GPT-6.2 Astra vs Claude Fable 5.2
Updated 2026-11-18
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, eight days after Anthropic released Claude Fable 5.2. Astra replaces GPT-6 Astra as OpenAI’s top model and lists 20% lower, at $8 / $40 per million tokens, which is exactly what Anthropic charges for Fable 5.2. Both are reasoning models with a context window of about a million tokens, up to 128,000 output tokens, and text and image input. Whether you search it as GPT-6.2 Astra vs Fable 5.2 or Fable 5.2 vs Astra 6.2, this is the comparison.
We run apimodels.app, which offers both models through one API key below the vendors’ list prices. We did not run our own tests for this page. The independent scores come from Artificial Analysis and Vals AI, the vendor scores from OpenAI’s and Anthropic’s launch posts and docs, and every source is listed with its check date at the end. Where the two companies report different benchmarks we say so, rather than setting their numbers side by side as if they came from one test.
GPT-6.2 Astra
Wins terminal, science and documents at a fraction of the cost per task
- ·Terminal-Bench 4.0 64.1% vs 59.6%; Terminal-Bench-Science 78.1% vs 73.3%
- ·GDP.pdf 29.8% vs 24.2%; AA-3D Elo 1318 vs 1271
- ·65.8M output tokens to run the Artificial Analysis index, against 275M
- ·About $2.00 per Artificial Analysis task at our prices
Claude Fable 5.2
Leads both independent indexes
- ·Artificial Analysis Intelligence Index 65 (Astra 62)
- ·Vals Index 71.31%, first of 41 (Astra 67.66%, seventh)
- ·HLE-Diamond 80.2% vs 72.2%; AA-Omniscience 67 vs 50
- ·One price across the full 1M context; about $15.92 per task at our prices
Pick GPT-6.2 Astra for terminal, science and document agents and for anything you run at volume. Pick Claude Fable 5.2 when research-grade reasoning or factual accuracy decides the result and the cost per task comes second.
Quick verdict by use case
Each row points to the test behind it; the tables further down give the full numbers and who ran them.
| If you need… | Pick | Because |
|---|---|---|
| A coding or science agent that works in a terminal | GPT-6.2 Astra | Terminal-Bench 4.0 64.1% vs 59.6%; Terminal-Bench-Science 78.1% vs 73.3% (Artificial Analysis) |
| Repository changes held to a maintainer’s standard | Claude Fable 5.2 | FrontierCode-AA 2.0 at max effort 48.8% vs 45.4%; on OpenAI’s own chart 49.1% (xhigh) vs 44.2% |
| Long multi-step agent workflows | GPT-6.2 Astra | Agents’ Last Exam 77.6% vs 72.8% at max effort (OpenAI’s chart); AutomationBench-AA 77.2% vs 76.4% |
| Hard research and knowledge questions | Claude Fable 5.2 | HLE-Diamond 80.2% vs 72.2%; CritPt 76.5% vs 71.5%; AA-Omniscience 67 vs 50 |
| Reports and PDFs as the deliverable | GPT-6.2 Astra | GDP.pdf 29.8% vs 24.2% (Artificial Analysis) |
| Prompts above 272K tokens | Claude Fable 5.2 | One price across 1M; Astra bills input ×2 and output ×1.5 above 272K and caps input at 922K |
| High volume, or a fixed cost per task | GPT-6.2 Astra, or GPT-6.2 Sol | $4.00 vs $26.54 per Artificial Analysis task at list; 10.7 vs 40.7 minutes |
| Exploit work, audio or video input | Neither | Both restrict high-risk cyber requests; both take text and images only |
Side by side: price, limits, controls and scores
The list price per token is now identical, so the differences that change a bill are elsewhere: Fable 5.2 charges half as much for cached input and keeps one rate across its whole window, while Astra writes far fewer output tokens per task. Through apimodels.app, build GPT-6.2 agent loops on function calling: OpenAI’s hosted Responses tools, such as computer use, file search and hosted shell, have not been verified on our endpoint. Bold marks the better value in each row.
| GPT-6.2 Astra | Claude Fable 5.2 | |
|---|---|---|
| Maker, release | OpenAI, 17 Nov 2026 | Anthropic, 9 Nov 2026 |
| List price per 1M tokens (input / output) | $8 / $40 | $8 / $40 |
| Cached input, list | $0.40 | $0.20 |
| Cache write, list | $10 (30-minute retention) | $10 (5-minute), $16 (1-hour) |
| Prompts above 272K tokens | Input and cache ×2, output ×1.5 | Same price across the full 1M |
| Price on apimodels.app (input / output / cached) | $4 / $20 / $0.20 | $4.80 / $24 / $0.12 |
| Example job, 50K in and 5K out: list / apimodels.app | $0.60 / $0.30 | $0.60 / $0.36 |
| Cost per Artificial Analysis task, list | $4.00 | $26.54 |
| Output tokens to run the Artificial Analysis index | 65.8M | 275M |
| Context window | 1.05M (input up to 922K) | 1M |
| Max output | 128K | 128K |
| Knowledge cutoff | 18 May 2026 | June 2026 |
| Input / output | Text, image / text | Text, image / text |
| Reasoning control | reasoning_effort low to max; none is rejected | Thinking always on; effort low to max, API default high |
| Tool calling (vendor API) | Function calling; OpenAI-hosted tools on the Responses API | Tool use; forced tool choice not supported |
| Endpoints on apimodels.app | /v1/chat/completions, /v1/responses | /v1/messages, /v1/chat/completions |
| Tools on apimodels.app | Function calling, including parallel calls; OpenAI’s hosted tools not verified here | Tool use; a forced tool choice is sent as auto plus an instruction |
| Not offered on apimodels.app | Batch API, Fast and Ultrafast modes | Batch API, Fast mode |
| Artificial Analysis Intelligence Index | 62 | 65 |
| Vals Index | 67.66% (7th of 41) | 71.31% (1st) |
| Output speed (Artificial Analysis) | 64 tokens/s | 112 tokens/s |
| Time per Artificial Analysis task | 10.7 min | 40.7 min |
| Cybersecurity safeguards | Tightened for high-risk tasks; OpenAI says this limits some legitimate vulnerability research | Blocks most cybersecurity tasks; blocked requests are answered by Opus 4.8 |
Independent benchmarks
Artificial Analysis and Vals AI both put Fable 5.2 first overall and GPT-6.2 Astra behind it. Inside the Artificial Analysis index the split is closer than the headline: of the nine evaluations it reports for both at max effort, Fable 5.2 leads five and Astra four. Fable’s leads are on knowledge and research (HLE-Diamond, CritPt, AA-Omniscience), on AA-Briefcase knowledge work and on FrontierCode; Astra’s are on Terminal-Bench 4.0, Terminal-Bench-Science, GDP.pdf and, narrowly, AutomationBench-AA. Artificial Analysis’s separate AA-3D and AA-Robotics tests also favour Astra.
The effort setting moves some rows. At xhigh, Astra scores 49.0% on FrontierCode-AA against Fable’s 47.6%, the reverse of the max-effort result. Artificial Analysis ran Fable 5.2 with Anthropic’s default fallback switched on.
| Benchmark (max effort) | Run by | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|---|
| Intelligence Index v5.0.1 | Artificial Analysis | 62 | 65 |
| Vals Index | Vals AI | 67.66% (7th of 41) | 71.31% (1st of 41) |
| Terminal-Bench 4.0 | Artificial Analysis | 64.1% | 59.6% |
| Terminal-Bench-Science 0.1 | Artificial Analysis | 78.1% | 73.3% |
| GDP.pdf | Artificial Analysis | 29.8% | 24.2% |
| AutomationBench-AA | Artificial Analysis | 77.2% | 76.4% |
| FrontierCode-AA 2.0 | Artificial Analysis | 45.4% | 48.8% |
| HLE-Diamond | Artificial Analysis | 72.2% | 80.2% |
| CritPt v2 (hard subset) | Artificial Analysis | 71.5% | 76.5% |
| AA-Omniscience v1.1 | Artificial Analysis | 50 | 67 |
| AA-Briefcase v1.1 (Elo) | Artificial Analysis | 1643 | 1785 |
| AA-3D (Elo, blind human votes) | Artificial Analysis | 1318 | 1271 |
| AA-Robotics v0.1 | Artificial Analysis | 79 | 75 |
| Cost per Vals Index test | Vals AI | $10.68 | $17.94 |
| Time per Vals Index test | Vals AI | 52 min 16 s | 43 min 42 s |
What OpenAI reports, and what Anthropic has not answered yet
OpenAI’s launch post plots Fable 5.2 next to GPT-6.2 Astra on nine charts, one of which repeats the Artificial Analysis index above. On the other eight, by OpenAI’s numbers, Astra leads six and Fable 5.2 two: FrontierCode 2.0 and BenchCAD. Two of OpenAI’s footnotes matter: on Agents’ Last Exam it left out nine Fable 5.2 tasks that were answered by Anthropic’s fallback model, and on Terminal-Bench-Science its Fable 5.2 point is Artificial Analysis’s run, while Anthropic reports 74.6% in Claude Code at max effort. GDPval v2 is OpenAI’s internal test. Read all of these as OpenAI’s claims.
Anthropic’s Fable 5.2 launch table came out on 9 November and compares against the older GPT-6 Astra, so there is no Anthropic number for GPT-6.2 Astra yet. The Fable 5.2 results it reports on tests OpenAI did not chart, such as FrontierSWE v2 at 69.9%, OSWorld 2.1 at 89.3% (offline set, partial credit) and Remote Labor Index at 41.3%, have no GPT-6.2 Astra counterpart to compare with.
| Benchmark (OpenAI’s launch charts) | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|
| Agents’ Last Exam v1.1 | 78.3% at xhigh, $2.89 a task (77.6% at max) | 72.8% at max, $60.05 a task |
| Terminal-Bench-Science 0.1 | 77.1% at max, $7.15 | 73.3% at max, $45.21 (Artificial Analysis run) |
| Vals Science Index v1 | 68.4%, $4.17 | 63.7%, $8.79 |
| AutomationBench 1.0.6 | 52.4% at max, $0.83 | 51.4% at max, $4.64 |
| Harvey LAB-AA v1.1 (legal) | 15.0% at max, $6.35 | 7.2% at max, $39.84 |
| GDPval v2 (OpenAI internal) | 50.4% at max, 35,812 output tokens | 44.8% at max, 222,344 output tokens |
| FrontierCode 2.0 | 44.2% at max, $5.97 | 49.1% at xhigh, $14.19 |
| BenchCAD 2.1 Core | 82.5% at max, $5.43 | 83.5% at max, $50.77 |
Price and what a job costs
With the same $8 / $40 list price, a job with the same token counts costs the same on both: 50,000 input and 5,000 output tokens is $0.60 either way. Three things then pull the bills apart. Thinking tokens bill as output, and Astra writes far fewer of them: 65.8 million output tokens to run the whole Artificial Analysis index at max effort against 275 million for Fable 5.2, which made a task $4.00 against $26.54. Vals measured a narrower gap on its own tasks, $10.68 against $17.94 per test. Cached input is $0.40 on Astra and $0.20 on Fable 5.2, which matters for agents that resend a long context. And above 272,000 input tokens Astra bills the whole request at twice the input rate and 1.5 times the output rate, while Fable 5.2 keeps one price to 1M.
On apimodels.app GPT-6.2 Astra is 50% below OpenAI’s list and Fable 5.2 40% below Anthropic’s, on every token type, so the measured task costs scale directly. Astra’s 272K step applies here too. Failed requests are not charged.
| Job | Astra, list | Fable 5.2, list | Astra on apimodels.app | Fable 5.2 on apimodels.app |
|---|---|---|---|---|
| 50K input, 5K output | $0.60 | $0.60 | $0.30 | $0.36 |
| 400K input, 10K output (over 272K) | $7.00 | $3.60 | $3.50 | $2.16 |
| One Artificial Analysis task | $4.00 | $26.54 | about $2.00 | about $15.92 |
| One Vals Index test | $10.68 | $17.94 | about $5.34 | about $10.76 |
Speed
Fable 5.2 streams faster once it starts, 112 tokens per second against 64. Astra starts sooner and, because it writes less, finished an Artificial Analysis task in 10.7 minutes against 40.7. At medium effort Astra’s first token arrives in 2.7 seconds against 9.9. Vals measured the opposite on its own task set, 52 minutes per test for Astra and 44 for Fable 5.2, so time your own workload before you route by speed.
| Measured by Artificial Analysis | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|
| Output speed | 64 tokens/s | 112 tokens/s |
| Time to first token, max effort | 147.65 s | 214.95 s |
| Time to first token, medium effort | 2.70 s | 9.87 s |
| End-to-end time, single request, max effort | 316.23 s | 397.83 s |
| Time per index task, max effort | 10.7 min | 40.7 min |
GPT-6.2 Sol, the lower-priced option
GPT-6.2 Sol is OpenAI’s everyday model from the same release, at a quarter of Astra’s list price: $2 / $10 per million tokens, $1 / $5 on apimodels.app. It has the same 1.05M context, 128K output and text-and-image input, with a 1 June 2026 knowledge cutoff. It scores 57 on Artificial Analysis, five points under Astra and eight under Fable 5.2, at $1.01 per task.
On some agent work it holds its own. OpenAI’s chart has Sol within half a point of Astra on FrontierCode 2.0 (43.8% against 44.2%) at about a fifth of the cost, and level with Fable 5.2 on Agents’ Last Exam (73.0% at $1.32 a task against 72.8% at $60.05). The caveat comes from Artificial Analysis: Sol’s Coding Agent Index fell three points against GPT-6.1 Sol, driven by a 14-point drop on SWE-Atlas-QnA-AA, so test it on your own code questions before switching a coding agent to it.
| GPT-6.2 Sol | GPT-6.2 Astra | Claude Fable 5.2 | |
|---|---|---|---|
| List price per 1M tokens | $2 / $10 | $8 / $40 | $8 / $40 |
| Price on apimodels.app | $1 / $5 | $4 / $20 | $4.80 / $24 |
| Artificial Analysis Intelligence Index | 57 | 62 | 65 |
| Cost per Artificial Analysis task, list | $1.01 | $4.00 | $26.54 |
| Vals Index | 66.06% (9th) | 67.66% (7th) | 71.31% (1st) |
| Cost per Vals Index test | $3.18 | $10.68 | $17.94 |
| FrontierCode 2.0 (OpenAI’s chart) | 43.8% | 44.2% | 49.1% |
| Agents’ Last Exam v1.1 at max (OpenAI’s chart) | 73.0% | 77.6% | 72.8% |
| Output speed (Artificial Analysis) | 85 tokens/s | 64 tokens/s | 112 tokens/s |
When neither is the right answer
Volume work where a few points do not matter: GPT-6.2 Sol, above. An Artificial Analysis task costs a quarter of Astra’s and about a twenty-sixth of Fable 5.2’s.
A Claude-based stack at half the token price: Claude Opus 5.5, at $4 / $20 list ($2.40 / $12 on apimodels.app). It scores 59 on Artificial Analysis at max effort, at $10.04 a task, and 70.14% on the Vals Index, fourth and above Astra. It runs in Claude Code and on the Anthropic Messages API unchanged.
Security research and exploit development: both models restrict it. Anthropic says Fable 5.2’s safeguards block most cybersecurity tasks and hand them to Opus 4.8; OpenAI says GPT-6.2’s tighter safeguards also limit some legitimate vulnerability research, and points defenders to its separate Daybreak programme, which does not include GPT-6.2 yet.
Audio or video as input: neither accepts it; both take text and images only. Gemini 4 Carbon accepts text, image, video, audio and PDF.
Caveats
GPT-6.2 Astra is one day old, and more independent results will arrive over the next weeks; Anthropic has not yet published any comparison with it. Artificial Analysis ran Fable 5.2 with Anthropic’s fallback on, and OpenAI removed the fallback-answered tasks from Fable’s Agents’ Last Exam score, so the two sources handle the same issue differently. Artificial Analysis measured a 59% hallucination rate for Astra at max effort on AA-Omniscience (55% for Sol) and gives Fable 5.2 the higher score on that test, 67 against 50. OpenAI also reports that both GPT-6.2 models were more likely than GPT-6 Astra to leave out blockers such as a missing file or an unavailable search tool in two of its tests, so check tool results in your own code rather than trusting the summary.
Sources
All read on 18 November 2026. Independent scores, speeds and task costs: Artificial Analysis (comparison page, model pages and its GPT-6.2 launch article) and Vals AI (Vals Index, updated 16 November, and the GPT-6.2 Astra page). Vendor figures: OpenAI’s GPT-6.2 launch post and model pages, and Anthropic’s Fable 5.2 launch post, models overview and pricing page. apimodels.app prices are our own. We did not run paid tests for this page.
Início rápido
cURL
curl https://api.apimodels.app/v1/chat/completions \
-H "Authorization: Bearer $APIMODELS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6.2-astra",
"messages": [{"role": "user", "content": "Read the attached test log and tell me which change broke the build: ..."}]
}'
# same key, same endpoint: change "model" to "claude-fable-5-2" for Fable 5.2, or "gpt-6.2-sol" for SolPerguntas frequentes
Is GPT-6.2 Astra better than Claude Fable 5.2?
Not overall. Claude Fable 5.2 scores higher on both independent indexes: 65 against 62 on Artificial Analysis, and 71.31% against 67.66% on the Vals Index, where it ranks first and Astra seventh. GPT-6.2 Astra is better at specific work: terminal coding (Terminal-Bench 4.0, 64.1% against 59.6%), science agents (Terminal-Bench-Science, 78.1% against 73.3%) and generated documents (GDP.pdf, 29.8% against 24.2%). It also finishes an Artificial Analysis task for $4.00 against $26.54 at list prices.
Which AI model is better than Claude Fable 5.2?
On the overall leaderboards, none yet. As of 18 November 2026 Claude Fable 5.2 is first on both the Artificial Analysis Intelligence Index (65) and the Vals Index (71.31%). GPT-6.2 Astra is the closest OpenAI model, at 62 and 67.66%, and beats Fable 5.2 on several single tests, including Terminal-Bench 4.0, Terminal-Bench-Science, GDP.pdf and AA-3D. If you mean the original Claude Fable 5, Fable 5.1 and Fable 5.2 are newer versions of it.
How much better is GPT-6.2 Astra than GPT-6 Astra?
About five points on Artificial Analysis’s Intelligence Index (62 against 57) and 2.6 points on the Vals Index (67.66% against 65.02%). OpenAI reports larger gains on its own charts: 5.2 points on Agents’ Last Exam at under a third of the cost, 9 points on Terminal-Bench-Science, and 6.2 points on its internal GDPval v2 with 22% fewer output tokens. The list price also fell 20%, to $8 / $40 per million tokens, and cached input fell 60%, to $0.40.
What are GPT-6.2 Astra’s benchmark scores?
Independent results at max effort: 62 on the Artificial Analysis Intelligence Index (Claude Fable 5.2: 65), 67.66% on the Vals Index (seventh of 41), 64.1% on Terminal-Bench 4.0, 78.1% on Terminal-Bench-Science 0.1, 72.2% on HLE-Diamond and an AA-3D Elo of 1318, first on that leaderboard. OpenAI’s own charts add 77.6% on Agents’ Last Exam v1.1 (78.3% at xhigh), 82.5% on BenchCAD 2.1 Core and 15.0% on Harvey LAB-AA v1.1.
GPT-6.2 Astra vs Claude Opus 5.5: which should I use?
GPT-6.2 Astra scores higher on Artificial Analysis (62 against 59 at max effort), costs less per task there ($4.00 against $10.04 at list) and leads on OpenAI’s Agents’ Last Exam chart (77.6% against 71.6%). Opus 5.5 ranks higher on the Vals Index (70.14%, fourth, against 67.66%, seventh) and has half Astra’s list price per token, $4 / $20 against $8 / $40. On apimodels.app Opus 5.5 is $2.40 / $12 and Astra $4 / $20. Choose Opus 5.5 if your tooling is built on Claude Code or the Anthropic Messages API; otherwise Astra did more per dollar on Artificial Analysis’s tasks.
GPT-6.2 Sol vs Claude Opus 5.5: which should I use?
GPT-6.2 Sol costs far less: $2 / $10 against $4 / $20 at list ($1 / $5 against $2.40 / $12 on apimodels.app), and an Artificial Analysis task costs $1.01 against $10.04. Opus 5.5 scores higher overall, 59 against 57 on Artificial Analysis and 70.14% against 66.06% on Vals, while Sol edges it on OpenAI’s Agents’ Last Exam chart (73.0% against 71.6%). Use Sol for high-volume work and Opus 5.5 when the extra points matter or you work in Claude Code.
Is GPT-6.2 Sol close to Claude Fable 5.2?
On overall scores, no: Sol has 57 on Artificial Analysis against Fable 5.2’s 65, and 66.06% on the Vals Index against 71.31%. On some agent tests it is level at a small fraction of the cost: on OpenAI’s Agents’ Last Exam chart Sol scores 73.0% at $1.32 a task and Fable 5.2 72.8% at $60.05. The list price is $2 / $10 against $8 / $40.
