Fable 5.1 launches in Claude Code at Fable 5 pricing

• Fable 5.1 is live in Claude Code and the Claude Platform at Fable 5 input and output pricing, while ClaudeDevs' price update cuts API cache reads from $1 to $0.25 per million tokens.
• Anthropic's benchmark chart puts Terminal-Bench 4.0 at 42.0% → 55.8%, a +13.8-point jump over Fable 5, according to ClaudeDevs' launch post.
• The release changes agent economics: Anthropic estimates 25% lower token-billed costs for typical work and up to 45% for highly agentic work in its launch announcement.
• API transcript edits are a migration gotcha, because the preserved-thinking documentation says altered prior prompts, tools, or messages can now cause an error.
• Creative tests are already centering on self-checking loops: om_patel5's city-builder account describes a 14-agent browser build that used screenshots, performance telemetry, and a separate critic agent.

One reported two-prompt Rocket League-style game took about 20 minutes, versus more than an hour on Fable 5, while LLMJunky's game-design test says the agent found and fixed most of its own visible bugs before returning control. linear's announcement puts the model in Linear coding sessions. The model page gives the API model a 1M-token context window and 128K-token maximum output.

• Fable 5.1 is generally available in Claude Code and the Claude Platform at Fable 5's input and output rates, per ClaudeDevs' launch post.
• Claude Code users select it in the model picker; API users call claude-fable-5-1, as ClaudeDevs' migration post specifies.
• The official model page lists $10 per MTok input, $50 per MTok output, 1M context, and 128K maximum output.
• Cache reads fell $1 → $0.25 per MTok, a 75% cut, in ClaudeDevs' price update.
• The new cache rate applies to API, SDK, and Enterprise customers; bcherny's reply says Pro, Max, and Team subscriptions were unchanged because they already have discounted cache reads.
• Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguards; the launch announcement makes Fable general availability and keeps Mythos in trusted-access programs.
• Anthropic reset five-hour and weekly limits for all users, ClaudeDevs' limit reset says.
• Claude Cowork and Claude Tag also received the model, according to _catwu's product post.

First-party

• Terminal-Bench 4.0: 42.0% → 55.8%, +13.8 points, per ClaudeDevs' benchmark chart.
• Terminal-Bench-Science 0.1: 24.7% → 52.6%, +27.9 points, per ClaudeDevs' benchmark chart, which lists a ±3.5 to 4.5-point standard error for this test.
• OSWorld 2.0, partial: 72.9% → 77.9%, +5.0 points, per ClaudeDevs' benchmark chart.
• AutomationBench: 17.1% → 31.4%, +14.3 points, per ClaudeDevs' benchmark chart.
• CursorBench 3.2.0: 70.5% → 73.4%, +2.9 points, per ClaudeDevs' benchmark chart.

Third-party evaluators

Customer-reported

• Comparable Slack-agent results: Opus 5 baseline token use → about half as many tokens, about 50% fewer tokens, in the Every hands-on test.
• Comparable Slack-agent runtime: Opus 5 baseline → about 60% of the time, about 40% faster, according to the Every hands-on test.

Fable's safeguarded configuration scored 55.8% on Terminal-Bench 4.0, versus 60.9% for Mythos 5.1, a 5.1-point gap between two configurations of the same model. Anthropic attributes the split to safeguards in its launch announcement.

Preserved-thinking binding is a concrete API compatibility break. The official documentation says changing an earlier system prompt, tool definition, or message defaults to an error for new accounts created after August 31, unless prefix_mismatch_behavior: "drop_block" removes the affected thinking blocks.

The migration guide also flags forced tool choice: it returns HTTP 400 on 5.1. A commenter in the Hacker News launch discussion separately reported prompt instructions being forgotten after context compaction.

The desktop announcement says Fable is included for up to 50% of Max-plan usage the in-app announcement screenshot. Separately, thekitze's limit report says a $200-plan session hit its cap in under 30 minutes.

The prompting guide lists five effort settings, with high as the default:

• low
• medium
• high
• xhigh
• max

At low effort, the guide says Fable calls search and retrieval tools less often. At xhigh and max, it can spend longer thinking before returning long outputs in the prompting guide. bcherny's effort-level reply describes medium as less intelligent but substantially lighter on tokens and latency.

The safety stack also changed:

• Cyber-related fallbacks to Opus are down about 40% from Fable 5's current rate, and 55% from Fable 5's launch rate, per ClaudeDevs' safeguard update.
• Benign biology safeguards intervene 85% less often, and Claude Code users should see about 60% fewer cyber interventions per session, bcherny's safeguard update says.

bcherny said he had run the model for weeks in bcherny's reply and called it Anthropic's best model for coding, data analysis, computer use, design, presentations, Tag, and long-running work in bcherny's launch post. Independent hands-on posts are more useful for the workflow details:

• The Every team says one prompt rebuilt its Proof document editor, while a separate one-shot built a Mac computer-use app called Hands in the Every hands-on test.
• LLMJunky says a two-prompt Rocket League-style browser build ran in roughly 20 minutes rather than over an hour on Fable 5, then reviewed earlier flawed passes and removed most bugs before yielding LLMJunky's game-design test.
• trq212 reported that low effort suits tasks with fewer edge cases, and that changing effort no longer breaks the prompt cache trq212's early test.
• petergyang used /claude-api prompt-audit on existing skills to find redundant rules for the new models petergyang's prompt-audit post.

According to om_patel5's city-builder account, a clean Claude Code install received one prompt, spawned 14 agents, and spent about $203 over roughly 90 minutes building the first wave of a browser city builder. The account describes five controls around the agent swarm:

1. Write an architecture document and create one folder per subsystem before feature code.
2. Use headless Chrome to screenshot the running game and log frame rate, draw calls, and console errors.
3. Dispatch one builder agent per module in dependency waves.
4. Give a critic agent screenshots at different times of day, compared against Cities: Skylines references; a module needed an 8.5/10 and zero errors or received a ranked issue list.
5. Restrict core-file changes to a dedicated integrator agent.

The same account says the build sourced free PBR textures, tested materials in Blender, generated terrain, water, sky, cloud, and particle systems in code, and synthesized nearly all sound. The first wave still had no buildings or traffic.

Within a day of launch, minchoi's example roundup collected Fable-attributed game and 3D demos with linked clips:

• A Mario Kart-style game minchoi's Mario Kart post
• A medieval kingdom with 1,500 soldiers minchoi's medieval-kingdom post
• A Minecraft recreation minchoi's Minecraft post
• A Call of Duty-style browser shooter minchoi's shooter post
• A Rocket League-style game in two prompts minchoi's Rocket League post
• A code-generated fantasy Mount Fuji world minchoi's Mount Fuji post
• An Airbus helicopter recreation in Three.js minchoi's helicopter post

Linear announced availability in its coding sessions linear's announcement. LiteLLM also published day-zero support across Anthropic, Bedrock, Gemini Enterprise Agent Platform, and Azure, including spend controls, rate limits, fallbacks, and logging.