Claude Fable 5.2 Benchmarks
A reference guide to Anthropic's reported scores, model comparisons, cost claims, and testing limits.
On the performance evaluation table published by Anthropic on November 9, 2026, Claude Fable 5.2 holds the highest score of the listed models on 10 of the 12 non-Mythos benchmark rows. The model posts scores including 49.1% on FrontierCode 2.0 at extra-high effort, 69.9% on FrontierSWE v2, and 74.2% on Humanity's Last Exam with tool access. It trails Gemini 4 Carbon on AutomationBench by 2.5 percentage points and sits 15 points below Claude Opus 5.5 on the AA-Briefcase Elo rating.
Every published score comes directly from Anthropic's release evaluations rather than independent audits or third-party laboratory replications. The underlying architecture is identical between Claude Fable 5.2 and Claude Mythos 5.2. The benchmark table separates standard model runs from specialized cybersecurity and biology evaluations that include restricted model configurations.
Claude Fable 5.2 Benchmark Scores From Anthropic
Anthropic reported official performance benchmarks for Claude Fable 5.2 on November 9, 2026, comparing the release against Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra from OpenAI, and Gemini 4 Carbon. Per Anthropic's published footnotes, scores for Claude Fable 5.2, Claude Fable 5.1 and Claude Opus 5.5 are evaluated with standard safeguards turned on unless explicitly noted. In addition, all reported scores reflect each model's highest available effort level unless a different effort configuration is specified in the benchmark label.
See Claude Fable 5.2 explained for the broader technical context. The table below reproduces the complete non-Mythos results in the exact column sequence confirmed from the served release document.
| Benchmark | Claude Fable 5.2 | Claude Fable 5.1 | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% (xhigh) | 32.4% (xhigh) | 45.5% | 42.8% | 41.1% (medium) |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
| WANDR (relaxed score) | 82.1% | 68.7% | 72.3% | n.a. | n.a. |
| OSWorld 2.1 (offline set, partial) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim (run internally) | 64.9% | 27.6% | 39.4% | 29.8% | n.a. |
| AA-Briefcase v1.1 (rubric) | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1 (Elo) | 1785 | 1669 | 1800 | 1569 | 1611 |
| Remote Labor Index (RLI) | 41.3% (high) | 17.9% | 21.3% | 20.8% | n.a. |
| Humanity's Last Exam (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | n.a. |
| Chartography (no tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Anthropic notes that for Terminal-Bench-Science 0.1, the specialized sibling model Claude Mythos 5.2 achieved a slightly higher score of 75.3%, compared to 74.6% for Claude Fable 5.2. Across all 12 non-Mythos rows, Claude Fable 5.2 scores higher than Claude Fable 5.1 in every single head-to-head evaluation.
Where Claude Fable 5.2 Leads and Where It Trails
The 12 rows are counted by treating the AA-Briefcase v1.1 rubric percentage and the AA-Briefcase v1.1 Elo score as two separate benchmark entries, while using Claude Fable 5.2's 74.6% result for the Terminal-Bench-Science row. Across these 12 rows, Claude Fable 5.2 achieves the highest score among all listed models on exactly 10 tests.
Claude Fable 5.2 leads its predecessor, Claude Fable 5.1, across every tracked category. Gains include RoboDojo Sim (64.9% against 27.6%) and the Remote Labor Index (41.3% against 17.9%). On FrontierCode 2.0, Claude Fable 5.2 posts 49.1% compared to 32.4% for Claude Fable 5.1 at the extra-high effort tier.
When compared directly to Claude Opus 5.5, Claude Fable 5.2 leads on 11 of the 12 non-Mythos rows. Detailed baseline metrics for Claude Opus 5.5 appear in the Claude Opus 5.5 benchmarks reference. The single row where Claude Opus 5.5 outperforms Claude Fable 5.2 is the AA-Briefcase v1.1 Elo rating, where Claude Opus 5.5 achieves an Elo of 1800 against Claude Fable 5.2's 1785, establishing a 15-point margin.
The second row where Claude Fable 5.2 trails a competitor is AutomationBench. On AutomationBench, Gemini 4 Carbon secures the leading score of 53.9%, placing Claude Fable 5.2 in second place at 51.4%, a gap of 2.5 percentage points. Claude Fable 5.2 remains ahead of Claude Opus 5.5 (42.5%), GPT-6 Astra (41.4%), and Claude Fable 5.1 (31.4%) on the same benchmark.
Against GPT-6 Astra, Claude Fable 5.2 leads on every shared evaluation. On FrontierSWE v2, Claude Fable 5.2 reaches 69.9% compared to 65.5% for GPT-6 Astra. On Humanity's Last Exam with tool assistance, Claude Fable 5.2 records 74.2% while GPT-6 Astra records 57.2%, reflecting a 17.0 percentage point difference under tool-augmented reasoning conditions.
- Highest score on 10 of 12 non-Mythos evaluation rows
- Trails Gemini 4 Carbon by 2.5 percentage points on AutomationBench (51.4% vs 53.9%)
- Trails Claude Opus 5.5 by 15 points on AA-Briefcase v1.1 Elo (1785 vs 1800)
- Outperforms Claude Fable 5.1 on 12 of 12 rows
- Outperforms GPT-6 Astra across all 11 shared benchmark rows
Cost-Adjusted Results
Raw accuracy percentages do not reflect production inference expenses. Anthropic published comparative cost claims showing how adjusting the inference effort parameter on Claude Fable 5.2 alters the price-to-performance ratio against earlier models. Earlier baseline pricing and efficiency numbers can be reviewed in the Claude Fable 5.1 benchmarks analysis.
On FrontierCode 2.0, Anthropic reports that Claude Fable 5.2 operating at High effort scores 45.3%, placing it approximately level with Claude Opus 5.5 operating at Max effort (45.5%). Anthropic states that Claude Fable 5.2 completes this work for roughly half the cost per task compared to Claude Opus 5.5 at maximum effort, while outperforming GPT-6 Astra's highest reported score (42.8%) at approximately equivalent task expense. When raised to extra-high effort, Claude Fable 5.2 reaches 49.1%, which Anthropic characterizes as the highest score of any released model it has evaluated.
Similar cost efficiencies are reported on non-coding reasoning tasks. On WANDR, Claude Fable 5.2 set to Medium effort matches the top score achieved by Claude Opus 5.5 while consuming roughly 40% of the cost per task. On Humanity's Last Exam (HLE), Claude Fable 5.2 configured at Low effort matches Claude Opus 5.5's top score for under one-tenth of the task expense.
On Chartography, Claude Fable 5.2 operating at Medium effort achieves 66.4% accuracy, outpacing Claude Opus 5.5 at Max effort (64.4%) for under one-tenth the cost. Anthropic reports that Claude Fable 5.2 at Max effort achieves 83.5 on BenchCAD 2.1, compared to 65.3 for Claude Opus 5.5 and 44.4 for Claude Fable 5.1.
Per Anthropic's standard API pricing table, Claude Fable 5.2 costs $8 per million input tokens (MTok) and $40 per million output tokens. Prompt cache reads cost $0.20 per MTok. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Because token counts per task differ between models as effort behavior differs, actual per-task costs depend on measured tokens.
What Is Mythos-Only
Anthropic's published evaluation grid contains three specialized security and life sciences rows that are not like-for-like with the standard Claude Fable 5.2 results. These rows are explicitly designated for Claude Mythos 5.2 and Claude Mythos 5.1 in the first two columns of the announcement data, reflecting restricted-access deployments rather than public API endpoints.
In those three rows, the comparative baseline for Claude Opus 5.5 is reported under a 'without safeguards' testing configuration, which makes the results methodologically distinct from the standard enterprise rows. On CyScenarioBench, Anthropic notes that the cross-laboratory comparison reflects each laboratory running on its own unique task suites, meaning the task distributions differ between model providers.
The scores recorded on the three specialized rows are:
| Specialized Benchmark | Claude Mythos 5.2 | Claude Mythos 5.1 | Claude Opus 5.5 (no safeguards) | GPT-6 Cyber |
|---|---|---|---|---|
| CyScenarioBench (unique lab tasks) | 76.3% | 61.7% | 67.6% | 61.0% |
| CVE Rediscovery | 89.6% | 78.3% | 81.3% | n.a. |
| De novo binder design (24 hr) | 85.1% | 79.1% | 82.6% | n.a. |
Claude Fable 5.2 and Claude Mythos 5.2 are the same underlying model. However, access to Claude Mythos 5.2 requires verification through specialized partner frameworks, including the Cyber Verification Program and the Life Sciences Verification Program. Anthropic limits Claude Mythos 5.2 to verified organizations.
What These Scores Do Not Tell You
Effort levels and safeguard settings differ across rows within the same published table. For example, Claude Fable 5.2's FrontierCode 2.0 score reflects extra-high effort, while Gemini 4 Carbon is reported at medium effort. On the Remote Labor Index, Claude Fable 5.2 was tested at high effort. Comparing rows directly without checking the associated effort annotations obscures differences in test-time token expenditure.
Read the table as Anthropic's evidence that Claude Fable 5.2 leads on long agentic and reasoning work: it has the highest score on 10 of the 12 non-Mythos rows, with the AutomationBench and AA-Briefcase Elo exceptions, so choose by the benchmark nearest your job and replicate on your own tasks before committing.
This guidance applies to plan buyers and engineering leads comparing frontier foundation models for production deployment. What would change it: independent evaluations that contradict Anthropic's published numbers, or official corrections issued to the launch evaluation table.
Frequently Asked Questions
- Yes, on Anthropic's published benchmarks Claude Fable 5.2 scores higher than Claude Opus 5.5 across all reported coding tasks. On FrontierCode 2.0 at extra-high effort, Claude Fable 5.2 achieves 49.1% compared to 45.5% for Claude Opus 5.5 at maximum effort. On FrontierSWE v2, Claude Fable 5.2 reaches 69.9% while Claude Opus 5.5 reaches 62.3%.
- No, all published scores for Claude Fable 5.2 are vendor-reported figures released directly by Anthropic on November 9, 2026. The launch materials do not include external third-party laboratory replications.
- They share the same underlying model architecture, but their reported benchmark scores differ slightly on specific tasks. On Terminal-Bench-Science 0.1, Claude Mythos 5.2 scored 75.3% while Claude Fable 5.2 scored 74.6%. In addition, Anthropic reported dedicated cybersecurity and biological design evaluations exclusively for Claude Mythos 5.2.
Request an AI Workflow Audit
Evaluate how Claude Fable 5.2 benchmark performance translates to your team's real production tasks. Contact Layer3Labs to audit model selection, compute costs, and integration requirements.
Book a Consultation