Claude Fable 5.1 vs Fable 5: What Changed and Should You Upgrade?

Claude Fable 5.1 vs. Fable 5: Should You Migrate?

Claude Fable 5.1 replaced Fable 5 on September 1, 2026, at the same $10 and $50 per million-token prices. Cache reads dropped from $1 to $0.25 per million tokens, and every benchmark Anthropic published improved. Fable 5 remains available, with Anthropic’s Fable 5 model page committing to support through at least June 9, 2027. The practical question is whether Fable 5.1’s gains justify the migration work—and whether its breaking changes affect your application.

Try Apidog today

This guide covers what stayed the same, what improved, what changed, and how to choose by workload. Sources include Anthropic’s launch post and What’s new in Claude Fable 5.1. For the full specification, see what Claude Fable 5.1 is.

Fable 5.1 vs. Fable 5

Specification	Claude Fable 5.1	Claude Fable 5
Model ID	claude-fable-5-1	claude-fable-5
Released	September 1, 2026	June 9, 2026 (suspended June 12, back July 1)
Retirement	Not sooner than September 1, 2027	Not sooner than June 9, 2027
Context / maximum output	1M / 128K	1M / 128K
Knowledge cutoff	June 2026	January 2026
Input / output price	$10 / $50	$10 / $50
Cache read	$0.25 per million	$1.00 per million
Cache writes	$12.50 (5m), $20 (1h)	Same
Batch	$5 / $25	Same
Tokenizer	Same (Opus 4.7 generation)	Same
Thinking	Adaptive, always on	Adaptive, always on
Forced tool_choice	400	Accepted
Thinking blocks readable by older models	No	Yes
History-editing check	Yes, enforced for accounts created on or after 2026-08-31	No
Priority Tier	Not supported	Supported
Per-message effort	Yes (beta)	No
Turn-scoped system messages	Yes (beta)	No
thinking.display: "updates"	Yes (beta)	Yes (beta)
Content provenance	Text watermark + C2PA on media	Text watermark
Data retention	30-day required, Covered Model	Same
Refusal categories	Same five	Same five

What stayed the same

The migration is relatively small for most applications. Both models retain:

• The same API surface
• A 1M-token context window and 128K maximum output
• The same input, output, cache-write, and batch prices
• The same tokenizer
• A 512-token cache minimum
• Always-on adaptive thinking
• "omitted" as the default for thinking.display
• No prefilling
• No sampling parameters
• Mid-conversation system messages
• The same refusal categories and fallback mechanics

Anthropic expects existing Fable 5 prompts to work well on Fable 5.1 without modification.

Existing guidance also remains relevant, including the rate limits guide, safeguards explainer, and cloud availability guide. The models also share one rate-limit pool.

What improved

Benchmark performance

Every benchmark row published by Anthropic improved:

Benchmark	Fable 5.1	Fable 5	Delta
Terminal-Bench-Science 0.1	52.6%	24.7%	+27.9
AutomationBench	31.4%	17.1%	+14.3
Terminal-Bench 4.0	55.8%	42.0%	+13.8
GDPval-AA v2	1853	1723	+130 Elo
OSWorld 2.0 (strict)	41.7%	36.1%	+5.6
OSWorld 2.0 (partial)	77.9%	72.9%	+5.0
Humanity’s Last Exam (no tools)	60.9%	57.8%	+3.1
CursorBench 3.2.0	73.4%	70.5%	+2.9
Humanity’s Last Exam (with tools)	65.0%	63.8%	+1.2

The largest gains are on long-horizon agentic tasks:

• Scientific research in a terminal
• Business workflow automation
• Agentic coding

Knowledge tasks and IDE-style coding show smaller improvements. These are vendor-run results and have not yet been independently reproduced. See the benchmarks breakdown for a row-by-row analysis.

Cache reads cost 75% less

According to the pricing page, cache reads are the most immediate cost improvement for many Fable 5 users.

A 150,000-token prefix reread on every turn costs:

• Fable 5: $0.15 per turn
• Fable 5.1: $0.0375 per turn

Anthropic estimates that typical workloads become 25% cheaper, with savings reaching 45% for agentic workloads. The pricing breakdown includes three worked examples.

Fewer safeguard false positives

Anthropic reports that Fable 5.1’s biology classifiers trigger 85% less often on benign elementary biology and medical requests than Fable 5’s did at launch.

Its cyber safeguards also produce approximately 60% fewer interventions per Claude Code session. Fable 5.1 can identify vulnerabilities in source code that Fable 5 previously declined to analyze. Exploit development remains refused.

More useful effort levels

Anthropic says medium effort on Fable 5.1 roughly matches Fable 5 at lower cost. The biggest quality gains appear at xhigh and max.

A practical migration strategy is to:

1. Move a Fable 5 workload from high to Fable 5.1 medium for similar quality and fewer output tokens.
2. Keep high when you want higher quality than Fable 5 provided.
3. Re-run your own evaluation suite rather than assuming effort levels map exactly between models.

Newer knowledge

Fable 5.1 has a June 2026 knowledge cutoff, compared with January 2026 for Fable 5. That adds five months of framework releases, API changes, and model launches.

New harness features

Fable 5.1 adds:

• Per-message effort changes without resetting the cache
• Turn-scoped system messages for per-turn reminders
• Readable progress updates between tool calls

Fable 5 supports the last feature through thinking.display: "updates"; the first two are new.

What changed

Forced tool use now returns an error

On Fable 5.1, tool_choice: "any" and tool_choice: "tool" return HTTP 400 errors.

If you used forced tools to guarantee JSON, migrate to either:

{
  "tool_choice": {
    "type": "auto"
  },
  "output_config": {
    "format": {
      "type": "json_schema",
      "strict": true
    }
  }
}

Or use structured outputs directly.

If your application requires a tool call on a particular turn, add a mid-conversation system message that explicitly requires it. Audit every forced-tool call site before switching models.

Thinking blocks are one-way

Fable 5.1 can read thinking blocks created by Fable 5, so conversations can move upward without losing reasoning.

The reverse is not true. Fable 5 cannot read Fable 5.1 thinking blocks. During a gradual migration, if a router sends a conversation back to Fable 5:

• The API drops the incompatible thinking blocks
• The dropped blocks are not billed
• Fable 5 replans the turn

Account for this behavior in routing and cost estimates.

History editing is validated

Fable 5.1 thinking blocks are valid only in the exact conversation that produced them. On accounts created on or after August 31, 2026, the next request fails if your application:

• Edits an earlier turn
• Rebuilds the system prompt
• Changes the tools array
• Injects or removes per-turn reminders
• Refreshes dates or other dynamic system-prompt content
• Compacts the client-side history while replaying recent turns verbatim

Run the three-step history check before migrating. Fix every prefix_binding_mismatch error. The preserved thinking guide provides a complete checklist.

Priority Tier is unavailable

Fable 5.1 does not support Priority Tier. If your enterprise workload depends on guaranteed capacity, remain on Fable 5 until Anthropic adds support.

Agent loops behave differently

At some effort levels, Fable 5.1 may:

• Issue one tool call per turn instead of batching several
• Write fewer progress updates
• Answer from memory more often instead of searching at low effort

Each behavior can be adjusted with a prompt change. See the prompting guide.

Migration decision rule

Upgrade now

Upgrade immediately for:

• Long-running agentic coding
• Multistep research
• Document generation
• Spreadsheet generation

These workloads benefit from the largest benchmark gains, cheaper cache reads, and fewer safeguard false positives. Cognition also said it was moving its Devin traffic to Fable 5.1 on launch day.

Upgrade after auditing your harness

If your application builds the messages array manually and does more than append new turns, audit history handling first:

1. Replay conversations with preserved thinking blocks.
2. Test system-prompt and tool-array changes.
3. Fix every prefix_binding_mismatch.
4. Re-run production-like evaluations.
5. Migrate only after the checks pass.

Move from high to medium

If Fable 5 at high handles routine tasks well, test Fable 5.1 at medium. You may get similar quality with fewer output tokens and lower total cost.

Stay on Fable 5

Stay on Fable 5 when:

• Priority Tier is required
• The workload is output-heavy with little cached input
• Migration overhead exceeds the expected quality improvement

Fable 5 is supported until at least June 9, 2027, on Anthropic-operated platforms. Bedrock and Google Cloud may set different retirement dates.

Consider Opus 5

If Fable 5 was already more capable than your workload required, evaluate Opus 5 instead. Anthropic recommends starting with Opus 5 and moving up only when it falls short. The Fable 5.1 vs. Opus 5 comparison covers the option at half the price.

Test both models in Apidog

The most reliable migration decision comes from your own production-like prompts.

In Apidog:

1. Save your ten hardest production prompts as requests.
2. Store the model ID in an environment variable.
3. Run the collection against claude-fable-5.
4. Run it again against claude-fable-5-1 at the same effort level.
5. Compare the response text, usage.output_tokens, and usage.cache_read_input_tokens on the second run.
6. Add a forced tool_choice request to confirm that your test suite catches the new 400 error.

Download Apidog to get started. The API walkthrough includes the request bodies.

FAQ

Is Claude Fable 5.1 a new model or an update to Fable 5?

It is a new model with its own ID, claude-fable-5-1, and its own retirement date. It uses the same tokenizer and per-token pricing as Fable 5 and is trained on data through June 2026.

Does Fable 5.1 cost more than Fable 5?

No. Input and output pricing remains $10 and $50 per million tokens. Cache writes and batch rates are also unchanged. Cache reads are cheaper: $0.25 per million tokens instead of $1.

Will my Fable 5 code work on Fable 5.1?

Most of it will. Forced tool_choice returns a 400, and applications that edit conversation history between requests can invalidate thinking blocks on accounts created on or after August 31, 2026. The remaining API behavior carries over.

Is Fable 5 being retired?

Not yet. Anthropic will not retire Fable 5 before June 9, 2027, on Anthropic-operated platforms. Bedrock and Google Cloud set their own dates.

Which model is better at coding?

On Anthropic’s benchmarks, Fable 5.1 performs better:

• Terminal-Bench 4.0: 55.8% vs. 42.0%
• CursorBench 3.2.0: 73.4% vs. 70.5%

The agentic-terminal improvement is large, while the IDE-style improvement is smaller. Both results are vendor-run.

Does Fable 5.1 refuse less often?

Anthropic reports 85% fewer biology false positives and approximately 60% fewer cyber interventions per Claude Code session. The refusal categories remain the same, so keep your refusal handling and fallback logic in place.