Claude Fable 5.2 benchmarks went live on November 9, 2026, in a 14-row table on Anthropic's launch page. Fable 5.2 leads 10 of the 11 rows where it has its own score. Gemini 4 Carbon beats it on the other one.
That table is useful, but most rows test things a buyer never runs. And every number in it is Anthropic's own.
This page ranks the 8 Fable 5.2 benchmarks that tell you the most before you pay for it. For each one you get the official score, what the test measures, the effort level behind the number and the catch. Then you get the independent leaderboards: Vals AI, Arena and Artificial Analysis.
TL;DR: Read FrontierSWE v2 first. Fable 5.2 scores 69.9% there (Anthropic), ahead of GPT-6 Astra (65.5%) and Opus 5.5 (62.3%). If you automate business workflows, check AutomationBench, the one row where Gemini 4 Carbon wins (53.9% vs 51.4%).
Claude Fable 5.2 benchmarks: a quick overview
If you are buying a model for coding work:
- FrontierSWE v2: best for teams picking a coding agent. Fable 5.2 has the top score of the five models Anthropic tested.
- FrontierCode 2.0: best for judging cost per coding task. Anthropic gives a cost comparison for this row.
If you are buying for research, analysis or office work:
- Humanity's Last Exam (with tools): best for research and analysis teams. Fable 5.2 leads GPT-6 Astra by 17 points.
- Remote Labor Index: best for estimating real project work. Fable 5.2 nearly doubles Opus 5.5 here.
- Chartography: best for teams that read charts and reports all day.
If you are building agents that act on their own:
- OSWorld 2.1: best for computer-use and browser agents.
- AutomationBench: best for business workflow automation, and the row Fable 5.2 loses.
- Terminal-Bench-Science 0.1: best for science and data teams that run code in a terminal.
All scores below are from Anthropic's launch table, read on Nov 9, 2026. Anthropic ran each model at its highest available effort unless a cell says otherwise. "Not shown" means the page gives no score for that model.
| Benchmark | What it tests (Anthropic's label) | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierSWE v2 | Agentic coding | 69.9% | 62.3% | 65.5% | 61.2% |
| FrontierCode 2.0 | Agentic coding | 49.1% (xhigh) | 45.5% | 42.8% | 41.1% (medium) |
| Humanity's Last Exam, with tools | Multidisciplinary reasoning | 74.2% | 67.7% | 57.2% | Not shown |
| OSWorld 2.1, offline set, partial | Computer use | 89.3% | 87.2% | 79.5% | 81.5% |
| Remote Labor Index | Knowledge work | 41.3% (high) | 21.3% | 20.8% | Not shown |
| AutomationBench | Business workflows | 51.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | Agentic scientific research | 74.6% | 58.7% | 68.1% | 64.8% |
| Chartography, no tools | Visual chart recognition | 73.4% | 64.4% | 71.0% | 72.5% |
1. FrontierSWE v2, best for teams picking a coding agent
FrontierSWE v2 is the agentic coding benchmark Anthropic reports for Fable 5.2, and it is the closest thing to a "Fable 5.2 SWE-bench score" on the launch page. Anthropic lists it under "Agentic coding". The page does not report SWE-bench Verified or SWE-bench Pro.
Fable 5.2 scores 69.9% on Anthropic's table. That is 4.4 points above GPT-6 Astra, 7.6 above Opus 5.5 and 13.6 above Fable 5.1 (56.3%). It is the cleanest coding row in the table: no special effort note, and all five models have a score.
Key features
- Score: 69.9% for Fable 5.2, the highest of the five models shown
- Category: agentic coding, per Anthropic's table
- Effort: highest available for each model, no exceptions noted
- Fable 5.1 scored 56.3%, so the jump in one release is 13.6 points
Best for
- Engineering leads choosing a default model for a coding agent
- Teams on Opus 5.5 asking whether Fable 5.2 is worth the switch
Pricing Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, twice Opus 5.5 at $4 and $20 (as of Nov 9, 2026, Anthropic models overview). Anthropic gives no cost per task for this row.
Pros
- Every competitor has a score, so the comparison is complete
- The lead over GPT-6 Astra holds without an effort footnote
Cons
- Anthropic ran it, not an independent lab
- A 7.6-point lead over Opus 5.5 comes at 2x the token price
2. FrontierCode 2.0, best for judging cost per coding task
FrontierCode 2.0 is the second agentic coding benchmark in Anthropic's table, and the one where Anthropic says what the score costs. That makes it the most useful row for a budget decision.
At xhigh effort, Fable 5.2 scores 49.1% (Anthropic), against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. The more telling number sits below the table. Anthropic says Fable 5.2 at high effort scores 45.3%, about level with Opus 5.5's best, "for roughly half the cost per task."
Key features
- Score: 49.1% at xhigh effort, 45.3% at high effort
- Fable 5.1 scored 32.4% at xhigh, a 16.7-point jump
- Gemini 4 Carbon's 41.1% is at medium effort, so that cell is not like for like
- Anthropic also says Fable 5.2 beats GPT-6 Astra's best "at about the same cost"
Best for
- Teams that pay per task and want the cheapest route to a given score
- Anyone running Opus 5.5 at max effort today
Pricing Per token, Fable 5.2 costs 2x Opus 5.5 ($8/$40 vs $4/$20 per million tokens, as of Nov 9, 2026). Per task on this benchmark, Anthropic says Fable 5.2 at high effort costs about half of Opus 5.5 at its best (launch page).
Pros
- The only coding row with a cost-per-task claim attached
- Shows that effort level changes the answer, not only the model
Cons
- The 49.1% headline needs xhigh effort, which costs more per task
- Cost per task is Anthropic's figure; your own tasks may differ
3. Humanity's Last Exam (with tools), best for research and analysis teams
Humanity's Last Exam is a multi-modal academic test from the Center for AI Safety and Scale AI, with 2,500 questions across more than a hundred subjects. Anthropic reports the "with tools" version, so the model can search and run code while it answers.
Fable 5.2 scores 74.2% on Anthropic's table. Opus 5.5 scores 67.7%, Fable 5.1 65.6% and GPT-6 Astra 57.2%. Anthropic shows no Gemini 4 Carbon score for this row.
Key features
- Score: 74.2% with tools
- Lead over GPT-6 Astra: 17 points
- Anthropic says Fable 5.2 at low effort matches Opus 5.5's top score
- Scored with tools, so it also tests search and tool use
Best for
- Analysts and researchers who ask hard, multi-step questions
- Teams comparing Claude with GPT-6 Astra for research assistants
Pricing Anthropic says Fable 5.2 at low effort matches Opus 5.5's top score here "for less than one-tenth of the cost per task" (launch page). List price is $8/$40 per million tokens (as of Nov 9, 2026).
Pros
- A 17-point margin over GPT-6 Astra
- The low-effort result means you may not need the expensive setting
Cons
- No Gemini 4 Carbon score, so the field is incomplete
- Tool-assisted scores depend on the tools Anthropic gave the model
4. OSWorld 2.1, best for computer-use and browser agents
OSWorld is a benchmark where an AI agent works a real computer desktop to finish tasks. Anthropic reports OSWorld 2.1 on the "offline set, partial", labeled "Computer use".
Fable 5.2 scores 89.3% (Anthropic), ahead of Opus 5.5 (87.2%), Gemini 4 Carbon (81.5%) and GPT-6 Astra (79.5%). The gap to Opus 5.5 is 2.1 points, the smallest in the coding and agent rows.
Key features
- Score: 89.3% on the offline, partial set
- Opus 5.5 is 2.1 points behind
- GPT-6 Astra trails by nearly 10 points
- Fable 5.1 scored 83.5%
Best for
- Teams building agents that click through apps and websites
- Ops teams automating back-office screens with no API
Pricing No cost-per-task figure on the launch page. At list price Fable 5.2 is 2x Opus 5.5 per token ($8/$40 vs $4/$20, as of Nov 9, 2026, models overview).
Pros
- Clear lead over GPT-6 Astra and Gemini 4 Carbon
- High absolute score on a task type where agents often fail
Cons
- "Offline set, partial" is a subset, not the full benchmark
- At 2.1 points over Opus 5.5, the cheaper model may be enough
5. Remote Labor Index, best for estimating real project work
Remote Labor Index (RLI) is the knowledge work benchmark in Anthropic's table. It measures how many real freelance projects an AI agent can complete to a standard a client would accept.
Fable 5.2 scores 41.3% at high effort on Anthropic's table. Opus 5.5 scores 21.3%, GPT-6 Astra 20.8% and Fable 5.1 17.9%. No Gemini 4 Carbon score appears. Fable 5.2's score is nearly double every other model in the row.
Key features
- Score: 41.3% at high effort
- Nearly 2x Opus 5.5 (21.3%) and GPT-6 Astra (20.8%)
- Fable 5.1 to 5.2: up 23.4 points
- Effort is high, not max, for Fable 5.2
Best for
- Agencies and freelancers pricing work an agent could take on
- Leaders estimating how much project work to hand to AI
Pricing No cost-per-task figure for this row. Fable 5.2 lists at $8/$40 per million tokens (as of Nov 9, 2026, models overview).
Pros
- The largest lead over the next-best model in the table, in relative terms
- Measured at high effort, so not the most expensive setting
Cons
- 41.3% still means most projects are not completed
- No Gemini 4 Carbon score to compare
6. AutomationBench, best for business workflow automation
AutomationBench is the business workflows row in Anthropic's table, and the one row where Fable 5.2 does not win. Gemini 4 Carbon scores 53.9% against Fable 5.2's 51.4% (Anthropic).
On the same table, Fable 5.2 still beats Opus 5.5 (42.5%), GPT-6 Astra (41.4%) and Fable 5.1 (31.4%). If your main use is multi-step business processes, this is the row to test against your own workflows before you commit.
Key features
- Score: 51.4% for Fable 5.2, 53.9% for Gemini 4 Carbon
- Fable 5.2 leads Opus 5.5 by 8.9 points
- Fable 5.1 to 5.2: up 20 points
- Anthropic published the loss in its own table
Best for
- RevOps and ops teams automating CRM, billing or support workflows
- Buyers choosing between Claude and Gemini for automation
Pricing No cost-per-task figure for this row. Fable 5.2 lists at $8/$40 per million tokens (as of Nov 9, 2026, models overview).
Pros
- Honest signal: the vendor shows where a rival wins
- Still a large gain over every other Claude model
Cons
- Gemini 4 Carbon is ahead by 2.5 points
- Roughly half the tasks still fail for every model shown
7. Terminal-Bench-Science 0.1, best for science and data teams
Terminal-Bench-Science 0.1 is the row Anthropic labels agentic scientific research. It is the only Fable 5.2 row where Anthropic also gives a Mythos 5.2 score.
Fable 5.2 scores 74.6% and Mythos 5.2 scores 75.3% (Anthropic). GPT-6 Astra scores 68.1%, Gemini 4 Carbon 64.8% and Opus 5.5 58.7%. The 15.9-point lead over Opus 5.5 is one of the largest gaps between the two Claude models.
Key features
- Score: 74.6% for Fable 5.2, 75.3% for Mythos 5.2
- Lead over Opus 5.5: 15.9 points
- Lead over GPT-6 Astra: 6.5 points
- Version 0.1, so the benchmark itself is early
Best for
- Research labs and data teams running analysis code
- Teams deciding whether Fable 5.2 is worth it over Opus 5.5 for science work
Pricing No cost-per-task figure for this row. List price is $8/$40 per million tokens (as of Nov 9, 2026). Mythos 5.2 is not sold publicly; access goes through Anthropic's verification programs (launch page).
Pros
- Wide margin over Opus 5.5 on a hard task type
- Fable 5.2 lands within 0.7 points of the gated Mythos 5.2
Cons
- A 0.1 version benchmark may change a lot
- Vals AI's own Terminal-Bench-Science run puts Fable 5.2 at 55.71%, a different setup and number
8. Chartography, best for teams that read charts and reports
Chartography is the row Anthropic labels visual chart recognition, run with no tools.
Fable 5.2 scores 73.4% (Anthropic), a narrow lead over Gemini 4 Carbon (72.5%) and GPT-6 Astra (71.0%). Opus 5.5 scores 64.4% and Fable 5.1 44.8%. The 28.6-point jump from Fable 5.1 is second only to RoboDojo Sim (up 37.3 points).
Key features
- Score: 73.4% with no tools
- Lead over Gemini 4 Carbon: 0.9 points
- Fable 5.1 to 5.2: up 28.6 points
- At medium effort, Fable 5.2 scores 66.4%, ahead of Opus 5.5 at max
Best for
- Finance and analytics teams pulling numbers from PDFs and slides
- Anyone feeding dashboards or screenshots to a model
Pricing Anthropic says Fable 5.2 at medium effort beats Opus 5.5 at max effort here "for less than one-tenth of the cost per task" (launch page). List price is $8/$40 per million tokens (as of Nov 9, 2026).
Pros
- Strong cost case at medium effort
- Second-largest gain over Fable 5.1 in the table
Cons
- Gemini 4 Carbon and GPT-6 Astra are within 2.4 points
- Chart reading is one narrow skill, not general vision
What about the other rows on Anthropic's page?
Anthropic's table has 14 rows. The 6 rows left out above are either narrow or not Fable 5.2 scores:
| Row | Fable 5.2 | Why it is not ranked above |
|---|---|---|
| WANDR (agentic search, relaxed score) | 82.1% | Only Claude models shown. Anthropic says medium effort matches Opus 5.5's top score for about 40% of the cost per task |
| AA-Briefcase v1.1 (knowledge work, rubric) | 66.1% | Rubric-scored; Gemini 4 Carbon 61.8%, Opus 5.5 61.2% |
| RoboDojo Sim (robotics) | 64.9% | Run internally by Anthropic; few buyers need robotics |
| CyScenarioBench | 76.3% (Mythos 5.2) | A Mythos 5.2 score, not Fable 5.2 |
| CVE Rediscovery | 89.6% (Mythos 5.2) | A Mythos 5.2 score, not Fable 5.2 |
| De novo binder design (24 hr) | 85.1% (Mythos 5.2) | A Mythos 5.2 score, not Fable 5.2 |
The page also shows an Elo row (Opus 5.5 at 1800, Fable 5.2 at 1785) without naming the benchmark. We left it out until Anthropic labels it.
Fable 5.2 SWE-bench: which coding benchmarks were reported?
Anthropic did not report a SWE-bench score for Fable 5.2. Its launch table coding rows are FrontierSWE v2 (69.9%) and FrontierCode 2.0 (49.1% at xhigh). If you search for "Fable 5.2 SWE-bench", those are the official numbers.
The only SWE-bench variant with a public Fable 5.2 score on launch day is on Vals AI: SWE-bench Multimodal, 62.29%, rank 2 of 22 (checked Nov 9, 2026). Pages that quote a SWE-bench Verified number for Fable 5.2 are not quoting Anthropic.
Fable 5.2 leaderboard: where it ranks outside Anthropic's table
Three independent leaderboards matter for buyers. Each uses its own tests, so a rank on one does not carry to another. We checked Vals AI and Arena on Nov 9, 2026, at 19:27 UTC, and Artificial Analysis again at 22:20 UTC the same day.
| Leaderboard | Fable 5.2 status | Key number |
|---|---|---|
| Vals AI | Listed | #1 of 36 on the Vals Index, 71.31% |
| Arena (formerly LMArena) | Not listed yet | Fable 5.1 (Max) is #4 in Agent "Best Overall" at 14.34% |
| Artificial Analysis | Listed | Fable 5.2 (max) is #1 of 74 on the Intelligence Index, 65 |
Vals AI. On the Vals Index, Fable 5.2 ranks #1 of 36 at 71.31%, at $17.94 per test and 43 minutes 42 seconds of latency. It is also #1 of 127 on Vibe Code Bench v1.1 (93.77%) and #1 of 63 on Terminal-Bench 4.0 (65.15%). It has weak spots on the same page: 5.00% on HLAB-Vals-1.1 (30 of 37) and 32.09% on Tax Agent Bench (27 of 84). The Vals Index is a different index from Artificial Analysis's Intelligence Index, so do not compare the two numbers.
Arena. Arena ranks models by human votes. Fable 5.2 had no entry at our check. Gemini 4 Carbon (High) was #1 in the Text arena.
Fable 5.2 on Artificial Analysis: #1 of 74
Claude Fable 5.2 (max) scores 65 on the Artificial Analysis Intelligence Index v5.0.1 and ranks #1 of 74 models (Artificial Analysis model page, checked Nov 9, 2026, at 22:20 UTC). That puts it 5 points ahead of Gemini 4 Carbon (high) and 6 ahead of Claude Opus 5.5 (max).
The catch is cost and speed. Fable 5.2 (max) costs $26.54 per Intelligence Index task and takes 44.1 minutes per task. It writes 103 tokens per second, #72 of 73 on speed (Artificial Analysis). Opus 5.5 (max) costs $10.04 per task at 59 points.
The top of the Intelligence Index leaderboard at that check:
| Model (effort) | Intelligence Index | Cost per task | Time per task |
|---|---|---|---|
| Claude Fable 5.2 (max) | 65 | $26.54 | 44.1 min |
| Gemini 4 Carbon (high) | 60 | $7.76 | 27.2 min |
| Claude Opus 5.5 (max) | 59 | $10.04 | 25.2 min |
| Gemini 4 Carbon (medium) | 58 | $6.59 | 23.7 min |
| Claude Opus 5.5 (xhigh) | 57 | $6.05 | 16.3 min |
| GPT-6 Astra (max) | 57 | $6.24 | 22.3 min |
Lower effort costs points. On the same leaderboard Fable 5.2 (xhigh) and (high) both score 62, (medium) 56 and (low) 54. Mythos 5.2 was not listed at our check.
How to choose which Fable 5.2 benchmark to trust
A benchmark helps only when it looks like your work. Start from the job, then pick the row.
1) Do you write code or automate business processes?
If you write code: read FrontierSWE v2 and FrontierCode 2.0. Fable 5.2 leads both. If you automate CRM, billing or support flows: read AutomationBench, where Gemini 4 Carbon is ahead. Run both models on 10 of your own workflows before you pick.
2) Do you need the top score or the lowest cost per task?
If you need the top score: run Fable 5.2 at xhigh. If cost per task matters more, start at high or medium effort. Anthropic's own notes on FrontierCode, Humanity's Last Exam and Chartography all show a lower effort level matching Opus 5.5's best for half the cost or less.
3) Do you trust vendor numbers or independent ones?
Anthropic ran every number in its table. Vals AI and Artificial Analysis ran their own tests, and both put Fable 5.2 first. For a procurement document, quote the Artificial Analysis Intelligence Index (65, #1 of 74) with its cost per task ($26.54), so the price of that score is on the same page.
4) Is Opus 5.5 good enough?
Anthropic's own docs say to start with Opus 5.5 for most workloads and use Fable 5.2 "when your evals on Claude Opus 5.5 at higher effort still fall short" (models overview). OSWorld 2.1 is a row where the gap is 2.1 points. Remote Labor Index is a row where Fable 5.2 nearly doubles it.
Buyers read pages like this one before they pick a tool, and so do ChatGPT and Perplexity. If you sell a SaaS product, AY Listicle gets it into the "best of" lists those buyers read. Request a spot on AY Listicle. More guides are on the AY Listicle blog.
FAQ
What are the main Claude Fable 5.2 benchmarks?
The main Claude Fable 5.2 benchmarks are FrontierSWE v2 (69.9%), FrontierCode 2.0 (49.1% at xhigh), Humanity's Last Exam with tools (74.2%) and OSWorld 2.1 (89.3%). All four come from Anthropic's launch table of Nov 9, 2026. Fable 5.2 leads 10 of the 11 rows where it has its own score. What is Fable 5.2's SWE-bench score? Anthropic did not report a SWE-bench score for Fable 5.2 on its launch page. The official coding numbers are FrontierSWE v2 at 69.9% and FrontierCode 2.0 at 49.1%. Vals AI lists a SWE-bench Multimodal score of 62.29%, rank 2 of 22. What did Fable 5.2 score on Humanity's Last Exam? Fable 5.2 scored 74.2% on Humanity's Last Exam with tools, per Anthropic's launch page. Opus 5.5 scored 67.7% and GPT-6 Astra 57.2%. Anthropic says Fable 5.2 at low effort matches Opus 5.5's top score for less than a tenth of the cost per task. Is Fable 5.2 on the Artificial Analysis leaderboard? Yes. Claude Fable 5.2 (max) ranks #1 of 74 on the Artificial Analysis Intelligence Index v5.0.1, with a score of 65 (Artificial Analysis, checked Nov 9, 2026). It costs $26.54 per index task, against $10.04 for Opus 5.5 (max) at 59. Where does Fable 5.2 rank on leaderboards? Fable 5.2 ranks #1 of 36 on the Vals AI Vals Index at 71.31%, and Fable 5.2 (max) ranks #1 of 74 on the Artificial Analysis Intelligence Index at 65. It was not yet listed on Arena on launch day. Each leaderboard uses its own tests, so ranks do not transfer between them. Does Fable 5.2 beat Opus 5.5 on benchmarks? Fable 5.2 beats Opus 5.5 on every row in Anthropic's table where both have a score. The smallest gap is OSWorld 2.1 (89.3% vs 87.2%). The largest relative gap is Remote Labor Index (41.3% vs 21.3%). Where does Fable 5.2 lose? Fable 5.2 loses AutomationBench to Gemini 4 Carbon, 51.4% to 53.9%, in Anthropic's own table. On Vals AI it scores 5.00% on HLAB-Vals-1.1 and 32.09% on Tax Agent Bench. How much does Claude Fable 5.2 cost? Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, as of Nov 9, 2026 (models overview). That is 20% below Fable 5.1 and twice Opus 5.5 per token. Batch API requests are 50% off. Are Mythos 5.2 benchmark scores the same as Fable 5.2? Mythos 5.2 scores are not Fable 5.2 scores. Anthropic reports cybersecurity and protein design rows only for Mythos 5.2, which is available through its verification programs. On Terminal-Bench-Science 0.1, where both appear in Anthropic's table, Mythos 5.2 scores 75.3% and Fable 5.2 74.6%.
Sources: Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, Anthropic models overview, Vals AI, Claude Fable 5.2, Artificial Analysis leaderboard, Artificial Analysis, Claude Fable 5.2, Arena leaderboard, Humanity's Last Exam, OSWorld. All checked Nov 9, 2026.
