by Anthropic · Claude 5 family · best for long-horizon agent work where Opus 5 falls short
Claude Fable 5.1 is Anthropic's top generally available model, released 2026-09-01 as the direct successor to Claude Fable 5 in the Fable tier that sits above Opus. It is the same underlying model as Claude Mythos 5.1; Fable 5.1 carries safety classifiers for cybersecurity, biology and model-distillation risk, while Mythos 5.1 is offered only to vetted organizations through Project Glasswing and Anthropic's trusted-access programs. Anthropic's own tables put it ahead of Fable 5, Claude Opus 5 and GPT-5.6 Sol on long-running agentic coding, scientific research, knowledge work and computer use. The list price is unchanged at $10 / $50 per million tokens, but cache reads fall 75% to $0.25, which Anthropic says cuts typical bills by about 25% and agent-heavy bills by up to 45%. The sentence a buyer needs: this is the model to escalate to when Opus 5 at high effort still falls short, as long as you can accept 30-day data retention.
stop_reason: "refusal" with a named category, optional server-side fallback (beta) to Opus 4.8 or Opus 5, and no charge for a refusal that arrives before any output.tool_choice (any, tool) returns a 400; thinking blocks can be read only by Fable 5.1 or a newer model; editing earlier turns invalidates later thinking blocks (enforced as a 400 for accounts created on or after 2026-08-31).display: "updates" mode that returns progress updates between tool calls as text.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 60.9% | anthropic.com 2026-09-01T00:00:00.000Z |
| LMArena Elo | 1498 | lmarena.ai 2026-09-13T00:00:00.000Z |
| LMArena Coding Elo | 1519 | lmarena.ai 2026-09-13T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The strongest model you can buy outright, now with a zero-retention path for eligible firms. Default to Opus 5 and escalate the hard work here.”
Fable 5.1 addresses the two objections that kept Fable 5 out of many procurement reviews. Effective cost falls about 25% on typical workloads through cheaper cache reads, and eligible enterprises can run it under zero data retention until Enterprise Frontier Safeguards, which keep data in the customer's own cloud, arrive in phases from this fall. The default is still 30-day retention, and Anthropic's own guidance is to start with Opus 5 and escalate only when evals fall short, so plan a two-model estate. Vendor risk is low: five platforms on day one, a retirement commitment to at least 2027-09-01, and a clear tiering strategy.
“Anthropic's lead now rests on long unattended agent work, not chat preference: it tops agent evals while LMArena voters still favor Fable 5.”
Fable 5.1 keeps Anthropic on top of the agentic-work evaluations it publishes: Terminal-Bench-Science 52.6% against 22.4% for GPT-5.6 Sol, GDPval-AA v2 1853 against 1711, AutomationBench 31.4% against 19.6%. Even xAI's Grok 4.7 launch table concedes it the lead on CursorBench 4.0 and Terminal-Bench 4.0. The moat is long unattended runs, and partner reports from Cognition, Ramp and Shopify describe exactly that. Two pressures matter: GPT-6 Astra arrived two days later at the identical $10 / $50 list price, and the human-preference arenas show no gain over Fable 5. The position is strong but no longer uncontested.
“Same $10/$50 rate card, but $0.25 cache reads cut agent-heavy bills by up to 45%. It is still twice Opus 5 per fresh token, so route with care.”
The rate card did not move; the cache price did. At $0.25 per million cached tokens, Fable 5.1 reads cache at half the Opus 5 rate and a quarter of Fable 5's, and Anthropic measured roughly 25% lower bills on typical August 2026 workloads and up to 45% on context-heavy agent loops. Fresh input and output still cost double Opus 5, and higher effort multiplies output spend. Batch halves both rates; US-only inference adds 10%. The Pro plan reaches Fable only through usage credits, so seat budgets need a metered line. Refusals before output are free, and fallback attempts bill at the answering model's rates. Build a router: Opus 5 by default, Fable 5.1 for designated hard work.
“Per-message effort and cheap cache reads suit agents, but forced tool_choice is gone and edited histories now fail with 400s. Test before you switch.”
For agent builders this upgrade is not a pure model-ID swap. Forced tool_choice returns a 400, so schema enforcement moves to strict tools or structured outputs. Thinking blocks are bound to their history: harnesses that edit earlier turns or rebuild system prompts get 400s on accounts created from 2026-08-31, and fallback models cannot read Fable 5.1 reasoning. Sampling parameters and prefill stay blocked. In exchange you get per-message effort without cache loss, turn-scoped system messages, readable progress updates, explicit refusal categories with server-side fallback, and a model partners describe as coherent across day-long runs. Parallel tool calls are less consistent; add the batching instruction Anthropic documents.
“The best finished documents and long answers Claude has produced, but on Pro it costs extra credits and at max effort you wait minutes, not seconds.”
In the Claude apps Fable 5.1 runs at medium effort by default, and the difference from Fable 5 shows in finished work: Canva preferred its writing in blind tests, Hebbia rated its decks and citation recall best of any model, and Glean's judges preferred its answers over Fable 5's about two to one. Chat replies carry fewer headers and bullets than older Claude models. Access depends on plan: Max and premium Team seats can spend up to half their weekly limits on it, Pro needs usage credits, and Free has none. It is slow at depth; Artificial Analysis measured about 67 tokens per second and roughly five minutes to the first answer token at max effort.
“Nearly every headline number is Anthropic's, on brand-new benchmark versions, and LMArena voters rank it below Fable 5. Trust your own evals.”
Three cautions. First, almost every performance figure is vendor-reported, on new benchmark versions (Terminal-Bench 4.0, Terminal-Bench-Science 0.1, OSWorld 2.0's August task release) with few outside baselines, and competitors publish their own versions of the same tables. Second, the independent signal is mixed: on LMArena's text board Fable 5.1 at max effort scores 1498 against Fable 5's 1506, and on the coding board 1519 against 1552, on far fewer votes. Third, safeguards still shape what you receive: cyber tasks can land on Opus 4.8 and life-sciences R&D on an Opus model. The agentic gains look real in named partner reports; confirm them on your workload.
Long-horizon engineering. Multi-hour or multi-day agent runs, codebase-wide migrations and root-cause debugging where one failed run costs more than the token premium; the cheaper cache reads matter most here.
Deep research and analysis. Multistep web research, document synthesis across the 1M window, and financial or scientific analysis where Hebbia, Samaya and Rakuten report gains in grounding and completeness.
Document, spreadsheet and slide production. Taking an analysis from a first question to a finished deck or live-formula spreadsheet, an area Anthropic singles out for improvement.
Escalation tier behind Opus 5. Route routine traffic to Opus 5 or Sonnet 5 and send only the tasks that fail there to Fable 5.1, lowering effort per message for routine steps.
The full research notes behind this review — verified against primary sources.
Anthropic discloses no architecture details: parameter counts, expert structure, layer counts and attention design are all unpublished, so those fields are null. What is disclosed: a 1M-token context window as both default and maximum, 128K max output tokens, and the tokenizer introduced with Claude Opus 4.7 (unchanged from Fable 5), which produces roughly 30% more tokens for the same text than pre-4.7 models. Thinking is adaptive and always on; depth is steered with the effort parameter, and the raw chain of thought is never returned. Anthropic's Transparency Hub says the model was pretrained on a proprietary mix of public web data, public and private datasets, user data and synthetic data, then post-trained toward the behavior described in Claude's constitution. Fable 5.1 and Mythos 5.1 share one set of weights; only the safeguards differ.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | Fable 5: 24.7% | Opus 5: 29.0%; GPT-5.6 Sol: 22.4% | Anthropic |
| Terminal-Bench 4.0 | 55.8% (Mythos 5.1: 60.9%) | Fable 5: 42.0% | Opus 5: 52.3%; GPT-5.6 Sol: 37.3% | Anthropic |
| GDPval-AA v2 (Elo) | 1853 | Fable 5: 1723 | Opus 5: 1824; GPT-5.6 Sol: 1711 | Anthropic |
| OSWorld 2.0 (partial / strict) | 77.9% / 41.7% | Fable 5: 72.9% / 36.1% | Opus 5: 75.4% / 39.6% | Anthropic |
| Humanity's Last Exam (no tools / with tools) | 60.9% / 65.0% | Fable 5: 57.8% / 63.8% | Opus 5: 56.6% / 63.6% | Anthropic |
| AutomationBench | 31.4% | Fable 5: 17.1% | Opus 5: 26.9%; GPT-5.6 Sol: 19.6% | Anthropic |
| CursorBench 3.2.0 (max effort) | 73.4% | Fable 5: 70.5% | Opus 5: 70.0%; GPT-5.6 Sol: 67.2% | Anthropic |
| LMArena Text Elo (style control, max effort) | 1498 (5th of 402, 5,783 votes) | Fable 5 (high): 1506 | Opus 5 (high): 1493; Gemini 3.1 Pro: 1487; GPT-6 Astra (max): 1480 | LMArena |
| LMArena Text Coding Elo (max effort) | 1519 (1,184 votes) | Fable 5 (high): 1552 | GPT-6 Astra (max): 1543; Opus 5 (high): 1533; Gemini 3.1 Pro: 1520 | LMArena |
Every row except LMArena is vendor-reported by Anthropic (2026-09-01) and was run with production safeguards on: tasks where the safeguards intervened scored zero on OSWorld 2.0, and other flagged tasks were completed by Opus 4.8 (cyber) or Opus 5 (biology). Terminal-Bench 4.0 and Terminal-Bench-Science are new versions with no comparable scores for older catalog models, so the compare table's Terminal-Bench column stays empty for this model rather than set a 4.0 score against 2.x scores. LMArena figures use the 2026-09-13 vote cutoff. xAI's Grok 4.7 launch table also scores Fable 5.1 at max effort: CursorBench 4.0 51.8%, Terminal-Bench 4.0 57.9%, DeepSWE v1.1 70.0% (competitor-run). Anthropic published no SWE-bench Verified, GPQA Diamond, MMLU-Pro or AIME score for this release.
Anthropic rates Fable 5.1's latency "Slower", against "Moderate" for Opus 5 and "Fast" for Sonnet 5. Artificial Analysis, measuring the Anthropic API at max effort with default fallback on a 10,000-input-token workload, reports a median 67.3 output tokens per second and about 291 seconds to the first answer token, most of it thinking time; a 500-token answer took about 299 seconds end to end. That figure describes max effort only. The API defaults to high effort and the Claude apps to medium, which think for less time, and no independent measurement at those settings was available, so time-to-first-token is left blank. Every's CEO reports it finished tasks about twice as fast as Opus 5 with half the tokens, a per-task claim rather than a per-token one. Plan batch or background execution for anything run at xhigh or max.
| Surface | Cost | Notes |
|---|---|---|
| API input | $10.00 / 1M tok | Flat across the full 1M window |
| API output | $50.00 / 1M tok | Effort level drives output spend |
| Cached input (read) | $0.25 / 1M tok | 0.025x base; Fable 5 is $1.00, Opus 5 is $0.50 |
| Cache write | $12.50 (5-minute) / $20.00 (1-hour) per 1M tok | Unchanged from Fable 5 |
| Batch (in/out) | $5.00 / $25.00 | 50% off |
| US-only inference | 1.1x | inference_geo: "us"; Bedrock and Google Cloud regional endpoints add 10% |
| Web search | $10 / 1,000 searches | Plus tokens |
| Direct UI | Max from $100/mo; Pro $20/mo plus usage credits | Max and premium Team seats: up to 50% of weekly limits; Enterprise: included |
| Free tier | None | Not on the Claude Free plan |
| Rate limits | Not listed on the model or pricing pages |
A refused request that stops before any output is not billed; when fallback runs, each attempt bills at the rates of the model that ran it. Artificial Analysis puts the blended price at $7.17 per 1M tokens on a 7:2:1 cache-read/input/output mix, which shows how much the cache cut matters for agent loops.
Proprietary and API-only: no weights, no self-hosting, no fine-tuning. Model IDs are claude-fable-5-1 on the Claude API, Claude Platform on AWS, Google Cloud and Microsoft Foundry, and anthropic.claude-fable-5-1 on Amazon Bedrock. Microsoft Foundry and Claude Platform on AWS run on Anthropic infrastructure and bill through the Azure and AWS marketplaces. Bedrock and Google Cloud offer global and regional endpoints, and the first-party API offers US-only inference at 1.1x. The 30-day retention requirement applies on every platform; on Bedrock and Google Cloud the retained data stays inside the cloud provider's environment. Claude Mythos 5.1, the same model with more permissive cyber and life-sciences safeguards, is invite-only through Project Glasswing, the Cyber Verification Program and the Life Sciences Verification Program, currently for US organizations, and also powers Claude Security. It is not a general procurement option.
Deployed under Anthropic's Responsible Scaling Policy with ASL-3 protections and treated as having CB-1 capabilities; Anthropic's evaluations put the underlying model short of the next biology risk tier and in the lower category of its Frontier Compliance Framework for cyber, with no critical-severity jailbreak found in external testing. Real-time classifiers cover cyber, biology, frontier-model development, reasoning extraction and general harms. It is a Covered Model: 30-day data retention is required, and zero data retention needs express authorization from Anthropic. Organizations with ZDR can enable 30-day retention for a single workspace. Anthropic's commercial terms bar it from training models on customer content. New anti-distillation controls stop new API accounts (created on or after 2026-08-31) from editing Claude's prior turns while keeping its earlier thinking. Anthropic reports lower reward hacking, less sycophancy and fewer hallucinated inputs than Mythos 5, but notes the model can still sometimes bypass approvals and auto-mode classifiers, and that its audit has thin coverage of very long-context and multi-agent work.
SDKs cover Python, TypeScript, Java, Go, Ruby, C# and PHP, plus the ant CLI. Claude Code runs Fable 5.1 at high effort by default, and the Claude apps run it at medium. It works with the Claude Agent SDK and Claude Managed Agents, which keep the conversation prefix intact for its history-bound thinking. Launch partners include Cognition, which moved its Opus 5 traffic in Devin to Fable 5.1 on day one, and Glean, which now recommends it wherever Fable 5 was the choice; Jane Street, Millennium, Datadog, Hebbia, Canva and Shopify reported evaluation results. Availability on Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry from day one makes it reachable inside every major enterprise cloud.
Anthropic recommends starting with Opus 5 and moving to Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5 at higher effort still falls short. Fable 5.1 costs twice as much per fresh token but reads cache at half the Opus 5 price, so the gap narrows on long, cache-heavy agent runs.
They are the same model. Fable 5.1 is generally available with safety classifiers; Mythos 5.1 has more permissive cyber and life-sciences safeguards and is invite-only through Project Glasswing and Anthropic's trusted-access programs, currently for US organizations.
On the API the response comes back with stop_reason: "refusal" and a category such as cyber or bio. With server-side fallback (beta) the request is retried on Opus 4.8 or Opus 5 and the response names the model that answered. A refusal before any output is not billed; fallback attempts bill at the answering model's rates.
Not by default: it is a Covered Model that requires 30-day retention, and ZDR needs express authorization from Anthropic. Anthropic says eligible customers can use it with ZDR until Enterprise Frontier Safeguards, which store data in the customer's own cloud, roll out in phases from fall 2026. Organizations with ZDR can also enable 30-day retention for a single workspace.
List input and output prices are identical at $10 / $50. Cache reads fall from $1.00 to $0.25 per 1M tokens, which Anthropic measured as about 25% lower cost on typical workloads and up to 45% on highly agentic ones.
Forced tool_choice (any or tool) returns a 400, earlier models cannot read its thinking blocks, and editing earlier turns invalidates later thinking blocks, returning a 400 on accounts created from 2026-08-31. Parallel tool calls are less frequent by default. The tokenizer, context window and refusal handling are unchanged.
Yes on paid plans: Max and premium Team seats can use it for up to 50% of weekly limits, Enterprise includes it, and Pro requires usage credits. It is not available on the Free plan. The apps run it at medium effort; Claude Code runs it at high.
The Claude API, Amazon Bedrock (anthropic.claude-fable-5-1), Claude Platform on AWS, Google Cloud and Microsoft Foundry, all from launch day.
The closest rival: same $10 / $50 list price, 128K max output and text-and-image input, released two days later. Fable 5.1 wins on cache reads ($0.25 vs $1.00), flat pricing across its 1M window (Astra bills prompts over 272K tokens at 2x input and 1.5x output), a later knowledge cutoff (June vs April 2026) and LMArena text (1498 vs 1480). Astra leads on LMArena coding (1543 vs 1519, on only 645 votes), and Anthropic's published comparisons do not include it yet.
Half the rate card ($5 / $25), faster, available under zero data retention, and Anthropic's own recommended default. Fable 5.1 leads it on Terminal-Bench 4.0 (55.8% vs 52.3%), Terminal-Bench-Science (52.6% vs 29.0%), AutomationBench (31.4% vs 26.9%) and GDPval-AA v2 (1853 vs 1824), and its cache reads cost half as much ($0.25 vs $0.50). Opus 5 scores higher on LMArena coding (1533 vs 1519).
Far cheaper at $2 / $12, with audio and video input, Google Search grounding and a 2M window on Vertex AI, but still in preview with a January 2025 knowledge cutoff. The two are level on LMArena coding (1520 vs 1519) and close on text (1487 vs 1498); Fable 5.1 is the stronger pick for long, unattended agent work.
$2 / $6 with a 500K window and native X search, released 2026-09-21. On xAI's own launch table Fable 5.1 at max effort leads on CursorBench 4.0 (51.8% vs 46.3%), Terminal-Bench 4.0 (57.9% vs 38.0%) and GDPval (1735 vs 1695), while Grok 4.7 leads on EEBench (64.0% vs 56.4%) and Harvey's legal agent benchmark (19.6% vs 6.7%) at a fraction of the price.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21