Anthropic logo

Claude Fable 5.2

Multimodal
Anthropic

Claude Fable 5.2 is Anthropic's generally available, production-safeguarded deployment of the same underlying weights as the verification-only Claude Mythos 5.2. It targets demanding reasoning and long-horizon agentic work with a 1M-token context window, 128K max output, always-on adaptive thinking and text plus image input. First-party pricing is $8/$40 per million input/output tokens (down from $10/$50 on Fable 5.1), with cache reads at $0.20 and 50% off Batch; Anthropic reports outputs up to 30% faster than Fable 5.1 and roughly 55% fewer cyber and biology safeguard blocks.

Key Specifications

Parameters
-
Context
1.0M
Release Date
November 9, 2026
Average Score
73.1%

Timeline

Key dates in the model's history
Announcement
November 9, 2026
Last Update
November 10, 2026
Today
November 20, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
June 1, 2026
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$8.00
Output (per 1M tokens)
$40.00
Max Input Tokens
1.0M
Max Output Tokens
128.0K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
AA-Briefcase v1.1
Announcement + System Card §8.14.4. AA-Briefcase v1.1 Elo at max effort (max_score 3000). Rubric pass rate 66.1% not stored. • Self-reported
59.5%
ARC-AGI-3
System Card §8.15. Verified by the ARC Prize Foundation; Standard harness at high effort (max effort 65.66%). Provider Adapter harness scored 97.72% at high effort. • Self-reported
79.3%
AutomationBench
Fable 5.2 with production safeguards enabled. Announcement + System Card §8.14.5. Zapier private held-out set (AutomationBench 1.0.6, 657 tasks) at max effort with the API's default fallbacks. • Self-reported
51.4%
Chartography
System Card §8.13.1. With tools (container and image cropping tool) at max effort; 5-run average. • Self-reported
91.6%
Chartography (no tools)
Announcement + System Card §8.13.1. Without tools at max effort; 5-run average; Gemini 3.5 Flash grader as on the Surge AI leaderboard. • Self-reported
73.4%
CursorBench 4.0
System Card §8.8. CursorBench 4.0 at max effort, measured independently by Cursor in its production harness (xhigh 56.3%, high 56.9%). Cursor attributes part of the score to prompt and harness improvements, so it may not be directly comparable with earlier Claude 5.5 system cards. • Self-reported
60.0%
DeepSWE 1.1
System Card §8.3. DeepSWE v1.1 (113 tasks); 5-trial average. • Self-reported
74.3%
FrontierCode 2.0
Fable 5.2 with production safeguards enabled. Announcement + System Card Table 8.1.A / §8.4. Best across effort levels (xhigh); mean@5 run by Cognition. Max effort 45.7%, high 45.3%. Not comparable with FrontierCode 1.1. • Self-reported
49.1%
FrontierSWE V2
Fable 5.2 with production safeguards enabled. Announcement + System Card §8.7. Proximal harness; max effort; mean across five trials per task. • Self-reported
69.9%
GDPval-AA 2.1
System Card Table 8.1.A / §8.14.3. GDPval-AA v2.1 Elo at max effort (max_score 3000); xhigh scored 1751. • Self-reported
61.0%
Global-MMLU
System Card §8.17.1. Average accuracy across 42 languages; max effort; single trial; no tools; blocked or unanswered questions (0.5%) excluded. • Self-reported
96.0%
HealthBench
System Card §8.16.1. Length-adjusted score at max effort, run independently at Anthropic's request with GPT-4.1 grading and fallbacks on. Not comparable with earlier Claude system cards graded by Claude Opus 4.8. • Self-reported
59.6%
HealthBench Professional
System Card §8.16.2. Length-adjusted score at max effort; GPT-5.4 grader at low effort; fallbacks on. Not comparable with earlier Claude system cards. • Self-reported
67.7%
HealthBench Professional (raw)
System Card §8.16.2. Raw score before length adjustment at max effort. • Self-reported
71.1%
HealthBench (raw)
System Card §8.16.1. Raw score before length adjustment at max effort; same third-party run and grader as the length-adjusted score. • Self-reported
58.8%
Humanity's Last Exam
System Card Table 8.1.A / §8.11.1. Full 2,500-question HLE, reasoning only (no tools); 980k-token task budget, no compaction. • Self-reported
73.2%
Humanity's Last Exam (with tools)
Fable 5.2 with production safeguards enabled. Announcement + System Card Table 8.1.A / §8.11.1. Web search, web fetch, programmatic tool calling and code execution; HLE sources blocklisted; 980k-token task budget. • Self-reported
74.2%
MILU
System Card §8.17.2. Average accuracy across 11 languages; max effort; averaged over five trials; no tools. • Self-reported
94.1%
OfficeQA
System Card §8.14.1. OfficeQA full set; max effort; mean of five runs; agentic harness with extracted-text documents. • Self-reported
83.4%
OfficeQA Pro
System Card §8.14.1. OfficeQA Pro (133-question subset); max effort; mean of five runs. • Self-reported
72.9%
OSWorld 2.1 (partial)
Fable 5.2 with production safeguards enabled. Announcement + System Card §8.13.3. Official offline set (82 of 108 tasks); partial score, Pass@1 averaged over five attempts at max effort. • Self-reported
89.3%
OSWorld 2.1 (strict)
System Card §8.13.3. Official offline set (82 of 108 tasks); strict pass rate, Pass@1 averaged over five attempts at max effort. • Self-reported
54.6%
PhysicianBench
System Card §8.16.3. 100 EHR tasks, five attempts each, max effort; graded with Claude Opus 5. • Self-reported
64.4%
Program Bench
System Card §8.10.1. 166 golden tasks; mini-swe-agent harness; no 6-hour limit. 91.6% when the harness also shows remaining context tokens. • Self-reported
85.6%
SWE-bench Multilingual
System Card Table 8.1.A / §8.2. Adaptive thinking at max effort; 5-trial average. 300 problems across nine programming languages. • Self-reported
96.6%
SWE-Bench Multimodal
System Card Table 8.1.A / §8.2. Adaptive thinking at max effort; 5-trial average. • Self-reported
74.1%
SWE-Bench Pro
System Card Table 8.1.A / §8.2. Adaptive thinking at max effort; 5-trial average. • Self-reported
93.0%
Terminal-Bench 4.0
Fable 5.2 with production safeguards enabled. System Card §8.5. Claude Code --bare at max effort, no internet access (cached resources and offline search); 10 trials per task, SE ±1.8 pts. Safeguard fallbacks on 2.5% of requests. Mythos 5.2 without safeguards scored 70.5%. • Self-reported
67.6%
Terminal-Bench-Science 0.1
Fable 5.2 with production safeguards enabled. Announcement + System Card §8.6. Claude Code --bare at max effort, no internet access; 10 trials per task, SE ±4.3 pts. Mythos 5.2 without safeguards scored 75.3%. • Self-reported
74.6%

License & Metadata

License
proprietary
Announcement Date
November 9, 2026
Last Updated
November 10, 2026

Compare Claude Fable 5.2

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.