Claude Fable 5.2
Released
2026-11-09
Last refreshed
2026-11-09
Status
Researched today
ProprietaryCommercial use: restrictedMultimodalCodingRAGAgentsLong contextVisionJSON / Tool useAgentCodingHighlight
Anthropic releases · 16 in the last 12 months · this family litChangelog →
Specifications
- Family
- Claude Fable
- Released
- 2026-11-09
- Context
- 1M
- Max output
- 128K
- Architecture
- Decoder Only
- Knowledge cutoff
- 2026-06
- Specialization
- general
- Openness
- Proprietary
- License
- ProprietaryRestricted
- Weights
- Not released
- Code
- Unknown
- Training
- Pretrained
Created by
Pricing
Output / 1M
$40.00
Input / 1M
$8.00
Cheapest of 5 routes · Anthropic · cache read $0.200
Links
About
Claude Fable 5.2 is Anthropic's generally available Mythos-class model for demanding reasoning and long-horizon agentic work, released November 9, 2026. It is the same underlying model as Claude Mythos 5.2 with cybersecurity and biology safeguards; requests those safeguards catch can fall back to other Claude models. Claude API id claude-fable-5-2. 1M-token context, 128K max output, adaptive thinking always on (default effort high), text and image input to text output. Knowledge cutoff June 2026. Anthropic documents model IDs for the Claude API, Amazon Bedrock, Claude Models on Google Cloud (Google-managed and Anthropic-managed), Microsoft Foundry and Claude Platform on AWS.
Providers(5)
Compare all 5Compare API pricing across 5 tracked provider routes for input and output tokens, batch, and cached reads when available.
| Provider | Input / 1M | Output / 1M | Batch in / out | Cache | Route |
|---|---|---|---|---|---|
| Anthropic | $8.00 | $40.00 | $4.00 / $20.00 | read $0.200 / 5m $10.00 / 1h $16.00 | Serverless |
| OpenRouter | $8.00 | $40.00 | - | read $0.200 / 5m $10.00 / 1h $16.00 | Serverless |
| Vercel AI Gateway | $8.00 | $40.00 | - | read $0.200 / 5m $10.00 | Serverless |
| AWS Bedrock | - | - | - | - | ServerlessPartial |
| GCP Vertex AI | - | - | - | - | ServerlessPartial |
Available via routers & gateways(15)
LiteLLM
GatewayOpen-source Python SDK and proxy server that unifies 100+ LLM APIs behind a single OpenAI-compatible interface, with load balancing, cost tracking, and configurable failover.
Free OSSAnthropicGCP Vertex AI
OpenRouter
HybridUnified hybrid gateway to 500+ models from 80+ providers via a single OpenAI-compatible API, with an optional Auto Router that selects a model per prompt from market spend data.
Passthrough + feeAnthropicGCP Vertex AI
Portkey
GatewayProduction AI gateway routing to 1,600+ LLMs with failover, load balancing, caching, and guardrails; MIT-licensed open-source core is self-hostable. Now part of Palo Alto Networks as Prisma AIRS AI Gateway.
SubscriptionAnthropicGCP Vertex AI
AIRouter
RouterCommercial LLM router (branded "AI Router" on its site) that picks the best model per request via a Decision API or an OpenAI-compatible Chat Completions endpoint, with a private selection mode that routes on anonymized patterns; sold as BYOK subscription plans priced by routed tokens.
SubscriptionAnthropicGCP Vertex AI
Amazon Bedrock Intelligent Prompt Routing
RouterAWS Bedrock's native prompt router that predicts each model's performance per request and routes between two models of the same family (Anthropic Claude, Meta Llama or Amazon Nova).
Passthrough + feeAWS Bedrock
Helicone
GatewayObservability-first AI gateway with routing, caching, rate limiting, and request tracing; Apache-2.0 open-source platform with a managed hosted tier for logging and analytics.
SubscriptionAnthropicGCP Vertex AI
Capabilities
VisionMultimodalReasoningJSON / Tool useStructured OutputsPrompt CachingBatch API
Benchmark peer barsfor Coding
SWE-bench ProRank 1 of 62
Benchmark scores(24)
Scores are benchmark-specific and are direction-aware: the same numeric gap can mean very different outcomes across suites. Use the leaderboard context and this model's provider route to decide whether the winning margin is meaningful for your workload.
| Benchmark | Score | Version | Evaluation | Source |
|---|---|---|---|---|
| 60.0 | CursorBench 4.0Observed 2026-11-09 | Configuration: Fable 5.2 Max Harness: CursorBench 4.0 productized Cursor-agent workflow Evaluator: Cursor Confidence: confirmed Cost/task: $32.79 Tokens/task: 352,794 Steps/task: 218 Notes: Cursor vendor-reported CursorBench 4.0 configuration result; verified on https://cursor.com/cursorbench chart label / table row 'Fable 5.2 Max' (retrieved 2026-11-09): 60.0%, $32.79 avg cost per task, 352794 tokens, 218 steps. Not independently reproducible. Results are subject to variance, and small score differences may not be statistically meaningful. The Claude Fable 5.2 system card (§8.8) says Cursor measured and reported all scores independently, and also notes that Cursor attributes part of Fable 5.2's score to additional prompt and harness improvements, so the score may not be directly comparable with earlier system-card CursorBench figures. Timestamp is retrieval date (page carries no per-row date). | ||
| 56.3 | CursorBench 4.0Observed 2026-11-09 | Configuration: Fable 5.2 Extra High Harness: CursorBench 4.0 productized Cursor-agent workflow Evaluator: Cursor Confidence: confirmed Cost/task: $15.48 Tokens/task: 161,251 Steps/task: 108 Notes: Cursor vendor-reported CursorBench 4.0 configuration result; verified on https://cursor.com/cursorbench chart label / table row 'Fable 5.2 Extra High' (retrieved 2026-11-09): 56.3%, $15.48 avg cost per task, 161251 tokens, 108 steps. Not independently reproducible. Results are subject to variance, and small score differences may not be statistically meaningful. The Claude Fable 5.2 system card (§8.8) says Cursor measured and reported all scores independently, and also notes that Cursor attributes part of Fable 5.2's score to additional prompt and harness improvements, so the score may not be directly comparable with earlier system-card CursorBench figures. Timestamp is retrieval date (page carries no per-row date). | ||
| 56.9 | CursorBench 4.0Observed 2026-11-09 | Configuration: Fable 5.2 High Harness: CursorBench 4.0 productized Cursor-agent workflow Evaluator: Cursor Confidence: confirmed Cost/task: $9.70 Tokens/task: 99,259 Steps/task: 69 Notes: Cursor vendor-reported CursorBench 4.0 configuration result; verified on https://cursor.com/cursorbench chart label / table row 'Fable 5.2 High' (retrieved 2026-11-09): 56.9%, $9.7 avg cost per task, 99259 tokens, 69 steps. Not independently reproducible. Results are subject to variance, and small score differences may not be statistically meaningful. The Claude Fable 5.2 system card (§8.8) says Cursor measured and reported all scores independently, and also notes that Cursor attributes part of Fable 5.2's score to additional prompt and harness improvements, so the score may not be directly comparable with earlier system-card CursorBench figures. Timestamp is retrieval date (page carries no per-row date). | ||
| 54.8 | CursorBench 4.0Observed 2026-11-09 | Configuration: Fable 5.2 Medium Harness: CursorBench 4.0 productized Cursor-agent workflow Evaluator: Cursor Confidence: confirmed Cost/task: $4.66 Tokens/task: 44,103 Steps/task: 35 Notes: Cursor vendor-reported CursorBench 4.0 configuration result; verified on https://cursor.com/cursorbench chart label / table row 'Fable 5.2 Medium' (retrieved 2026-11-09): 54.8%, $4.66 avg cost per task, 44103 tokens, 35 steps. Not independently reproducible. Results are subject to variance, and small score differences may not be statistically meaningful. The Claude Fable 5.2 system card (§8.8) says Cursor measured and reported all scores independently, and also notes that Cursor attributes part of Fable 5.2's score to additional prompt and harness improvements, so the score may not be directly comparable with earlier system-card CursorBench figures. Timestamp is retrieval date (page carries no per-row date). | ||
| 53.2 | CursorBench 4.0Observed 2026-11-09 | Configuration: Fable 5.2 Low Harness: CursorBench 4.0 productized Cursor-agent workflow Evaluator: Cursor Confidence: confirmed Cost/task: $3.94 Tokens/task: 36,616 Steps/task: 30 Notes: Cursor vendor-reported CursorBench 4.0 configuration result; verified on https://cursor.com/cursorbench chart label / table row 'Fable 5.2 Low' (retrieved 2026-11-09): 53.2%, $3.94 avg cost per task, 36616 tokens, 30 steps. Not independently reproducible. Results are subject to variance, and small score differences may not be statistically meaningful. The Claude Fable 5.2 system card (§8.8) says Cursor measured and reported all scores independently, and also notes that Cursor attributes part of Fable 5.2's score to additional prompt and harness improvements, so the score may not be directly comparable with earlier system-card CursorBench figures. Timestamp is retrieval date (page carries no per-row date). | ||
| 93.8 | ARC-AGI-2 semi-privateObserved 2026-11-09 | Configuration: claude-fable-5-2 (XHigh) Harness: ARC Prize semi-private evaluation Evaluator: ARC Prize Foundation Confidence: confirmed Cost/task: $0.98 Notes: ARC Prize official leaderboard data (ARC-AGI-2 semi-private), verified in https://arcprize.org/media/data/leaderboard/v2.json (generatedAt 2026-11-09T15:06:07.767Z) for modelDisplayName 'Claude Fable 5.2 (XHigh)', modelId anthropic-claude-fable-5-2-xhigh, datasetId v2_Semi_Private, score 0.9375, costPerTask 0.9832189333333337. Stored as 93.75. Highest-scoring published effort selected; other official efforts in the same file: Max 0.925 (92.5), High 0.9041666666666667, Medium 0.8958333333333334, Low 0.8666666666666667. modelReleaseDate 2026-11-09T00:00:00.000Z. Retrieved 2026-11-09. This is an owner-board row and may move /best/reasoning. | ||
| 49.1 | FrontierCode 2.0Observed 2026-11-09 | Configuration: claude-fable-5-2 (XHigh) Harness: Claude Code (Cognition official evaluation) Evaluator: Cognition Confidence: confirmed Notes: Cognition official FrontierCode 2.0 board (https://cognition.com/frontiercode, retrieved 2026-11-09): JSON-LD ListItem position 1, name 'Claude Fable 5.2', description 'Score 49.1%'. The Anthropic system card §8.4 confirms Cognition ran the evaluation (official eval, no changes), 132-task rebuilt suite (FrontierCode 1.1 had 150) with updated scope grading — scores are not comparable to frontiercode-main (1.1). Mean@5 composite. Effort peak is xhigh 49.1%; other Cognition/card efforts: low 34.2, medium 37.6, high 45.3, max 45.7 (max includes 22/660 timed-out attempts still graded; if timed out as zero, 44.4%). Claude models run in Claude Code. Highest-scoring published effort selected. This is an owner-board row on a NEW slug; it must not be written onto frontiercode-main. A /best/agents key move from FrontierCode 1.1 Main to FrontierCode 2.0 would put Fable 5.2 first at 49.1 and is a separate FE PR, not this packet. | ||
| 93.0 | SWE-bench ProObserved 2026-11-09 | Configuration: claude-fable-5-2 (max effort, mean of 5 trials) Harness: Anthropic system-card standard configuration (scaffold not named); average of five trials Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.2 / Table 8.1.A: 93.0%. Adaptive thinking at max effort, default sampling, average of five trials. | ||
| 96.6 | SWE-bench MultilingualObserved 2026-11-09 | Configuration: claude-fable-5-2 (max effort, mean of 5 trials) Harness: Anthropic system-card standard configuration (scaffold not named); average of five trials Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.2: 300 problems across nine programming languages; 96.6%; average of five trials, max effort. | ||
| 74.1 | SWE-bench MultimodalObserved 2026-11-09 | Configuration: claude-fable-5-2 (max effort, mean of 5 trials) Harness: Anthropic system-card standard configuration (scaffold not named); average of five trials Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.2: visual context (screenshots and design mockups); 74.1%; average of five trials, max effort. | ||
| 74.3 | DeepSWE v1.1Observed 2026-11-09 | Configuration: claude-fable-5-2 (mean of 5 trials) Harness: Anthropic system-card standard configuration (scaffold not named); average of five trials Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.3: DeepSWE v1.1, 113 long-horizon tasks, 74.3% averaged over five trials. Evaluator is Anthropic (lab run), not the public Datacurve/mini-swe-agent leaderboard. Do not treat as a board restamp. | ||
| 67.6 | Terminal-Bench 4.0Observed 2026-11-09 | Configuration: claude-fable-5-2 (Claude Code --bare, max thinking, safeguards on) Harness: Claude Code --bare mode Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.5: 67.6% with safeguards enabled (fallback model answered flagged requests). Claude Code --bare mode at max thinking effort, averaged over 10 trials per task (660 trials), SE ±1.8 points. Mythos 5.2 without safeguards scored 70.5% (separate row). Do not merge the two. | ||
| 74.2 | HLE with toolsObserved 2026-11-09 | Configuration: claude-fable-5-2 (highest-available effort; with tools) Harness: Anthropic system-card configuration: web search, web fetch, programmatic tool calling, and code execution; thinking auto; 980k-token task budget; no context compaction; Opus 4.6 grader Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). Launch table and Table 8.1.A: 74.2% Humanity's Last Exam with tools. System card §8.11.1: with web search, web fetch, programmatic tool calling, and code execution; thinking auto; 980k-token task budget; no context compaction; Opus 4.6 grader; HLE-source blocklist and transcript screening. Table says all scores are each model's highest-available effort unless noted. | ||
| 73.2 | HLE no toolsObserved 2026-11-09 | Configuration: claude-fable-5-2 (highest-available effort; no tools) Harness: Anthropic system-card reasoning-only configuration (no tools); thinking auto; 980k-token task budget; no context compaction; Opus 4.6 grader Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). Table 8.1.A: 73.2% Humanity's Last Exam, no tools (reasoning-only). Same §8.11.1 setup minus tools. Do not write this onto hle or hle-with-tools. | ||
| 51.4 | AutomationBenchObserved 2026-11-09 | Configuration: claude-fable-5-2 (highest-available effort) Harness: Zapier AutomationBench 1.0.6 private held-out set (657 tasks); API default fallbacks; max effort Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). Launch table and Table 8.1.A: 51.4% AutomationBench. Self-reported; existing catalog slug automation-bench. | ||
| 91.6 | Chartography with toolsObserved 2026-11-09 | Configuration: claude-fable-5-2 (max effort, mean of 5 runs, with tools) Harness: Official Chartography tasks and expert ranges; tools: container, image file, standard libraries, and image cropping tool; Gemini 3.5 Flash grader; average of five runs Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.13.1 / Table 8.1.A: 91.6% with tools at max effort, average of five runs. Anthropic's own runs (not the Surge AI public figures). Existing catalog slug is chartography-with-tools. The no-tools 73.4% score is on the new slug chartography-no-tools. | ||
| 73.4 | Chartography no toolsObserved 2026-11-09 | Configuration: claude-fable-5-2 (max effort, mean of 5 runs, no tools) Harness: Official Chartography tasks and expert ranges; no tools; Gemini 3.5 Flash grader; average of five runs Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). Launch table and §8.13.1: 73.4% without tools at max effort, average of five runs, Anthropic's own runs. Medium-effort 66.4% is mentioned in the launch-post chart caption and is not this row. New slug: catalog only had chartography-with-tools. | ||
| 74.6 | Terminal-Bench-Science 0.1Observed 2026-11-09 | Configuration: claude-fable-5-2 (Claude Code --bare, max thinking, safeguards on) Harness: Claude Code --bare mode, no internet Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.6 and launch table: 74.6% with safeguards on. 70 Stanford-led tasks; 10 trials per task (700 trials); clustered SE ±4.3 points. Claude Code --bare, max thinking, no-internet-access setup as in Terminal-Bench 4.0. Fallback policy affected 1.3% of requests (4.1% of trials). | ||
| 69.9 | FrontierSWE v2Observed 2026-11-09 | Configuration: claude-fable-5-2 (max reasoning, Proximal harness, mean of 5 trials) Harness: Proximal agent harness Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.7 and launch table: 69.9%. Proximal FrontierSWE v2, 34 ultra-long-horizon tasks, 20-hour budget, max reasoning effort in Proximal's own agent harness, mean of five trials per task. Card says Fable 5.2 was first among the models Proximal evaluated. Owner site https://www.frontierswe.com/blog/v2 was captured today but did not list a Fable 5.2 score in the static HTML, so this row stays lab-published (evaluator Anthropic citing Proximal's eval) rather than a scraped owner-board restamp. | ||
| 82.1 | WANDR relaxed scoreObserved 2026-11-09 | Configuration: claude-fable-5-2 (highest-available effort; relaxed score) Harness: Anthropic modified WANDR run: web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget; offline frozen index; Opus 4.6 judge (soft F1) Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). Launch table and Table 8.1.A: 82.1% WANDR (relaxed score). Perplexity wide-search benchmark; Anthropic lab run (system card §8.11.3). Not an official Perplexity board capture today. | ||
| 89.3 | OSWorld 2.1 offline set (partial)Observed 2026-11-09 | Configuration: claude-fable-5-2 (max effort, Pass@1 mean of 5, 82-task official offline set) Harness: OSWorld 2.1 official offline set (82 of 108 tasks, no internet); default 1080p / 500-step settings; Pass@1 averaged over five attempts; Opus 4.8 model grader where needed Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.13.3 and launch table: 89.3% partial score on OSWorld 2.1 official offline set (82 of 108 tasks, no internet), max effort, Pass@1 averaged over five attempts, default 1080p / 500-step settings, Opus 4.8 as model grader where needed. Strict pass rate was 54.6% (not this row). Do NOT write onto osworld or osworld-verified — this is a 2.1 offline-set partial metric. | ||
| 41.3 | Remote Labor Index (RLI)Observed 2026-11-09 | Configuration: claude-fable-5-2 (default / high effort) Harness: Scale AI default configuration: high effort, 200k-token context window, 32k tokens per response Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). Launch table: 41.3% High effort. System card Table 8.1.A footnote: RLI uses default effort settings (Fable 5.2's default effort is high, so these agree). Lab-published. | ||
| 83.5 | BenchCAD 2.1 Core with toolsObserved 2026-11-09 | Configuration: claude-fable-5-2 (max effort, mean of 5, with tools) Harness: BenchCAD official scorer, tools (shell, file editor, image zoom) Evaluator: Anthropic Confidence: self-reported Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.13.2 and launch-post prose (not chart-only): 83.5 with tools at max effort on the Core 100-problem set, official scorer, average of five runs. Without tools the card reports 75.4% (not this row). Scores are not comparable with BenchCAD 1.0. New slug. | ||
| 79.3 | ARC-AGI-3 Standard harnessObserved 2026-11-09 | Configuration: claude-fable-5-2 (High, Standard harness) Harness: ARC Prize Standard harness Evaluator: ARC Prize Foundation Confidence: self-reported Notes: ARC Prize Foundation figures as reported in the Claude Fable 5.2 system card §8.15 (retrieved 2026-11-09): verified score 79.26% with the Standard harness at high effort (highest Standard-harness score published there). Other Foundation figures in the same section: Provider Adapter 97.72% at high; Standard 65.66% and Provider Adapter 94.90% at max. This is not ARC-AGI-2 and must not be written onto arc-agi-2. No separate arc-agi-3 file was found in the ARC v2.json dump (that file is ARC-AGI-2 semi-private). Owner attribution is the Foundation as quoted in the system card; treat as Foundation-reported rather than a live board scrape. |
Migration checks
No linked migration route is available for this model yet.
API versions
claude-fable-5-2Rankings & picks(4)
Created by
Pricing
Output / 1M
$40.00
Input / 1M
$8.00
Cheapest of 5 routes · Anthropic · cache read $0.200
Links