Anthropic releases Claude Fable 5.1 for long-running agents

• Anthropic shipped Fable 5.1 as its broadly available long-horizon agent model, while Mythos 5.1 exposes the same underlying model through restricted cyber and life-sciences programs, according to Anthropic's availability post.
• Agentic science posted the sharpest first-party jump: Terminal-Bench-Science moved from 24.7% to 52.6%, as Anthropic's benchmark post reported.
• API economics changed at the cache layer, not the sticker price: cached input fell from $1 to $0.25 per million tokens, which ClaudeDevs' pricing note says can cut long agentic runs by 25% to 45%.
• Migration has real API breaks: ClaudeDevs' preserved-thinking notice says altering context before a thinking block can now fail a request, while the official migration guide removes forced tool choice.

An independent Artificial Analysis evaluation found that max effort emits about 1.7 times Fable 5's output tokens. The system card records Mythos 5.1 producing working Firefox exploits in 245 of 250 trials when safeguards were disabled. Anthropic's launch also describes Fable using old Magellan radar data to map a third of Venus at two-to-three-kilometer resolution.

• Fable 5.1 is live in Claude, Claude Code, and the Claude Platform, according to ClaudeDevs' launch note.
• The API model ID is claude-fable-5-1, with a 1M-token context window, 128K-token maximum output, and unchanged $10/M input plus $50/M output prices, per Anthropic's model page.
• Cache-read pricing is $0.25/M, a 75% reduction from Fable 5's $1/M rate, as bcherny's pricing update put it.
• Mythos 5.1 shares Fable 5.1's weights but carries more permissive domain safeguards for vetted users, according to the official system card.

First-party

• Terminal-Bench-Science 0.1: 24.7% → 52.6%, +27.9 points, as reported in Anthropic's benchmark post.
• Terminal-Bench 4.0 in Claude Code: 42.0% → 55.8%, +13.8 points, per ClaudeDevs' Terminal-Bench result.
• AutomationBench: 17.1% → 31.4%, +14.3 points, in the launch benchmark table.
• CursorBench 3.2: 70.5% → 73.4%, +2.9 points, in the launch benchmark table.

Third-party evaluators

• Artificial Analysis Intelligence Index: 62 → 66, +4 points, per the Artificial Analysis report.
• Humanity's Last Exam: 55.5% → 59.1%, +3.6 points, according to the Artificial Analysis report.
• Vals Index: 66.04% → 67.87%, +1.83 points, in Vals AI's results.

Customer-reported

• Box's complex enterprise-work evaluation: 65% → 72%, +7 points, in Box's enterprise evaluation.

Max effort is the cleanest cost caveat. Artificial Analysis measured $3.76 per Intelligence Index task for Fable 5.1 Max versus $3.14 for Fable 5 Max, a 20% increase driven by roughly 1.7 times as many output tokens.

The cache cut therefore pays off differently by harness and task length. According to Vals AI's comparison, Fable 5.1 took about 73 minutes and $28.40 per Index test, versus Opus 5's 56 minutes and $19.28; where the two tied on Legal Research, Vals measured a 3.4-times higher cost for Fable.

Anthropic's internal CoBench also placed Mythos 5.1 at 57.6% versus Opus 5 at 59.6%, a -2.0-point result in a historical-snapshot debugging evaluation, as shown in the CoBench chart. The system card says the model is still well short of Anthropic's 85% threshold for fully substituting for its research staff.

Safeguards still interrupt some legitimate evaluation traffic. ARC Prize said its ARC-AGI-3 API calls were repeatedly misclassified as reverse-engineering attempts, preventing completion before launch; one developer separately reported rewriting a theoretical-math session's “military campaign” metaphor to avoid cyber safeguards.

Subscription economics remain separate from API economics. ClaudeDevs' usage note says Claude Code subscription usage stays unchanged, and a Claude app notification says Fable 5.1 draws down usage faster than Opus 5 on Pro; ClaudeDevs' limit-reset post announced a five-hour and weekly reset on release day, not a larger ongoing quota.

Fable and Mythos are policy configurations around identical weights, per the official system card. For cyber traffic, the card describes an activation probe that escalates flagged requests to a trained LLM classifier, as captured in a system-card excerpt.

The migration guide resolves the API questions:

• Adaptive thinking is always on. The model chooses its reasoning depth, with low, medium, high, xhigh, and max effort levels.
• tool_choice: {type: "auto"} and none remain supported, but any and a named forced tool now return HTTP 400.
• Editing an earlier system prompt, tool definition, or message invalidates preceding thinking. New accounts created on or after August 31 receive an error by default, while prefix_mismatch_behavior: "drop_block" continues without showing the invalidated reasoning to the model, per Anthropic's preserved-thinking documentation.
• Fable 5.1 can read thinking blocks produced by earlier Claude models, but earlier models cannot read Fable 5.1 blocks.
• Effort can change mid-conversation without invalidating prompt cache, according to RLanceMartin's migration tips.

Anthropic also documents behavioral shifts that appear without an API change: more variable parallel tool calls, fewer progress updates, less tool-based retrieval at low effort, denser chat prose, and more whole-file rewrites for small edits.

Every's early-access testing found comparable Slack-agent results to Opus 5 with about half the tokens and roughly 60% of the runtime, while danshipper's early-access report also described one-shot builds and multi-day coding jobs.

Simon Willison's SVG pelican test found its best result at Max effort, but his writeup logged a $3.30 run lasting nearly 14 minutes.

In T3 Code, theo's initial report said Fable 5.1 caught mistakes in earlier Fable 5 code, but also made an obviously wrong assumption about auto-settle logic. emollick's early-access note described a meaningful advance in long-run work requiring judgment and taste, with a smaller improvement in Claude's characteristic style.

• Cursor added Fable 5.1 on day one and reported 73.4% at max effort on CursorBench 3.2, per Cursor's launch note.
• Devin added Fable 5.1 to Fusion, where Cognition's announcement reported a 63.6 FrontierCode score at $2.68 per rollout, versus $5.84 for Fable 5 Medium.
• Lovable deployed it for live-app iteration, reporting a 17% lift on difficult tasks and 31% lower cost than Fable 5 in Lovable's deployment post.
• Perplexity Computer made it available to Pro and Max subscribers, according to Perplexity's launch note.
• Warp shipped it in Warp and the Warp Agent CLI, as Warp's release note announced; Vercel added the model to AI Gateway in Vercel's gateway post.

Anthropic says Enterprise Frontier Safeguards will roll out in phases this fall. Eligible customers can use Fable 5.1 with zero data retention during the transition, according to Anthropic's EFS announcement.

The system stores data in the customer's own cloud under the customer's encryption keys, access policies, and audit logs. Automated monitoring can correlate risky behavior across sessions and accounts, then sends signals to the customer's team for review rather than to Anthropic staff, as eugeneyan's EFS summary describes.