Overview
Claude Fable 5.1 is Anthropic's September 1, 2026 successor to Claude Fable 5. It remains the premium Claude model for demanding reasoning, coding, research, and long-running agent work, but it changes the deployment decision in three practical ways: stronger results at the same base token rates, cache reads at one quarter of the previous price, and API behavior changes that require migration testing.
The model is generally available as claude-fable-5-1 with a 1M-token context window, up to 128K output tokens, and adaptive thinking that is always on. Anthropic still recommends Claude Opus 5 for most workloads; Fable 5.1 is the escalation path when Opus at higher effort does not meet a team's evaluations.
What's New
Improves Long-Running Coding and Research
Anthropic reports the largest gains in agentic coding, multistep research, document and spreadsheet work, vision, long-context reasoning, and computer use. On its published evaluations, Fable 5.1 rises from Fable 5's 24.7% to 52.6% on Terminal-Bench-Science 0.1 and from 17.1% to 31.4% on AutomationBench. CursorBench 3.2.0 improves from 70.5% to 73.4%, while strict OSWorld 2.0 rises from 36.1% to 41.7%.
These are vendor-run results, not independent tests. They are still useful for deciding where to rerun internal evaluations: long autonomous coding sessions, research that follows multiple leads, and browser or desktop workflows with recovery steps.
Cuts Cache-Read Cost by 75%
Base input and output prices stay at $10 and $50 per million tokens. Cache reads fall from the Fable 5 rate of $1 to $0.25 per million tokens, while cache writes remain $12.50 for five-minute storage and $20 for one-hour storage. Anthropic estimates that the change reduces typical token-billed workload cost by about 25% and highly agentic workload cost by up to approximately 45%.
The saving matters most when an agent repeatedly reads a large cached prefix. It does not make every Fable task 25% cheaper: jobs without meaningful cache reuse will remain close to the previous base price.
Adds Per-Message Effort and Visible Progress Updates
Fable 5.1 can change effort during a conversation without invalidating the prompt cache through a beta API. It also supports turn-scoped system messages and a beta thinking.display: "updates" option that exposes short progress updates between tool calls while keeping private reasoning hidden. These controls let an application spend more reasoning on difficult steps, lower effort for routine turns, and show users that a long agent run is still active.
Reduces Safeguard False Positives
Fable 5.1 can identify software vulnerabilities in source code, although penetration testing, exploit generation, and binary-based vulnerability scanning still route to Opus models. Anthropic says Claude Code users should see around 60% fewer cyber-safeguard interventions per session, while Fable 5.1 inherits the biology-safeguard update already rolled out to Fable 5 in August 2026; that update reduced interventions on benign elementary-biology and medical requests by about 85% versus Fable 5's launch safeguards.
Adds Content Provenance
Text generated by Fable 5.1 carries Anthropic's statistical watermark. Supported image and video files that Claude produces can carry signed C2PA Content Credentials when retrieved through the Claude API Files API. Anthropic says the text watermark adds no tokens, hidden characters, or customer information and does not change output meaning or readability.
Performance & Cost Benchmarks
| Metric | Fable 5.1 | Fable 5 | Decision value |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | Re-evaluate scientific and terminal research agents |
| Terminal-Bench 4.0 | 55.8% | 42.0% | Test long tool-driven coding sessions |
| AutomationBench | 31.4% | 17.1% | Re-test multi-application business workflows |
| CursorBench 3.2.0 | 73.4% | 70.5% | Smaller but measurable coding-agent gain |
| OSWorld 2.0, strict | 41.7% | 36.1% | Better desktop task completion under strict scoring |
| Cache read price | $0.25/MTok | $1/MTok | 75% lower for reusable cached prefixes |
Anthropic evaluated Fable 5.1 with production safeguards enabled. Safeguard interventions can lower scores on affected cyber and biology tasks, so teams should not treat this table as a forecast for every workload.
Compared With Previous Version
| Area | Fable 5 | Fable 5.1 |
|---|---|---|
| Model ID | claude-fable-5 |
claude-fable-5-1 |
| Base API price | $10 input / $50 output per MTok | Unchanged |
| Cache reads | $1 per MTok | $0.25 per MTok |
| Tool choice | auto, none, any, and named tool |
auto and none; forced choices return HTTP 400 |
| Effort | Request-level control | Adds per-message effort in beta |
| Progress visibility | Progress blocks omitted or mixed with summaries | Adds thinking.display: "updates" in beta |
| Thinking-block binding | Earlier Fable behavior | Earlier turns, tools, and system prefixes must remain stable where enforcement applies |
| Cyber safeguards | More defensive-security false positives | Vulnerability discovery allowed; about 60% fewer interventions per Claude Code session |
| Output provenance | No 5.1 provenance behavior | Statistical text watermark and C2PA for supported files |
Migration Guide
Update the Model ID
Change claude-fable-5 to claude-fable-5-1, then test the complete agent loop rather than only a single response. Managed Agents require no integration change beyond the model name, but direct Messages API clients need the checks below.
Remove Forced Tool Choice
Fable 5.1 does not support tool_choice with type any or a named tool; those requests return a 400 invalid_request_error. Keep tool_choice: {"type": "auto"}, name the required tool in the instruction, and use strict tool schemas or structured outputs when schema-valid JSON is the real requirement.
Keep Thinking Histories Append-Only
Fable 5.1 can read thinking blocks from earlier Claude models, but earlier models cannot read Fable 5.1 thinking blocks. Where conversation-prefix enforcement applies, editing earlier messages, rebuilding the top-level system prompt or tools array, or replaying different bytes for a prior file can invalidate later thinking blocks.
Applications that construct their own message history should preserve prior turns byte-for-byte, move temporary instructions to turn-scoped system messages, and use server-side compaction or context editing. During migration, Anthropic recommends logging input_transformations with prefix_mismatch_behavior: "drop_block" to identify history edits before choosing an error or drop policy for production.
Re-Tune Effort and Tool Parallelism
The default effort is high on the API and Claude Code, but medium in Claude Cowork and claude.ai. Re-run quality, latency, and cost evaluations at the effort levels you plan to use. Fable 5.1 may issue fewer parallel tool calls when independence is only implied, so agent harnesses should explicitly request batching where extra round trips matter.
Availability & Access
Fable 5.1 is available to Pro, Max, Team, and Enterprise users and to API customers on the Claude Platform, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Platform identifiers differ: the Claude API uses claude-fable-5-1, while Amazon Bedrock uses anthropic.claude-fable-5-1.
The model carries 30-day data retention for safety monitoring by default. Eligible enterprise customers can use zero data retention until Enterprise Frontier Safeguards becomes available; Anthropic plans to roll out that customer-controlled storage and review system in phases beginning in fall 2026.
Pricing & Plans
| Item | Price |
|---|---|
| Input | $10 per million tokens |
| Output | $50 per million tokens |
| Five-minute cache write | $12.50 per million tokens |
| One-hour cache write | $20 per million tokens |
| Cache read | $0.25 per million tokens |
| Batch API | $5 input / $25 output per million tokens |
| US-only inference | 1.1x input and output pricing |
Fable 5.1 costs twice as much as Opus 5 on base input and output. Route routine traffic to Opus or Sonnet and reserve Fable 5.1 for jobs where internal evaluations justify the premium or cache reuse closes the cost gap.
Who Should Upgrade / Who Should Wait
Upgrade if you run multi-hour coding or research agents that already justify Fable pricing, reuse large cached prefixes, need better document or computer-use performance, or want finer effort and progress controls.
Wait if Opus 5 meets your evaluations, your workload rarely reuses prompt caches, you depend on forced tool choice, or your conversation store rewrites earlier messages and is not ready for thinking-block binding checks.
Known Limitations
- Forced
tool_choicevaluesanyand namedtoolare breaking API incompatibilities. - Long agent loops may use fewer parallel tool calls and therefore more turns unless prompts make batching explicit.
- Low effort can rely on memory more often instead of calling search or retrieval tools.
- Fable 5.1 may produce denser prose, less formatting, and more whole-file rewrites than Fable 5 without prompting adjustments.
- Safeguards still redirect penetration testing, exploit generation, binary vulnerability scanning, and some life-sciences research to Opus models.
- Default 30-day retention may not fit sensitive workloads without explicit enterprise authorization.
FAQ
Is Fable 5.1 cheaper than Fable 5?
Base input and output prices are unchanged. Cache reads are 75% cheaper, so the total saving depends on how often the workload reuses cached prefixes. Anthropic estimates about 25% lower cost for typical workloads and up to approximately 45% for highly agentic workloads.
Can Fable 5.1 replace Opus 5 for every request?
No, not for every request. Anthropic recommends Opus 5 for most workloads, and Fable 5.1 can refuse or fall back on safeguarded cyber and biology requests; its base input/output rates are also 2x Opus 5's. Use Fable 5.1 when Opus at higher effort still misses your quality target or when the new cache economics make the premium acceptable.
Does the Fable 5 API integration work unchanged?
Not always. The model ID changes, forced tool choice is unsupported, and applications that preserve thinking blocks may need append-only conversation handling. Re-run integration and behavioral evaluations before switching production traffic.
