Claude Fable 5.1: Pricing, Benchmarks, and Migration Risks
Anthropic's Claude Fable 5.1 cuts cache costs 75%, beats Opus 5 on agentic benchmarks, and ships three breaking changes for Fable 5 API users.
Anthropic shipped Claude Fable 5.1 on September 1, 2026, alongside Claude Mythos 5.1. The two share one underlying model with different safeguards: Fable 5.1 is generally available, while Mythos 5.1 is limited to trusted access programs, with safeguards designed for cybersecurity and life sciences. For builders, the short version: per-token prices are unchanged from Fable 5, cache reads dropped 75%, and three breaking changes will bite unprepared migrations.
What happened
Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Fable 5.1 is generally available under model ID claude-fable-5-1 on the Claude API, Amazon Bedrock (anthropic.claude-fable-5-1), Google Cloud (claude-fable-5-1), Microsoft Foundry (claude-fable-5-1), and Claude Platform on AWS (claude-fable-5-1), plus marketplaces, and for Pro, Max, Team, and Enterprise users. Mythos 5.1 (claude-mythos-5-1) is restricted to Project Glasswing participants.
1M token context and 128K max output, with the full 1M window billed at standard per-token pricing
$10/MTok input, $50/MTok output — unchanged from Fable 5
Cache reads at $0.25/MTok, 75% less than Fable 5
Estimated 25% less than Fable 5 for typical token-billed workloads because cache reads are cheaper; highly agentic work often up to approximately 45% less
Adaptive thinking always on; default effort high
Defaults inside Anthropic's own surfaces differ: High effort in Claude Code, Medium in Claude Cowork and on Claude.ai.
Why it matters
The model docs are unusually direct about who should buy this. Per Anthropic: for most workloads, start with Claude Opus 5. Use Fable 5.1 when evals on Opus 5 at higher effort still fall short. Fable 5.1 runs $10/$50 per MTok against Opus 5's $5/$25, and Anthropic rates its latency as slower than Opus 5's moderate. The premium buys demanding reasoning and long-horizon agentic capability.
The cost story is the real news. Cache reads on Fable 5.1 and Mythos 5.1 cost 2.5% of base input price versus 10% on other Claude models. That is why Anthropic estimates typical token-billed workloads at roughly 25% less than Fable 5, and highly agentic work often up to approximately 45% less. Cache-heavy agentic loops, where every step re-reads context, are where the savings compound.
On benchmarks with production safeguards enabled, Fable 5.1 posts higher numbers than Opus 5 everywhere Anthropic reported both: 55.8% vs 52.3% on Terminal-Bench 4.0, 73.4% vs 70.0% on CursorBench 3.2.0, 31.4% vs 26.9% on AutomationBench, and 77.9% vs 75.4% on OSWorld 2.0 partial. The widest margin is Terminal-Bench-Science 0.1 (52.6% vs 29.0%); the narrowest is Humanity's Last Exam with tools (65.0% vs 63.6%). The largest gap is on agentic scientific research.
Safeguards shape what you can build. Safeguards may auto-route many cyber and biology queries to less capable models. Users are not charged Fable prices for rerouted requests — but rerouted requests are also not answered by Fable. Fable 5.1 can discover software vulnerabilities but cannot develop exploits.
The details
Model basics. Per the official overview: released September 1, 2026; knowledge cutoff June 2026; context 1M tokens; max output 128K tokens; input text and images, output text; adaptive thinking always on; default effort high; latency moderately slower than Opus 5, while Sonnet 5 is the fast option; fallback targets Opus 4.8 and Opus 5; retirement not sooner than September 1, 2027.
Pricing. Input $10/MTok, output $50/MTok. Cache read $0.25/MTok. Cache writes: $12.50/MTok for 5-minute and $20/MTok for 1-hour. Batch API is 50% off — $5 input and $25 output per MTok. US-only inference is priced at 1.1x on input and output. The rest of the lineup, for comparison: Opus 5 at $5/$25, Sonnet 5 at $2/$10, and Haiku 4.5 at $1/$5 with 200K context and 64K max output.
Effort and new API surface. Effort levels run low, medium, high, xhigh, and max; thinking is adaptive, always on, and steered by effort. Three betas ship alongside: per-message effort (mid-conversation-output-config-2026-07-01), turn-scoped system messages (mid-conversation-system-clear-at-2026-08-21), and progress display updates (thinking-display-updates-2026-08-18). Content provenance is also new: a statistical text watermark, plus C2PA on supported media via the Files API.
Breaking changes vs Fable 5 — three, per the what's-new docs:
Forced tool use errors.tool_choice of type any or tool now returns a 400. auto and none are unchanged. Reason: thinking is always on, and forced tools would skip thinking.
One-way thinking interop. Fable 5.1 can read earlier models' thinking blocks; earlier models cannot read Fable 5.1's. Pipelines that hand thinking blocks between models break in one direction.
Editing earlier turns invalidates thinking. Editing the system prompt, tools, or earlier messages invalidates thinking blocks. The check is enforced for accounts created on or after August 31, 2026; older accounts hit it only when prefix_mismatch_behavior is set. Mythos 5.1 does not run this check.
Behavior shifts vs Fable 5. More variable parallel tool calling; fewer progress updates; answers from memory more often at low effort; denser prose; less chat formatting; unmarked quotations in summaries; whole-file rewrites for small edits.
What did not change. The enabled/disabled thinking type still returns a 400; display is omitted by default; no prefilling; non-default temperature/top_p/top_k still return a 400; 512-token cache minimum. The tokenizer is unchanged from Fable 5 — the Opus 4.7 tokenizer — which yields roughly 30% more tokens than older models, so token budgets ported from pre-Fable-5 models will not stretch as far as character counts suggest.
Data handling. The sources word this differently, so here are both, precisely. Anthropic's Fable product page describes default 30-day data retention for safety monitoring. The what's-new docs list Fable 5.1 under Covered Models with 30-day data retention and state it is not available under zero data retention (ZDR) unless expressly authorized by Anthropic. The announcement adds the bridge: until Enterprise Frontier Safeguards (EFS) is available, eligible customers can use Fable 5.1 with zero data retention. EFS stores data on customer-controlled cloud infrastructure, with human review handled by the customer by default, and rolls out in phases beginning later this fall.
Benchmarks. Scores with production safeguards enabled, per Anthropic:
Benchmark
Fable 5.1
Mythos 5.1
Opus 5
Fable 5
GPT-5.6 Sol
Terminal-Bench 4.0
55.8%
60.9%
52.3%
42.0%
37.3%
Terminal-Bench-Science 0.1
52.6%
—
29.0%
24.7%
22.4%
GDPval-AA v2
1853
—
1824
1723
1711
OSWorld 2.0 (partial)
77.9%
—
75.4%
72.9%
—
OSWorld 2.0 (strict)
41.7%
—
39.6%
36.1%
—
Humanity's Last Exam (no tools)
60.9%
—
56.6%
57.8%
—
Humanity's Last Exam (with tools)
65.0%
—
63.6%
63.8%
—
AutomationBench
31.4%
—
26.9%
17.1%
19.6%
CursorBench 3.2.0
73.4%
—
70.0%
70.5%
67.2%
Caveats Anthropic flags: Terminal-Bench-Science 0.1 carries a standard error of ±3.5–4.5 points, and public leaderboard numbers differ — Opus 5 at 30.0% and Fable 5 at 21.4%, against Anthropic's setup at 29.0% and 24.7%, within noise. OSWorld 2.0 has no competitor score because the August 2026 task release is not comparable to prior OSWorld 2.0 numbers. Multilingual performance is on par with Fable 5. Gains concentrate at higher effort across long-session agentic coding, document/spreadsheet/slide work, research and search, vision, long-context, and computer use. Mythos 5.1 scored higher still on Terminal-Bench 4.0, at 60.9%.
What to watch
EFS rollout. Phases begin later this fall. Until EFS reaches a given account, eligible customers can run Fable 5.1 with zero data retention.
Mythos 5.1 access. A biology access program with the US government covers Mythos 5.1, with enrollment for scientists expected soon. The Cyber Verification Program will include Mythos in the near future; cyberdefenders can apply.
Cyber safeguards. Anthropic says its newest cybersecurity safeguards block 60% fewer false positives than before. Fable 5.1 can discover software vulnerabilities but not develop exploits — plan security tooling accordingly.
Rerouting. Many cyber and biology queries may be auto-routed to less capable models, without Fable pricing on rerouted requests. Factor this into eval design: a passing suite does not prove Fable answered every request.
Lifecycle. Retirement is not sooner than September 1, 2027. Anthropic has also published a system card for Fable 5.1 and Mythos 5.1.
FAQ
Who should use Fable 5.1 instead of Opus 5?
Anthropic's guidance: start with Opus 5 for most workloads; move to Fable 5.1 when evals on Opus 5 at higher effort still fall short. Fable 5.1 targets demanding reasoning and long-horizon agentic work — long-session agentic coding, multistep research, document/spreadsheet/slide work, and computer use. The trade-offs are real: double the per-token price ($10/$50 vs $5/$25 per MTok) and slower latency. Cache-heavy agentic pipelines are the case where total cost can still drop, because $0.25/MTok cache reads are 75% cheaper than Fable 5's.
What does Fable 5.1 actually cost?
List price: $10/MTok input and $50/MTok output — unchanged from Fable 5. Cache reads cost $0.25/MTok, which is 2.5% of base input price versus 10% on other Claude models. Cache writes run $12.50/MTok for 5-minute and $20/MTok for 1-hour. Batch API is 50% off: $5 input and $25 output per MTok. US-only inference is billed at 1.1x on input and output. For context: Opus 5 is $5/$25, Sonnet 5 is $2/$10, and Haiku 4.5 is $1/$5 with 200K context and 64K max output.
What breaks if you already call Fable 5?
Three breaking changes. First, forced tool use: tool_choice of type any or tool returns a 400 (auto and none unchanged), because thinking is always on and forced tools would skip it. Second, earlier models cannot read Fable 5.1's thinking blocks, though Fable 5.1 can read earlier models' blocks. Third, editing the system prompt, tools, or earlier messages invalidates thinking blocks — enforced for accounts created on or after August 31, 2026, and for older accounts only when prefix_mismatch_behavior is set (Mythos 5.1 skips the check). Expect behavior shifts too: more variable parallel tool calling, fewer progress updates, denser prose, less chat formatting, unmarked quotations in summaries, whole-file rewrites for small edits, and more answers from memory at low effort. Unchanged: enabled/disabled thinking types still 400, no prefilling, non-default temperature/top_p/top_k still 400, the 512-token cache minimum, and the same tokenizer as Fable 5.
