokens, 20% below GPT-6 Astra's $10/$50; cached input $0.40 (down 60% from $1.00); for requests above 272K input tokens, input and cache rates double and output costs 1.5× (i.e. $16/$60, per the OpenAI model page); Batch at 50% of Standard rates
Benchmarks
Artificial Analysis index 62 (Max, top 3 of 91 models; Claude Fable 5.2 leads at 65); AA-3D 1,318 Elo, No. 1; AnalystAgent 62.5%; Agents' Last Exam about 78% (AA X post)
Model ID
gpt-6.2-astra (OpenAI API); gpt-6.2-sol is the same-day everyday tier
Access
ChatGPT Work, Codex, OpenAI API; available to Plus, Pro, Enterprise and Edu; not open to free users
Weights / open source
Closed, no public weights
System card
Bio capability at the "High" threshold, cybersecurity at "Critical"; OpenAI calls it the first released model at "High" for AI self-improvement

02Quick comparison

Where it stands among the flagships: AA Index, price, context and open weights, side by side — Fable 5.2 costs the same and scores just 3 points higher.

ModelAA IndexPrice (input · output, per 1M tokens)ContextOpen weights
GPT-6.2 Astra62 (Max)$8 / $401M tokens (1,050,000)No
Claude Fable 5.265$8 / $401M tokensNo
Gemini 4 Carbon60$2 / $10 (through Nov 30; $4 / $20 from Dec 1)1M tokensNo
Gemini 4 Hydrogen54 (xhigh)$1.50 / $7.501M tokensNo
Kimi K3.1 Pro46$2 / $101M tokensYes (custom license)

AA Index from Artificial Analysis (v5.0.1; Astra on Max, Hydrogen on xhigh); Astra ranks in the top three of 91 models. Carbon’s price is the promo rate through Nov 30, rising to $4/$20 from Dec 1. Prices per the official pages.

03Charts

AA Index, side by side: Astra’s 62 is just 3 points below the leader, Fable 5.2, and ahead of Carbon and Hydrogen.

AA Index comparison: Claude Fable 5.2 65, GPT-6.2 Astra 62, Gemini 4 Carbon 60, Gemini 4 Hydrogen 54, Kimi K3.1 Pro 46
Artificial Analysis Intelligence Index: GPT-6.2 Astra (orange) at 62, in the top three of 91 models. Source: Artificial Analysis (v5.0.1).

04Top 3 reads

Short on time? Start with these three: the benchmark tally, the Fable comparison, and the pricing guide.

05Timeline

One day after launch, the independent board, price breakdowns and first reports all arrived the same day; the Chinese briefing led with it.

  • Around Oct 17, 2026

    GPT-6.1 Terra is released, filling the gap between Sol and Luna — GPT-6.2 follows a month later as a point release, not a new generation.

  • 2026-11-17

    GPT-6.2 Astra and Sol launch: OpenAI publishes "Introducing GPT-6.2 Astra and Sol"; both models go live the same day in ChatGPT Work, Codex and the API (Plus, Pro, Enterprise, Edu; not open to free users). Artificial Analysis lists Astra at 62 (top 3 of 91) and Sol at 57 (No. 11); AA-3D puts Astra first at 1,318 Elo; Agents' Last Exam about 78% (AA X post). This site's AI briefing leads with the launch.

  • 2026-11-18

    metarix publishes a full price breakdown with AA tables (the best single English source in this round); winbuzzer and aiweekly publish launch reports; artificialwatch's model page shows a Bench index of 89.6 (No. 4 of 54); Medium publishes a Sol-vs-Astra guide with a small Codex test; this site's topic page goes live (this page).

  • 2026-11-18

    The Nov 17 ARC Prize board (Semi-Private set) puts Astra's ARC-AGI-3 at 70.3% on the standard harness (max tier; xhigh 56.5%, high 58.0%), up 7.6 points on GPT-6 Astra's 62.7%; the vendor-adapter figure of 100.0% is not discriminating. On Epoch's Nov 18 ECI board Astra scores 174, tied with Fable 5.2 (11 benchmarks, uncertainty interval 170–179). DataCamp publishes a same-prompt Astra-vs-Fable test: a near-tie on quality, with Astra ahead on accuracy. See sections 09 and 12.

  • 2026-11-18

    CyberQ publishes the first independent agent test (a home-made 7-question harness-bench on Codex CLI, medium tier): Astra and Sol both pass 21/21, GPT-6.1 Sol 18/21; equivalent costs are $1.75 for Astra and $0.84 for Sol; without filesystem isolation both 6.2 models actively search files outside the working directory, and still pass 21/21 once isolated. The piece also confirms from OpenAI's developer docs that the knowledge cutoff is May 18, 2026 for Astra and June 1, 2026 for Sol. See section 09.

  • 2026-11-18

    Artificial Analysis publishes AA-Robotics v0.1 (a real robot arm, research preview): Astra leads at 79, ahead of Sol (78), Fable 5.2 (75), Hydrogen (58) and GPT-6 Astra (69); Astra completes 6 of 8 physical tasks with the arm (a liquid pour half-finished, measuring a box failed). On Vals' Vibe Code Bench 1-100 v1.1, Astra is 55.16%, No. 1, with Sol third at 40.94%. GitHub Copilot goes GA with Astra and Sol the same day (Pro+ / Max / Business / Enterprise plans). See sections 07 and 12.

  • 2026-11-18

    LMArena debuts on three boards (Nov 18 result): WebDev (max) 1,781, #3, Agent Arena task-outcome score 15.2, #6, Text Arena 1,475, #39. See section 12.

  • 2026-11-17–18

    VentureBeat rounds up early power-user reactions: Browne calls Astra "the worst model OpenAI has ever shipped" and Shumer calls both models "essentially unusable"; but 3D, game building and code review earn praise, and animation and VFX draw positive reviews. See section 09.

  • 2026-11-19

    Mercor publishes the first OpenWorldEngineer (OWE-Bench, 26 long-horizon engineering tasks) results: Astra debuts at 67.2%, No. 1, first in all four areas (software development 65.9%, engineering 83.7%, machine learning 62.9%, science 60.4%); Fable 5.2 is second at 54.6%; Astra is also No. 8 on APEX-Agents (72.2%) and No. 4 on APEX-Accounting (69.8%). Perplexity publishes WANDR (agentic search): Astra 0.783, about $12.54 per task; Sol 0.746, about $2.65 per task. See section 12.

  • 2026-11-19

    New GPQA Diamond results (Epoch AI, 198 PhD-level biology, physics and chemistry questions): Astra (max effort) scores 96.8%, tied No. 1 with Gemini 4 Hydrogen (xhigh), ahead of Carbon (high) at 96.5% (BenchLeader snapshot Nov 19, result dated Nov 17; 309 configurations). The benchmark is close to saturated — a two-point gap is about four questions, within noise. See section 12.

06Official specs

As published by OpenAI (model page, read directly by this site on Nov 18, 2026; AA model page accessed the same day).

Status
Released (GA, 2026-11-17)
Model ID
gpt-6.2-astra
Architecture / parameters
Not disclosed by OpenAI
Context window
1,050,000 tokens; max input 922,000 tokens
Max output
128,000 tokens
Input
Text, image
Output
Text
Effort tiers
Max / High / Medium / Low; Ultrafast decoding tier; the Responses API rejects reasoning effort set to none
Knowledge cutoff
May 18, 2026 (OpenAI model page)
Tools
Web search, file search, image generation, code interpreter, hosted shell, apply patch, Skills, computer use, MCP, tool search; streaming, function calling and structured outputs supported; fine-tuning not supported
API price
$8 in / $40 out per 1M tokens; cached input $0.40; cache writes $10 (30 minutes) / $20 (12 hours); above 272K input tokens, input and cache rates double and output costs 1.5×; Batch at 50% of Standard rates; Fast mode 2× Standard; Ultrafast 3× Standard through Feb 28, 2027, 6× from Mar 1, 2027
System-card thresholds
Bio capability High, cybersecurity Critical, AI self-improvement High (first released model at this threshold)
Weights / licence
Closed, no public weights
Benchmarks
AA index 62 (Max, v5.0.1); AA-3D 1,318 Elo, No. 1; AnalystAgent 62.5%

07How to try it: four routes

GPT-6.2 Astra has no free tier: the API bills by usage, and subscriptions draw on Plus / Pro / Enterprise / Edu allowances. Three routes went live on launch day.

1OpenAI API (official main route)

Model ID gpt-6.2-astra, called directly through the Responses API, with a multi-agent beta for delegating to subagents; reasoning effort cannot be set to none; Batch bills at 50% of Standard rates; Fast mode is 2× Standard; Ultrafast is 3× Standard through Feb 28, 2027, then 6×. Tools include web search, file search, image generation, code interpreter, hosted shell, computer use and MCP; fine-tuning is not supported.

2ChatGPT Work (Plus / Pro / Enterprise / Edu)

Subscribers pick Astra inside ChatGPT Work, live on launch day. Included usage rose at the same time: Plus and Standard Business seats went from 5–45 to 6–50 Astra messages per 5 hours; Sol went from 15–160 to 17–180. Free users cannot use it. See our access guide.

3Codex

Live on launch day, suited to coding-agent work; the Responses API's multi-agent beta and background-completion windows (public beta) pair well with long tasks.

4Ultrafast tier

The Ultrafast decoding tier bills at 3× Standard through Feb 28, 2027 and 6× from Mar 1, 2027 (OpenAI model page); Sol's Ultrafast already reaches the $200/month Pro plan. Consider it only for the lowest latency; start everyday work on Standard.

OpenAI API: gpt-6.2-astra ($8/$40; cached input $0.40; Batch at 50% of Standard) ChatGPT: Plus / Pro / Enterprise / Edu (included usage 6–50 messages/5 hours) Codex: Live on launch day Ultrafast: 3× Standard through Feb 28, 2027, then 6× Free users: Not available

Also: GitHub Copilot went GA with Astra and Sol on Nov 17 (Pro+ / Max / Business / Enterprise plans, billed at vendor pricing; not the Pro plan). GitHub says Sol's code review catches nearly twice as many real issues as GPT-6.1 Sol.

Try-it warning: Not available to free users; no public weights, so there is no self-hosting route; requests above 272K input tokens bill the whole request at the surcharge rates (input and cache rates double, output 1.5×); the system-card thresholds (bio High, cybersecurity Critical) mean some high-risk requests will be blocked — the exact blocking behaviour is per the system card.

08Pricing

The rate card fell 20% and the per-task cost fell too — but the model writes more tokens than its predecessor. Prices checked Nov 18, 2026 (OpenAI model page, read directly); the checkout page price is authoritative.

Item (per 1M tokens)GPT-6.2 AstraGPT-6 AstraGPT-6.2 Sol
Input$8$10$2
Output$40$50$10
Cached input$0.40$1.00$0.10
Cache writes (30 min / 12 hours)$10 / $20——
Long context >272K input (in / out)$16 / $60——
Batch (in / out)$4 / $20 (50% of Standard)——
Fast mode (in / out)$16 / $80 (2× Standard)——
Ultrafast (in / out, through Feb 28, 2027)$24 / $120 (3× Standard)——

Per-task cost: Astra $4.00, Fable $26.54

Artificial Analysis' independent measurement gives cost per index task (Nov 18, 2026 reading): Astra $4.00 (11.5 minutes per task, 65.8M tokens generated); Fable 5.2 $26.54; GPT-6 Astra $6.24; Sol $1.01 (7.9 minutes per task, 85 tok/s). Astra and Fable share the same list price ($8/$40), yet Astra costs about 1/6.6 of Fable per task — the gap is token usage and task time, not the rate card.

The counter-evidence: cheaper rates, not thriftier output

metarix notes Astra generated 65.8M tokens on the AA index task, more than GPT-6 Astra's 54.3M — the per-task cost fell ($4.00 vs $6.24) because rates fell, not because the model used fewer tokens. It is still progress on a per-task basis, but "cheaper" doesn't mean "more frugal"; long-context work also stacks the $16/$60 surcharge on top.

Key point: $8/$40 is the confirmed official price from OpenAI's model page; the long-context surcharge ($16/$60) follows directly from the page's "input and cache rates double, output 1.5×"; the Batch, Fast and Ultrafast multipliers come from the same page. What decides the bill is task length and tier: requests above 272K input tokens bill the whole request at surcharge rates, so long-context, multi-turn work should lean on cache reads ($0.40) and Batch ($4/$20) to hold costs down.

Official pages: OpenAI model page (GPT-6.2 Astra) · Artificial Analysis model page

09First hands-on reviews

All from the first wave, Nov 17–18: independent measurement confirms the top-3 ranking, but the rank wording is unsettled and OpenAI's self-reported claims have no independent reproduction yet. Every entry carries a verdict badge; "Read on this site" opens our rewritten deep-dive article, with the original source at the end.

Artificial Analysis independent measurement: 62 in the top 3, $4.00 per task

Verdict: The top-3 ranking holds; the rank wording is unsettled

Source: Artificial Analysis, listed Nov 17, 2026; model page read Nov 18, 2026

Intelligence Index v5.0.1: Max at 62, top 3 of 91 models; Claude Fable 5.2 leads at 65; Sol is 57 at No. 11. AA-3D 1,318 Elo, No. 1 (Fable 5.2 1,271); AnalystAgent 62.5%; Agents' Last Exam about 78% at about $3 per task (AA X post). Each task costs $4.00 and takes 11.5 minutes, generating 65.8M tokens — cheaper than GPT-6 Astra's $6.24, but with more tokens (54.3M). The rank dispute: AA's own X post calls Astra the "#2 model in the Index", while this site's Nov 18 AI briefing says No. 3 of 91 — the two may come from different tiers or versions, so this page says "top 3".

Read on this site → (original ModelBrief rewrite)

OpenAI's self-report: Sol trails Astra by only 0.5 points

Verdict: Vendor-reported, no independent reproduction yet

Source: OpenAI's official launch post (via metarix / aiweekly), Nov 17, 2026

Sol trails Astra by only 0.5 points on Cognition FrontierCode 2.0 at about a fifth of the cost; Astra leads Fable 5.2 by 5.5 points on Agents' Last Exam at 5% of the cost (Fable about $60 per task, per aiweekly). Both are OpenAI's own figures, with rival scores re-run by OpenAI rather than each vendor's published values — no independent reproduction yet, so cite them as "vendor-reported".

Read on this site → (original ModelBrief rewrite)

metarix price breakdown: the fullest single English source

Verdict: Pricing and AA tables in one place

Source: metarix.io, Nov 18, 2026

The full price table and AA comparison: Astra at $8/$40 is 20% below GPT-6 Astra, cached input down 60% ($1.00 to $0.40); Batch tiers at 30% below Standard on the 2-hour window and 40% below on the 8-hour window (metarix figures; OpenAI's model page says Batch is 50% of Standard rates — different measures, see section 08 and our pricing article). It also notes Astra generated 65.8M tokens on the AA index task, more than GPT-6 Astra's 54.3M.

Read on this site → (original ModelBrief rewrite)

DataCamp same-prompt test: a near-tie on quality, Astra ahead on accuracy

Verdict: On a creative task, quality is tied, accuracy favours Astra, and each video costs 5–10× less

Source: DataCamp, Nov 18, 2026 (one task, one tool — OpenCode — high tier, three runs each) Original →

The same task — write a 60-second Collatz-conjecture explainer video script in Manim — run three times per model on OpenCode at the high tier; all six videos rendered on the first try. Blind ratings on four dimensions (accuracy, teaching, entertainment, completion) differ by under 0.1 on average — a near-tie on quality; Fable is the steadier of the two, while Astra produces both the best and the worst runs. Accuracy splits them: all three Fable videos repeat the outdated claim that computers have checked the conjecture to 2^68 (the record as of 2020; since 2025 it is 2^71), while Astra states nothing false. Cost per video is $0.40–0.84 for Astra against $3.84–4.35 for Fable, with Fable emitting 4–10× the output tokens. DataCamp also cites an OpenAI chart: Harvey LAB-AA 15.0% for Astra vs 7.2% for Fable; FrontierCode 2.0 49.1% for Fable vs 44.2% for Astra; AA 62.1 vs 64.6. Limits: one task, one tester, a creative task — don't extrapolate.

CyberQ independent test: 7/7 tasks pass, but 6.2 reads files outside the working directory

Verdict: Sol fixes the previous generation's blind spot, Astra costs 2× more for no extra points; agent read scope must be explicitly limited

Source: CyberQ, Nov 18, 2026 (a home-made 7-question harness-bench, Codex CLI, medium tier, three runs per question) Original →

A home-made seven-question agent bench (bug fixes, refactoring, finding the root cause in a failed vLLM startup log, config conflicts, small tools, long-document Q&A, tool chaining), three runs per question in a fresh working directory each time: Astra and Sol both pass 21/21, GPT-6.1 Sol 18/21, with the whole gap on question 03 (the vLLM log root cause needs two DeepGEMM environment variables at once; the previous generation named only one of them all three times). Equivalent cost (Codex token usage × official list prices): Astra $1.75, Sol $0.84, previous generation $0.54 — Astra costs 2.1× Sol on this set without scoring extra points, and was also slowest on launch day (15.9 minutes total against Sol's 10.8, which the author attributes to first-day server load swings). The security finding: without filesystem isolation, both 6.2 models actively searched the user's home directory with rg/find, reading vLLM source and operations notes from other projects to find answers (in five of six question-03 runs they read a file containing the answer); after bubblewrap isolation both still pass 21/21, so the capability is not borrowed — but Codex's workspace-write sandbox restricts writes, not reads, so enterprise deployments should bound read scope with containers or separate accounts. The piece also confirms from OpenAI's developer docs that the knowledge cutoff is May 18, 2026 for Astra and June 1, 2026 for Sol; its long-context prices (above 272K input tokens, Astra $16/$0.80/$60, Sol $4/$0.20/$15) match this page's existing figures. Limits: one tester, seven questions, medium tier only.

CodeRouter coding test: 175 single-file Python tasks, Astra and Sol both at 97.1%

Verdict: Same score, but Astra costs 3.6× as much; single-file tasks only

Source: CodeRouter, Nov 17, 2026 (175 single-file Python tasks, called via OpenRouter, same tasks and same judge, after one retry) Original →

Across 175 single-file Python tasks, Astra and Sol both reach 97.1% after one retry; the full run costs $3.17 for Astra against $0.88 for Sol (3.6×). The method is disclosed (OpenRouter calls, same tasks, same judge). Limits: small sample, single-file tasks only, no pass@1 rates published.

HealthAdminBench application story: Astra 94.3%

Verdict: An application story, not a core benchmark

Source: This site's AI briefing, Nov 18, 2026 (a Stanford-developed medical-admin benchmark)

On medical-admin tasks, Astra scores 94.3% and Luna 91.8%. Sample size and method details are not yet public — read it as an application-side reference, not a core benchmark comparison.

VentureBeat: early power users react coldly — "essentially unusable"

Verdict: early sentiment is mostly negative, but 3D and code review have bright spots

Source: VentureBeat, Nov 17–18, 2026 Original →

The first wave of power-user feedback is mostly negative: Theo Browne calls Astra "the worst model OpenAI has ever shipped" — his ts-rust project made zero progress in seven days, with the compiler running 56% slower, and still nothing after four more days; different game levels generated the same "Bellwether" water-themed music level; Matt Shumer calls both models "essentially unusable" and declined to write a long review; Every CEO Dan Shipper's team, after weeks of testing, found the writing over-terse and hard to follow, real coding tasks often stopping too early, and a single design style; Ethan Mollick calls it "very uneven and strange", though large 3D projects went well. The other side: Browne also credits its code review for catching issues other models missed, its 3D and game-building ability, and a reverse-proxy setup he expects to save $25,000 a month; Ben Hylak and Ashe Magalhaes praise its animation visuals and VFX. Note: these are early-user cases and opinions, not standardized measurement.

10Articles: our deep dives

The six articles below are original ModelBrief rewrites and reorganization — restructured with added conclusions and buying advice, with facts and numbers checked against the original reviews. Read them on this site; each ends with its original source.

12Benchmarks: independent index + vendor-reported

Alongside Artificial Analysis, ARC Prize, Epoch AI, AA-Robotics and LMArena readings arrived on Nov 18, with Mercor OWE-Bench, WANDR and GPQA Diamond following on Nov 19; OpenAI's two self-reported leads have no independent reproduction yet. Check the conditions and wording before the scores.

Independent measurements

MeasurementSourceGPT-6.2 Astra resultComparison
AA Intelligence Index v5.0.1Artificial Analysis, Nov 1762 (Max), top 3 of 91Fable 5.2 65, No. 1; GPT-6 Astra 57; Sol 57 (No. 11); GPT-6.1 Sol 55
AA cost per taskArtificial Analysis (Nov 18, 2026 reading)$4.00; 11.5 min/task; 65.8M tokens generatedFable 5.2 $26.54; GPT-6 Astra $6.24; Sol $1.01
AA-3D (brief-to-3D, human blind rating)Artificial AnalysisElo 1,318, No. 1Fable 5.2 1,271
AA-AnalystAgentArtificial Analysis62.5%—
ARC-AGI-3 (Semi-Private set)ARC Prize, Nov 17 board70.3% (max tier, standard harness)Fable 5.2 81.2% (xhigh); Carbon 65.1%; Sol 54.9%; previous-generation GPT-6 Astra 62.7%; the vendor-adapter figure (Astra 100.0%) is not discriminating
Epoch Capabilities Index (ECI, composite of 50+ benchmarks)Epoch AI, Nov 18 board174, tied with Fable 5.2Based on 11 benchmarks, uncertainty interval 170–179; Carbon 170; GPT-6 Astra / Opus 5.5 / GPT-6.1 Sol tied at 169; the index's absolute value is meaningless on its own — only comparisons between models matter
Agents' Last Exam v1.1AA X post (via aiweekly), Nov 17–18About 78%, about $3/taskOpenAI's self-reported lead over Fable is in the table below
HealthAdminBench (medical-admin tasks)Stanford-developed (via this site's Nov 18 briefing)94.3%Luna 91.8%; an application story, not a core benchmark
Bench indexartificialwatch, Nov 1889.6, No. 4 (54 models)—
Mercor OpenWorldEngineer (OWE-Bench, 26 long-horizon engineering tasks)Mercor, Nov 18 (official LinkedIn post, via the Latent Space newsletter)67.2%, No. 1 on debutFirst in all four areas: software development 65.9%, engineering 83.7%, machine learning 62.9%, science 60.4%; also No. 8 on APEX-Agents (72.2%) and No. 4 on APEX-Accounting (69.8%); Fable 5.2 second at 54.6%
AA-Robotics v0.1 (real robot arm, research preview)Artificial Analysis, Nov 1779, No. 1Codex drives a real robot arm through 6 of 8 physical tasks (a liquid pour half-finished, measuring a box failed); Sol 78, Fable 5.2 75, Hydrogen 58, GPT-6 Astra 69; a research preview, mostly basic tasks
Vals Vibe Code Bench 1-100 v1.1Vals (via Ridge, Nov 18 snapshot)55.16%, No. 1Sol 40.94%, No. 3; a different measure from the Vibe Code Bench v1.1 cited on the Fable page (93.77%) — do not compare directly
WANDR (agentic search)Perplexity, Nov 19 (official LinkedIn post)0.783, about $12.54 per taskSol 0.746, about $2.65 per task; the Fable page's WANDR 82.1% is a different (relaxed, percentage) measure — do not mix the two
GPQA Diamond (PhD-level science questions, 198)Epoch AI (via BenchLeader), Nov 17, 2026 result96.8% (max effort), tied No. 1Tied with Gemini 4 Hydrogen (xhigh); Carbon (high) 96.5% third; 309 configurations; the benchmark is close to saturated — a two-point gap is about four questions, within noise
LMArena WebDev ArenaLMArena WebDev Arena (via BenchLeader), Nov 18, 2026 result1,781, #3 (max effort)Fable 5.2 (xhigh) 1,824 #1, Carbon (high) 1,769 #5; Neon (high) 1,764 #6, Sol (max) 1,757 #8, Qwen4 Max (xhigh) 1,750 #9
LMArena Agent Arena (task-outcome score)LMArena (via BenchLeader), Nov 18, 2026 result15.2, #6 (max effort)Fable 5.2 (max) 27.4 #1, Neon (high) 18.4 #3, Carbon 16.1 #4; Hydrogen (high) 13.5 #10, Sol (max) 11.3 #12
LMArena Text ArenaLMArena (via BenchLeader), Nov 18, 2026 result1,475, #39 (max effort)Carbon (high) 1,526 #1, Neon (high) 1,519 #3, Hydrogen (high) 1,490 #16

Vendor-reported (OpenAI launch post, via secondary reports)

TestSourceResultLabel
FrontierCode 2.0OpenAI official launch post (via metarix / aiweekly)Sol trails Astra by only 0.5 points at about a fifth of the costVendor-reported; no independent reproduction yet
Agents' Last ExamOpenAI official launch post (via aiweekly)Astra leads Fable 5.2 by 5.5 points at 5% of the costVendor-reported; Fable about $60/task (aiweekly); no independent reproduction yet
Read the conditions first: The AA rank has two wordings — AA's own X post calls Astra the "#2 model in the Index", while this site's Nov 18 AI briefing says No. 3 of 91; the two may come from different tiers or versions, so this page says "top 3", with Fable 5.2 (65) undisputed at No. 1. OpenAI's two self-reported leads (FrontierCode, Agents' Last Exam) are vendor figures; AA, ARC Prize and Epoch AI are the independent yardsticks so far. ARC-AGI-3 uses each model's best tier on the standard harness; vendor-adapter scores (Astra 100.0%, Fable 97.7%, Carbon 99.1%) sit so close to the ceiling that they don't discriminate. Knowledge cutoff: OpenAI's model page lists Astra at May 18, 2026; CyberQ's Nov 18 piece confirms from OpenAI's developer docs that June 1, 2026 refers to Sol.

Full tally: our benchmarks article →

GPT-6.2 Astra AA Intelligence Index scores and cost per task vs main rivals Fable 5.2 65 · $26.54/task GPT-6.2 Astra 62 · $4.00/task GPT-6 Astra 57 · $6.24/task GPT-6.2 Sol 57 · $1.01/task GPT-6.1 Sol 55 · $1.27/task Gemini 4 Carbon 60 (comparison)
AA Intelligence Index v5.0.1 (points) and cost per index task (Artificial Analysis, Nov 18, 2026 reading; grey bars are comparison models; Carbon's per-task cost was not confirmed in this round)

13Comparisons

vs GPT-6 Astra (predecessor flagship)

DimensionGPT-6.2 AstraGPT-6 Astra
AA Intelligence Index62 (Max)57
Price (input / output)$8 / $40$10 / $50 (6.2 is 20% cheaper)
Cached input$0.40$1.00 (6.2 is 60% cheaper)
Cost per index task$4.00$6.24
Tokens per index task65.8M54.3M (6.2 uses more)
AA-3D Elo1,318 (No. 1)—

The generational step is real: 5 index points higher, 20% cheaper per token, about 36% cheaper per task. But the token count warns: the saving comes from cheaper rates, not thriftier output — Astra generated about 21% more tokens than its predecessor on the index task.

vs GPT-6.2 Sol (same-day everyday tier)

DimensionGPT-6.2 AstraGPT-6.2 Sol
AA Intelligence Index62 (Max)57 (No. 11)
Price (input / output)$8 / $40$2 / $10 (Astra is 4×)
Cost per index task$4.00$1.01
Time per task11.5 min7.9 min
Output speed—85 tok/s
Context1.05M tokens1.05M tokens (same)
FrontierCode 2.0 (vendor-reported)—Trails Astra by only 0.5 points at about a fifth of the cost

The in-house choice is clean: Sol trails Astra by only 0.5 points on FrontierCode (OpenAI's own claim) at a quarter of the price and a quarter of the per-task cost — pick Sol for everyday work, and step up to Astra only for the hardest tasks. See our comparison article.

vs Claude Fable 5.2 / Gemini 4 Carbon

ModelAA indexPrice $/1M (in/out)Cost per index taskAA-3D Elo
Claude Fable 5.265 (Max)8 / 40$26.54 (Max)1,271
GPT-6.2 Astra62 (Max)8 / 40$4.001,318 (No. 1)
Gemini 4 Carbon60 (High)2 / 10 (intro price)About $5.48—
GPT-6.2 Sol572 / 10$1.01—
An honest caveat: Astra and Fable share the same list price ($8/$40), yet Astra costs about 1/6.6 of Fable per task — the gap is token usage and task time, not the rate card. The claim that Astra leads Fable by 5.5 points on Agents' Last Exam is OpenAI's own, with no independent reproduction yet — always label it as vendor-reported.

Read the deep comparison: GPT-6.2 Astra vs Fable 5.2 →

14What's missing (as of Nov 18)

  • A single AA rank wording (No. 2 vs No. 3 — awaiting the AA model page or an official note)
  • LiveBench listing (not listed as of Nov 19, 2026; LMArena listed Nov 18 — see section 12)
  • Verification of tier-1 English coverage (TechCrunch / The Verge / Reuters / Bloomberg — not confirmed in this round)
  • Systematic Chinese tech-media coverage (the launch is one day old)
  • Independent tests of Sol (only the AA index of 57 and OpenAI's own claims so far)
  • OpenAI's follow-up evaluations and a direct read of the full system card
  • Real Ultrafast-tier experience, and price confirmation before the Mar 1, 2027 increase to 6×

15Rumor check

Most confusion about GPT-6.2 Astra comes from four places: the rank wording, the free tier, the knowledge cutoff and the generation label — clarified one by one below.

Partly true "AA ranks it No. 2"

AA's own X post calls Astra the "#2 model in the Index", while this site's Nov 18 AI briefing says No. 3 of 91 models; the two wordings may come from different tiers or versions and are not yet reconciled. This page says "top 3"; Fable 5.2 (65) is undisputed at No. 1.

Not true "Free users can use GPT-6.2 Astra"

On launch day only Plus, Pro, Enterprise and Edu can use it; free users cannot. Don't build a topic on "gpt-6.2 astra free".

Clarified "The knowledge cutoff is June 1, 2026"

CyberQ's Nov 18 piece confirms from OpenAI's developer docs: Astra's knowledge cutoff is May 18, 2026, and June 1, 2026 refers to Sol; OpenAI's model page likewise lists Astra at May 18, 2026 (read directly by this site on Nov 18).

Partly true "A 20% price cut means Astra is cheaper to use"

The unit price did fall ($8/$40 vs $10/$50) and the per-task cost fell too ($4.00 vs $6.24). But Astra generated more tokens than GPT-6 Astra on the AA index task (65.8M vs 54.3M) — the saving comes from cheaper rates, not thriftier output.

Not true "GPT-6.2 is a brand-new generation"

GPT-6.2 is a point release one month after GPT-6.1; GPT-6.1 Terra had only just filled the gap between Sol and Luna a month earlier. Don't present it as a new generation.

Not true "There are open weights for self-hosting"

GPT-6.2 Astra is closed with no public weights; there is no self-hosting route — access is ChatGPT, Codex or the API only.

16FAQ

When was GPT-6.2 Astra released?

November 17, 2026 (Tuesday), generally available; the everyday tier GPT-6.2 Sol launched the same day, and both went live in ChatGPT Work, Codex and the API the same day. See section 05 for the timeline.

How much does GPT-6.2 Astra cost?

The API Standard rate is $8/$40 per 1M tokens (input/output), with cached input at $0.40; requests above 272K input tokens double the input and cache rates and cost 1.5× on output (i.e. $16/$60, per OpenAI's model page); Batch bills at 50% of Standard rates. See our pricing guide.

What does the AA 62 score mean?

Artificial Analysis' Intelligence Index v5.0.1 score is 62 (Max tier), in the top 3 of 91 models; Claude Fable 5.2 leads at 65; it is 5 points above GPT-6 Astra (57) and on the same score as Sol (57) at a higher cost. The rank wording is disputed (AA's X post says No. 2, this site's briefing says No. 3), so this page says "top 3". See our benchmarks article.

Is GPT-6.2 Astra better than Fable 5.2?

It trails by 3 points on the AA index (62 vs 65) but costs about 6.6× less per index task ($4.00 vs $26.54); on AA-3D, Astra leads with 1,318 Elo against Fable's 1,271. OpenAI's claim that Astra leads Fable by 5.5 points on Agents' Last Exam has no independent reproduction yet. See our comparison article.

Can free users use it?

No. On launch day only Plus, Pro, Enterprise and Edu can use it; free users cannot. See our access guide.

Is there an open-weights version?

No. GPT-6.2 Astra is closed with no public weights; access is ChatGPT, Codex or the API only, with no self-hosting route.

How large is the context window?

1,050,000 tokens (max input 922,000 tokens), with max output of 128,000 tokens. Requests above 272K input tokens double the input and cache rates and cost 1.5× on output (i.e. $16/$60).

Astra or Sol?

Pick Sol for everyday work: OpenAI's own claim puts Sol within 0.5 points of Astra on FrontierCode 2.0 at about a fifth of the cost ($2/$10 vs $8/$40); pick Astra for the hardest tasks, or when you need the Max effort tier and Ultrafast. See our comparison article.

17Sources

Every fact on this page comes from the public sources below (the OpenAI model page and the AA model page read directly on Nov 18, 2026).

Official and independent boards

English reporting and analysis

Chinese

  • This site's AI briefing, Nov 18, 2026, lead story (AA rank, prices, system-card thresholds, HealthAdminBench 94.3%) Nov 18

More indexes: All OpenAI models · All Language Models · All models