Quick Answer:
Claude Fable 5.2 is still the more capable model: it leads the Artificial Analysis Intelligence Index 65 to 62, ranks 1st on the Vals Index against Astra's 7th, and has the higher ceiling in coding and open-ended research. GPT-6.2 Astra is the better value: both cost exactly $8/$40 per million tokens, but Astra uses far fewer tokens, so Artificial Analysis puts its cost per task at about $4 against about $27 for Fable at maximum effort. Astra also leads on Terminal-Bench-Science, legal work, 3D design and robotics. Choose Fable for the hardest coding and research; choose Astra for science and professional work at volume.
Eight days apart, OpenAI and Anthropic released their flagship models at exactly the same list price. That makes this the cleanest head-to-head of the year: same $8 in, same $40 out. What separates them is not the price per token but how many tokens each spends to get the job done, and where each one wins.
Claude Fable 5.2 launched on 09/11/2026 and GPT-6.2 Astra on 17/11/2026. This comparison pulls together every published number we could verify: independent results from Artificial Analysis, Vals AI and Epoch AI; OpenAI's launch charts and system card; and Anthropic's launch figures. Values read from a vendor chart are marked with "~". Every chart is interactive: hover or tab to a bar or point for its value, and open "Show data table" for the raw figures. For the full reviews, see our GPT-6.2 Astra and Sol review and our Claude Fable 5.2 review.
65 vs 62
AA Intelligence Index
Fable 5.2 leads
~$27 vs ~$4
AA cost per task (max)
Astra about 7x cheaper
$8 / $40
List price, both models
per million tokens in / out
174 = 174
Epoch Capabilities Index
A dead heat
Source: Artificial Analysis, Epoch AI, OpenAI and Anthropic pricing, 17-18/11/2026
Wes Roth on the GPT-6.2 launch, the reaction to Astra and how it stacks up against Anthropic's Fable 5.2.
The Verdict in 60 Seconds
| Category | Winner | Why |
|---|---|---|
| Overall capability | Fable 5.2 | 65 vs 62 on Artificial Analysis; 1st vs 7th on Vals |
| Cost per task | Astra | ~$4 vs ~$27 on Artificial Analysis at max effort |
| Science benchmarks | Split | Astra: Terminal-Bench-Science, Vals Science. Fable: CritPt, HLE-Diamond |
| Open-ended research | Fable 5.2 | Epoch AI's real research tasks |
| Peak coding | Fable 5.2 | 49.1% vs ~44% on FrontierCode 2.0 |
| Coding value | Astra (and Sol) | ~44% at ~$6, or ~$1.30 with Sol |
| Legal and professional | Astra | ~15% vs ~7% on Harvey LAB-AA (OpenAI chart) |
| 3D design and robotics | Astra | AA-3D Elo 1,318 vs 1,271; AA-Robotics 79 vs 75 |
| Factual reliability | Fable 5.2 | Leads AA-Omniscience; Astra hallucination rate 59% |
Specs and Pricing Side by Side
| GPT-6.2 Astra | Claude Fable 5.2 | |
|---|---|---|
| Released | 17/11/2026 | 09/11/2026 |
| Input price (per million tokens) | $8 (~£6.32) | $8 (~£6.32) |
| Output price (per million tokens) | $40 (~£31.60) | $40 (~£31.60) |
| Cached input | $0.40 | See Anthropic pricing |
| Fast tier | Ultrafast (Pro 500 and API) | Fast mode: $16 / $80, up to 2.5x faster |
| Previous generation price | $10 / $50 (GPT-6 Astra) | $10 / $50 (Fable 5.1) |
| Consumer access | ChatGPT Work and Codex: Plus, Pro, Enterprise, Edu | Claude apps and major clouds, all users |
| Restricted sibling | Daybreak access for vetted defenders | Mythos 5.2 for vetted cyber and life-science organisations |
Both labs cut their flagship price by exactly 20% this month, from $10/$50 to $8/$40, and landed on the same number. That is unusual and it makes the comparison simple: whichever model finishes your task with fewer tokens is the cheaper one.
Overall Intelligence: Independent Indexes
Three independent evaluators have now scored both models. Two put Fable ahead; one calls it a draw.
Artificial Analysis Intelligence Index v5
Composite of nine evaluations, 0-100, max reasoning effort. Higher is better. (Epoch's Capabilities Index, on a different scale, scores both at 174.)
- GPT-6.2 Astra
- Claude Fable 5.2
- GPT-6.2 Sol
Intelligence Index score
Show data table
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 | GPT-6.2 Sol |
|---|---|---|---|
| Intelligence Index score | 62 | 65 | 57 |
Source: Artificial Analysis Intelligence Index v5.0.1, 17/11/2026; Epoch AI (via The Decoder)
On the Vals Index, Fable 5.2 ranks 1st and Astra 7th. Vals did not publish Fable's cost per task alongside, but it measured Astra at $10.68 per task with the fewest output tokens in its top 10, an average of 91,000 per task. On Artificial Analysis, Astra (max) sits second of all models, ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59). Astra is up 5 points on GPT-6 Astra; Fable 5.2 sits 6 points above Anthropic's own Opus 5.5.
Epoch AI also measured how much reasoning a model can do inside a single output token, without a visible chain of thought. Astra scored 12.3 steps against 4.2 for Claude Fable 5.1 (Epoch did not publish a Fable 5.2 figure). It is a striking capability gap, and one that matters for safety monitoring, as we discuss below.
Cost per Task: The Real Difference
Same list price, very different bills. Artificial Analysis runs its whole Intelligence Index suite on each model and reports the weighted average cost per task.
Cost per Artificial Analysis Intelligence Index task
US dollars at maximum reasoning effort. Lower is better.
- GPT-6.2 Astra
- Claude Fable 5.2
Average cost per task (max effort)
Show data table
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|
| Average cost per task (max effort) | ~$4 | ~$27 |
Source: Artificial Analysis cost per task, as reported by Trending Topics and Artificial Analysis's launch chart, 17/11/2026
At roughly $4 against $27, Astra completes the same suite for about a seventh of the cost while scoring 3 points lower. Fable 5.2 at Xhigh effort is cheaper than at max (around $15 on Artificial Analysis's chart) and scores almost as well, but is still several times Astra's cost. For high-volume workloads, the gap compounds: 10,000 tasks a month is roughly $40,000 on Astra against roughly $270,000 on Fable at max effort, before caching.
The caveat: cost per task depends heavily on the task. On FrontierCode, Fable's cheaper effort settings are competitive (see below), and on Anthropic's own figures Max effort scored lower than Xhigh at three times the price. Pick the effort level, not just the model.
Science Benchmarks
OpenAI calls Astra "the world's best model for scientists". Anthropic's Fable launch featured two solved open problems in theoretical physics. The benchmarks split between them.
Science benchmarks with published scores for both models
Higher is better. ~ marks values read from OpenAI's launch chart.
- GPT-6.2 Astra
- Claude Fable 5.2
Terminal-Bench-Science 0.1 · OpenAI's chart (Fable on mini-swe-agent harness)
Terminal-Bench-Science 0.1 · Anthropic's own run for Fable
GPQA Diamond (no tools) · Astra from OpenAI system card; Fable not published
Show data table
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|
| Terminal-Bench-Science 0.1 | ~77% | ~73% |
| Terminal-Bench-Science 0.1 | ~77% | 74.6% |
| GPQA Diamond (no tools) | 97% | n/a |
Source: OpenAI launch chart (via OfficeChai), OpenAI GPT-6.2 system card, Anthropic Fable 5.2 launch figures
Where Astra leads: Terminal-Bench-Science 0.1, on either Fable figure, and the Vals Science Index, where Astra ranks first and also tops VCB 1-100. On FrontierMath's Erdős problem set, Epoch reports Astra produced Lean-verified solutions to 7 of 68 problems, more than double the previous record of 3, including one previously unknown result.
Where Fable leads: Artificial Analysis says Astra ranks second behind Fable 5.2 on CritPt v2 (hard physics reasoning), HLE-Diamond and AA-Omniscience; it has not published the individual scores. Anthropic reports 74.2% on Humanity's Last Exam with tools for Fable 5.2. And on Epoch AI's real-world research tasks, Fable came out clearly ahead: Astra missed the central question on simpler projects and skipped key estimates.
Our reading: Astra is the better tool for well-specified computational science in a terminal: running analyses, writing solvers, checking results. Fable is the better research partner for open-ended questions where judging what matters is the hard part. For more on Astra's maths results, see Astra's maths breakthroughs.
Coding Benchmarks
Cognition's FrontierCode 2.0 measures merge-ready changes to real codebases. Both labs published results at several effort levels, so we can plot what each extra point costs. Fable's points use Anthropic's published figures, which match OpenAI's chart; Astra and Sol are read from OpenAI's chart.
FrontierCode 2.0: score vs cost per task
Each point is a reasoning-effort setting. Up and to the left is better.
- GPT-6.2 Astra
- Claude Fable 5.2
- GPT-6.2 Sol
Show data table
| Model | Setting | FrontierCode 2.0 score | Cost per task |
|---|---|---|---|
| GPT-6.2 Astra | low effort | ~38.1% | ~$1.4 |
| GPT-6.2 Astra | - | ~42.3% | ~$2.5 |
| GPT-6.2 Astra | - | ~44.3% | ~$4.2 |
| GPT-6.2 Astra | max | ~44.3% | ~$5.9 |
| Claude Fable 5.2 | Low | ~34.2% | ~$2.1 |
| Claude Fable 5.2 | Medium | 37.6% | $2.5 |
| Claude Fable 5.2 | High | 45.3% | $6.7 |
| Claude Fable 5.2 | Xhigh | 49.1% | $14 |
| Claude Fable 5.2 | Max | 45.7% | $45 |
| GPT-6.2 Sol | low effort | ~33.2% | ~$0.39 |
| GPT-6.2 Sol | - | ~38.4% | ~$0.67 |
| GPT-6.2 Sol | - | ~39.9% | ~$0.97 |
| GPT-6.2 Sol | max | ~43.9% | ~$1.3 |
Source: Anthropic Fable 5.2 launch figures; OpenAI GPT-6.2 launch chart (Astra and Sol values read from chart)
The chart tells the coding story clearly. Fable 5.2 has the highest ceiling, 49.1% at Xhigh for $14.19 a task, about 5 points above anything OpenAI ships. At equal spend, the two are close: Fable at High scores 45.3% for $6.72, Astra about 44% for about $6. GPT-6.2 Sol is the value outlier, reaching Astra's score for about $1.30. And Max effort is a trap on Fable: it scores below Xhigh at three times the cost.
Independent agentic coding results lean towards Fable. On Vals AI's Terminal-Bench 4 ranking, Astra is 5th. On Artificial Analysis's Coding Agent Index, Astra did not improve on GPT-6 Astra at all. Developer Theo Browne reported that on a TypeScript-to-Rust rewrite, Astra ran for seven days and produced a compiler 56% slower. For long autonomous coding runs, Fable is the safer pick; see our best AI coding agents guide for the tools around them.
Matthew Berman tests Claude Fable 5.2 on coding tasks and explains why its launch numbers stood out.
Agents and Professional Work
OpenAI's launch charts include Fable 5.2 on four agent and professional benchmarks. These are OpenAI's runs, and Fable may score differently in Anthropic's own harness, so treat them as indicative.
Agent and professional benchmarks (OpenAI's runs)
Best score for each model. Higher is better. All values read from OpenAI's launch charts.
- GPT-6.2 Astra
- Claude Fable 5.2
- GPT-6.2 Sol
Agents' Last Exam v1.1
BenchCAD 2.1 Core
Harvey LAB-AA v1.1 (legal)
Show data table
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 | GPT-6.2 Sol |
|---|---|---|---|
| Agents' Last Exam v1.1 | ~78% | ~73% | ~73% |
| BenchCAD 2.1 Core | ~83% | ~84% | ~73% |
| Harvey LAB-AA v1.1 (legal) | ~15% | ~7% | ~11% |
Source: OpenAI GPT-6.2 launch charts (values via OfficeChai and The Decoder)
Agents' Last Exam v1.1: score vs cost per task
Astra and Sol effort sweeps; Fable 5.2's best published point. Up and to the left is better.
- GPT-6.2 Astra
- Claude Fable 5.2
- GPT-6.2 Sol
Show data table
| Model | Setting | Agents' Last Exam score | Cost per task |
|---|---|---|---|
| GPT-6.2 Astra | low effort | ~65.8% | ~$0.40 |
| GPT-6.2 Astra | - | ~72.1% | ~$1 |
| GPT-6.2 Astra | - | ~76% | ~$2 |
| GPT-6.2 Astra | xhigh | ~78.4% | ~$2.9 |
| GPT-6.2 Astra | max | ~77.6% | ~$4.5 |
| Claude Fable 5.2 | best published | ~72.8% | ~$60 |
| GPT-6.2 Sol | low effort | ~66.7% | ~$0.20 |
| GPT-6.2 Sol | - | ~70.3% | ~$0.40 |
| GPT-6.2 Sol | max | ~73% | ~$1.2 |
Source: OpenAI GPT-6.2 launch chart (values read from chart)
Agents' Last Exam is OpenAI's strongest result: Astra scores about 5.5 points above Fable for about 5% of the cost per task. OpenAI notes that Fable's leaderboard entry excludes nine tasks that used a fallback model. On Harvey's LAB-AA legal tasks, Astra roughly doubles Fable's pass rate at about a sixth of the cost, and Vals AI independently confirmed that Astra roughly doubled its predecessor on legal work. On BenchCAD, the two are level, with Fable about a point ahead at roughly nine times the cost.
Anthropic's own agent figures for Fable are strong on tests OpenAI did not run: 51.4% on AutomationBench for business workflows, 89.3% on OSWorld 2.1 computer use and 41.3% on the Remote Labor Index. Without matching Astra numbers, we cannot call those categories.
3D, CAD and Robotics
Artificial Analysis robotics benchmark
AA-Robotics v0.1 score. Higher is better.
- GPT-6.2 Astra
- Claude Fable 5.2
- GPT-6.2 Sol
AA-Robotics v0.1
Show data table
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 | GPT-6.2 Sol |
|---|---|---|---|
| AA-Robotics v0.1 | 79 | 75 | 78 |
Source: Artificial Analysis, 17/11/2026 (via The Decoder)
On Artificial Analysis's blinded human-preference 3D test, AA-3D, Astra leads with an Elo of 1,318 against Fable's 1,271, with Sol at 1,259. A 47-point Elo gap means human judges preferred Astra's 3D output in roughly 57% of head-to-head comparisons. Astra also leads AA-Robotics. Fable's robotics claim rests on Anthropic's own RoboDojo Sim result of 64.9%, which OpenAI has not reported for Astra.
Results Only One Lab Published
Many launch numbers have no counterpart from the other side. We list them here rather than invent a comparison.
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 | Source |
|---|---|---|---|
| FrontierSWE v2 | n/a | 69.9% | Anthropic |
| OSWorld 2.1 (offline, partial credit) | n/a | 89.3% | Anthropic |
| AutomationBench | n/a | 51.4% | Anthropic |
| Remote Labor Index | n/a | 41.3% | Anthropic |
| Humanity's Last Exam (with tools) | n/a | 74.2% | Anthropic |
| Chartography (no tools) | n/a | 73.4% | Anthropic |
| RoboDojo Sim | n/a | 64.9% | Anthropic |
| GPQA Diamond (no tools) | 97.0% | n/a | OpenAI system card |
| MLE-Bench Revised (percentile) | 96.06 | n/a | OpenAI system card |
| FrontierMath Erdős (Lean-verified) | 7 / 68 | n/a | Epoch AI |
| Vals AI cost per task | $10.68 | n/a | Vals AI |
Safety: Two Different Philosophies
The starkest difference between the two models is not a benchmark but a decision. Anthropic says it deliberately held two AI research skills in Fable 5.2 at Opus 5.5 level, choosing not to make the model better at accelerating AI development. OpenAI went the other way: its system card says Astra is the first released model to reach its High threshold for AI self-improvement, and that it can meaningfully speed up AI research. On NanoGPT training optimisation, Astra saved 42.87% of baseline training time against 7.73% for GPT-6 Astra.
| Area | GPT-6.2 Astra | Claude Fable 5.2 |
|---|---|---|
| AI research acceleration | High (first released model) | Deliberately capped at Opus 5.5 level on two skills |
| Cybersecurity | Critical; tighter refusals; Daybreak for vetted defenders | Classifiers block most cyber tasks; Mythos 5.2 for vetted organisations |
| Biology | High; below Critical thresholds | Topic classifiers; Anthropic says about 40% fewer biology blocks in this release |
| Oversight tooling | Control Room can halt research workloads | Blocked requests fall back to older models |
| External testers named | UK AISI, SecureBio, Irregular and others | See Anthropic's system card |
For UK users, the practical difference is in refusals. Fable routes blocked cyber and biology queries to older models, and Anthropic says this release triggers about 55% fewer blocks overall. Astra refuses higher-risk cyber help by default unless you are approved for Daybreak. Security teams should test both against their own workloads. Our Astra cyber analysis and Claude Mythos explainer cover the background.
AI Revolution X, 19/11/2026: the Astra versus Fable 5.2 benchmark and cost claims, the tester backlash and the safety findings from OpenAI's own testing.
Reliability and Honesty
Artificial Analysis measured Astra's hallucination rate on its deliberately hard AA-Omniscience questions at 59% at maximum effort, and says both Fable 5.2 and Gemini 4 Carbon did better, without publishing their figures. Fable also outranks Astra on the AA-Omniscience knowledge score. OpenAI's system card adds an honesty regression: Astra is more likely to omit known obstacles, such as a missing file, from its reports, in about 22% of cases according to Trending Topics, twice GPT-6 Astra's rate. A good developer prompt reduces this.
Anthropic's figures for the Fable and Mythos 5.2 model show an automated behavioural audit score of 2.67 (lower is better) against 2.79 for Opus 5.5, and sandbox escape attempts using outside credentials in 3.1% of scenarios, down from 21% for Mythos 5.1. These are different tests, so they cannot be compared directly with OpenAI's, but both labs are now publishing measurable honesty and control data. Read our report on Astra's covert underperformance for the earlier history.
Which Should You Choose?
| If you need... | Choose |
|---|---|
| The best possible answer, cost no object | Claude Fable 5.2 at Xhigh |
| Long autonomous coding runs | Claude Fable 5.2 |
| Everyday coding at the lowest cost | GPT-6.2 Sol |
| Computational science, data analysis in a terminal | GPT-6.2 Astra |
| Open-ended research and judgement | Claude Fable 5.2 |
| Legal, finance and templated documents at volume | GPT-6.2 Astra |
| 3D, CAD and robotics work | GPT-6.2 Astra |
| Factual Q&A with fewer hallucinations | Claude Fable 5.2 |
| Fastest responses | Astra Ultrafast or Fable Fast mode (test both) |
Many teams will use both: Fable 5.2 for the hardest problems and final review, Astra or Sol for the bulk of the work. Our GPT-6 Sol vs Opus 5.5 comparison shows the same split one tier down.
How We Compiled These Numbers
- Exact figures come from published tables: Artificial Analysis index scores, AA-3D and AA-Robotics; Epoch's ECI; Vals AI rankings and costs; OpenAI's system card; Anthropic's FrontierCode effort table and launch benchmarks.
- Values marked ~ were read from OpenAI's published charts and are accurate to about one percentage point and 10% on cost.
- Vendor runs of rival models (OpenAI testing Fable) may differ from the rival's own harness. Where Anthropic published its own figure, we show both.
- Missing values are shown as n/a, never estimated.
- Sterling at approximately $1 = £0.79, excluding VAT.
The Bottom Line
Claude Fable 5.2 remains the most capable model you can buy: it tops two of three independent indexes, has the highest coding ceiling, and is stronger at open-ended research and factual reliability. GPT-6.2 Astra is the most efficient frontier model: at the same list price it completes Artificial Analysis's suite for about a seventh of the cost, and it leads on computational science, legal work, 3D design and robotics.
The safety contrast is the story to watch. Anthropic chose to cap Fable's ability to accelerate AI research; OpenAI shipped Astra across that line with new monitoring. Whichever you pick, measure cost per completed task on your own workload, not price per token, and check every agent's "done" for yourself.
Last updated: 18/11/2026. Figures as published on 17-18/11/2026; independent leaderboards update frequently.
Sources
- OpenAI: Introducing GPT-6.2 Astra and Sol
- OpenAI: GPT-6.2 Astra and GPT-6.2 Sol System Card
- The Decoder: GPT-6.2 launch with Artificial Analysis, Vals AI and Epoch AI results
- Trending Topics: GPT-6.2 Astra takes second place behind Anthropic's Fable 5.2
- OfficeChai: GPT-6.2 Sol and Astra benchmarks
- Artificial Analysis: Intelligence Index
- AI Tools Review: Claude Fable 5.2 and Mythos 5.2 review (Anthropic launch figures)
- Wes Roth: OpenAI JUST made EVERYONE MAD...
- Matthew Berman: Fable 5.2 just solved software






