AnthropicPrevious version · Claude Fable 5.1

Claude Fable 5

Eleven of twelve benchmark scores tracked for Claude Fable 5 come from independent evaluators rather than Anthropic's own runs (the exception: an OSWorld 2.0 figure from Anthropic's launch materials) — but that doesn't settle what they measured: Fable 5's safety classifiers can silently reroute a refused prompt to Claude Opus 4.8, and Anthropic has acknowledged some published Fable 5 scores already include those substituted answers, without disclosing which.

Claude Fable 5 benchmarks and pricing, every number sourced: 12 tracked Claude Fable 5 benchmark scores (11 independently run, 1 still resting on a vendor’s own claim), priced at $10.00 per million input tokens and $50.00 per million output.

Released
2026-06-09
License
proprietary
Context window
1M tokens
Knowledge cutoff
2026-01
Verified
sources checked 2026-08-17–2026-10-08
Parameters
Not disclosed
Architecture
Not disclosed

Claude Fable 5’s verified record

Claude Fable 5’s most-compared rival is Kimi K3: 4 leads, 2 ties, and 5 not callable across their 11 shared comparisons. Claude Fable 5 is priced at $10.00/$50.00 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.

Against the 290 head-to-head comparisons Claude Fable 5 shares with other tracked models:

79 real gaps46 inside the band165 not called

A gap counts for Claude Fable 5 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Claude Fable 5 trails on 11 of them.

HLE · no tools Reasoning

±2 is noise
Behind
Claude Opus 5.5 −5.9 · 1 previous: Claude Fable 5.1 −3.6
Ahead
Kimi K3.1 Pro +3.0 · Muse Wonder +5.4 · MiMo-V2.6-Pro +6.1 · Muse Spark 1.3 +6.8 · MiniMax M3.1 Flash +7.5 · Gemini 3.1 Pro Preview +8.5 · Kimi K3 +8.6 · Step 5 Preview +9.0 · Claude Haiku 5.5 +11.1 · DeepSeek V4.1 Flash +12.2 · GLM-5.3 +13.2 · DeepSeek V4 Pro (0813) +14.5 · GLM-5.3-Flash +15.6 · MiniMax M3 +16.5 · GPT-6 Luna +17.0 · Qwen3.8-Flash-Next +17.5 · Mistral Large 4 +20.5 · Beam 1.0 +20.8 · Nemotron 3.5 Super VL +24.8 · 9 previous: GPT-6.1 Sol +2.6 · GPT-5.6 Sol +6.0 · Claude Opus 4.8 +6.8 · GPT-6 Sol +7.6 · Qwen3.8-Max +12.4 · Claude Sonnet 5 +14.2 · GLM-5.2 +14.4 · GPT-5.6 Luna +16.0 · DeepSeek V4 Flash (0731) +16.9
Tie
3 models within ±2
Unverified
1 model — vendor-reported on one side
Setup-dependent
7 models — scored on a different harness or effort tier

LiveBench Composite score across 7 domains

±2.7 is noise
Ahead
Kimi K3 +3.8 · DeepSeek V4 Pro (0813) +5.6 · Qwen3.8-Flash-Next +6.8 · GLM-5.3 +6.9 · GPT-6 Luna +11.0 · GLM-5.3-Flash +11.4 · MiniMax M3 +15.7 · 8 previous: Claude Opus 5 +2.9 · GPT-6 Sol +3.7 · Qwen3.8-Max +4.5 · Grok 4.6 +5.0 · Claude Opus 4.8 +6.8 · DeepSeek V4 Flash (0731) +8.8 · GPT-5.6 Luna +9.4 · GLM-5.2 +9.8
Tie
5 models within ±2.7
Setup-dependent
9 models — scored on a different harness or effort tier

Terminal-Bench 4.0 · max Terminal ops

±12.4 is noise
Behind
Claude Sonnet 5.5 −22.7 · Claude Opus 5.5 −23.7 · 3 previous: GPT-6.1 Sol −13.6 · Claude Fable 5.1 −16.7 · GPT-6 Astra −18.2
Ahead
Muse Spark 1.3 +16.7 · Kimi K3 +24.2 · DeepSeek V4 Pro (0813) +27.3 · GPT-6 Luna +27.8 · 3 previous: Claude Opus 4.8 +18.2 · GPT-5.6 Luna +29.8 · Claude Sonnet 5 +31.8
Tie
5 models within ±12.4
Unverified
1 model — vendor-reported on one side
Setup-dependent
1 model — scored on a different harness or effort tier

AnalystAgent Spreadsheet & document analysis

±11.2 is noise
Ahead
Muse Spark 1.3 +11.3 · MiMo-V2.6-Pro +11.3 · Step 5 Preview +13.8 · DeepSeek V4.1 Flash +13.8 · GLM-5.3-Flash +23.8 · MiniMax M3 +38.8
Tie
16 models within ±11.2
Setup-dependent
5 models — scored on a different harness or effort tier

DeepSWE Long-horizon coding

±9.5 is noise
Ahead
6 previous: Claude Opus 4.8 +11.0 · Qwen3.8-Max +13.0 · Muse Spark 1.2 +15.0 · Claude Sonnet 5 +16.0 · DeepSeek V4 Flash (0731) +17.0 · GLM-5.2 +26.0
Tie
10 models within ±9.5
Unverified
19 models — vendor-reported on one side
Setup-dependent
2 models — scored on a different harness or effort tier

ARC-AGI-2 · max Compositional visual reasoning

±9.2 is noise
Ahead
GLM-5.3-Flash +23.4 · DeepSeek V4 Pro (0813) +27.9 · Kimi K3 +28.8 · GPT-6 Luna +29.9 · 2 previous: DeepSeek V4 Flash (0731) +27.8 · GPT-5.6 Luna +29.6
Tie
6 models within ±9.2

Agents' Last Exam Professional work

±3.2 is noise
Behind
Muse Spark 1.3 −6.5 · 3 previous: GPT-5.6 Luna −4.6 · GPT-5.6 Sol −4.9 · GPT-6 Sol −6.5
Tie
1 model within ±3.2
Unverified
7 models — vendor-reported on one side
Setup-dependent
9 models — scored on a different harness or effort tier

No verdict for Claude Fable 5 anywhere on GPQA Diamond, LiveCodeBench, SWE-bench Verified ( saturated); OSWorld 2.0 · v2026 08 08 full modified, Terminal-Bench 2.1 ( nothing independently confirmed on both sides).

Claude Fable 5 API pricing

$10.00 in / $50.00 out per 1M tokens — official pricing source

What Claude Fable 5 costs per job

Claude Fable 5 cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$1.50
A codebase review1,000K / 100K$15.00
A day of agent work10,000K / 1,000K$150.00

Computed from Claude Fable 5’s list rates above — cache discounts and batch tiers are not applied.

Claude Fable 5 is one of 9 Anthropic models tracked on this site, at these official list prices.

Anthropic model pricing, official list rates
ModelIn / 1MOut / 1M
Claude Fable 5.2$8.00$40.00
Claude Haiku 5.5$0.10$0.50
Claude Sonnet 5.5$2.00$10.00
Claude Opus 5.5$4.00$20.00
Claude Fable 5.1(previous version)$10.00$50.00
Claude Opus 5(previous version)$5.00$25.00
Claude Sonnet 5(previous version)$2.00$10.00
Claude Fable 5(previous version)$10.00$50.00
Claude Opus 4.8(previous version)$5.00$25.00

Claude Fable 5 benchmark scores

Claude Fable 5 benchmark scores, provenance, and source links
BenchmarkScore
HLE(no tools)[1]
Reasoning · ±2 is noise
Terminal-Bench 2.1[2]
Terminal ops · ±10.6 is noise
DeepSWE[3]
Long-horizon coding · ±9.5 is noise
GPQA Diamondsaturated[4]
Expert science Q&A — not ranked at any gap size
Agents' Last Exam[5]
Professional work · ±3.2 is noise
LiveCodeBenchsaturated[6]
Contest coding — not ranked at any gap size
SWE-bench Verifiedsaturated[7]
Bug fixing — not ranked at any gap size
ARC-AGI-2(max)[8]
Compositional visual reasoning · ±9.2 is noise
LiveBench[9]
Composite score across 7 domains · ±2.7 is noise
AnalystAgent[10]
Spreadsheet & document analysis · ±11.2 is noise
Terminal-Bench 4.0(max)[11]
Terminal ops · ±12.4 is noise
OSWorld 2.0(v2026 08_08_full_modified)[12]
Long-horizon computer use · ±9.7 is noise

Who ran these numbers: 11 of 12 independent — vals.ai (4), artificialanalysis.ai (2), tbench.ai (1), deepswe.datacurve.ai (1), snorkel.ai (1), arcprize.org (1), livebench.ai (1); vendor self-reported (1).

  1. HLE: AA live leaderboard value, 'Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)' — the shipped system: classifier-refused tasks fall back to Opus 4.8, ~9%, included in the score. Board rank 1 as of 2026-08-19. Corrected 2026-08-19: this row used to cite AA's Opus 5 launch article, which reports 53.0 — the 55.5 was always the board's number, so the citation pointed at a page that did not support it.
  2. Terminal-Bench 2.1: Fable 5's Terminal-Bench 2.1 row was re-verified 2026-09-03 in a browser after npm run check-sources flagged it: tbench.ai rebuilt the site around 2026-08-27 (new default board Terminal-Bench 4.0, version now a client-side selection on a single URL), so the old deep link no longer serves 2.1 in static HTML — see the claude-opus-4-8 row on this same benchmark for the full mechanism. Value confirmed unchanged: 83.8%, rank 1 of 17, 'Fable 5 (xhigh)' via Claude Code (CI now shown as ±2.3%, not ±1.2 — recomputed, not a re-run). Anthropic's own launch table claims 88.0. Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists it at 80.52, with Opus 4.8 as a refusal fallback on 44 of 267 tasks.
  3. DeepSWE: 70%±4 pass rate, mini-swe-agent harness; board updated 2026-08-13. Costs $21.63/task vs Opus 5's $11.84.
  4. GPQA Diamond: Major caveat, per vals.ai itself: 93.18% counts refusal-triggered fallbacks as successes — counting refusals as failures drops it to 55.56%. Read with the benchmark's saturation grade in mind.
  5. Agents' Last Exam: Overall pass rate 25.7 (Claude Code, XHigh; partial-credit score 48.7; $4,340 eval cost). Snorkel flags the served variant may understate the full capability tier. Board as_of 2026-08-14. Per-effort results on 2026-10-01: XHigh 25.7, Adaptive 22.0.
  6. LiveCodeBench: vals.ai top performer, updated 2026-08-15.
  7. SWE-bench Verified: vals.ai run, rank 7/86 shown, Mini-SWE-agent harness (board updated 2026-08-19). SWE-bench Verified is graded saturated on this site, so this row completes Claude Fable 5's tracked record here without settling any head-to-head comparison.
  8. ARC-AGI-2: Claude Fable 5's official ARC-AGI-2 leaderboard row, dated 2026-06-09 on arcprize.org. The top of five populated tiers (Max/XHigh/High/Medium/Low).
  9. LiveBench: Board row "Claude Fable 5 Max Effort" on the 2026-06-25 LiveBench release.
  10. AnalystAgent: AA's own run, board row 'Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)'; the board rounds to one decimal. pass@1 59.75, pass@5 68.75.
  11. Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 41.414 at effort max. Official board (grant-funded, Claude Code, max effort, 5-trial pass@1): 44.55 ± 3.85 CI — 3.1 points above this run.
  12. OSWorld 2.0: Anthropic's rerun of Fable 5 under the same modified protocol as its Fable 5.1 row (August release plus Anthropic's fixes, Opus 4.8 grader, five-run pass@1). Partial reward 72.9, strict pass rate 36.1. Safeguards-zeroing caveat applies here too: tasks where safeguards intervened were scored zero. Not comparable to any official-board row. Verified on Anthropic's launch grid and in the Fable 5 system card text.

Notes on the record

First "Mythos-class" model made generally available — a new tier above Opus, not a successor to Opus 4.8 (Anthropic kept developing and selling Opus 4.8 alongside it). A less-restricted sibling, Claude Mythos 5, shares Fable 5's specs and $10/$50 pricing but is invitation-only, reserved for vetted cybersecurity customers under Anthropic's Project Glasswing (source: platform.claude.com/docs/en/about-claude/models/overview).

Always-on adaptive thinking; safety classifiers can refuse requests (stop_reason "refusal"), in which case Claude Opus 4.8 serves as automatic fallback — some published benchmark scores include those fallback answers.

Suspended under US export controls 2026-06-12 to 06-30, restored to all users 2026-07-01 — see the FAQ below for the cause and the fix. Mythos 5, gated to vetted Project Glasswing organizations, had already returned earlier and more narrowly: roughly 100 approved US organizations from 2026-06-26 (sources: cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html; anthropic.com/news/redeploying-fable-5).

Subscription access has also changed since launch: an initial free window for Claude Pro/Max/Team/Enterprise users (2026-06-09 through 2026-06-22) was cut short by the export-control suspension; as of 2026-07-20, Claude Max and Team Premium plans include Fable 5 at no extra token charge up to 50% of weekly usage limits, while Claude Pro and Team Standard plans run Fable 5 entirely on pay-as-you-go usage credits at standard API rates, with eligible Pro/Team Standard users receiving a one-time promotional credit when the change took effect (source: support.claude.com/en/articles/15424964-claude-fable-5-on-your-plan).

Core API pricing ($10/$50 per MTok, 1M context, Jan 2026 knowledge cutoff) is unchanged and independently confirmed against Anthropic's current pricing and models-overview pages as of 2026-08-20. That $10/$50 price ties Claude Fable 5.1 for the highest of the nine Claude models tracked on this site, but not Anthropic's highest on the Claude API: legacy Opus 4.1/Opus 4 (still on Bedrock and Google Cloud) run $15/$75, and Opus 5/4.8 Fast Mode is also $10/$50.

A previous version since Claude Fable 5.1 shipped on 2026-09-01 at an unchanged $10/$50 — which Anthropic's own models table agrees with ("Successor to Claude Fable 5"). Fable 5 keeps an independent record — 11 of 12 tracked scores remain independent runs (one OSWorld 2.0 exception, Anthropic's launch materials) — while Fable 5.1 has picked up 8 independent scores of its own so far, with 5 of the boards this site tracks (DeepSWE, Toolathlon-Verified, Agents' Last Exam, SWE-bench Verified, HMMT Feb 2026) not yet showing a Fable 5.1 row as of 2026-09-02.

Compare with

FAQ

Has Claude Fable 5 been independently benchmarked?

Eleven of the twelve scores tracked for Claude Fable 5 on this page — HLE, Terminal-Bench 2.1, DeepSWE, GPQA Diamond, Agents' Last Exam, LiveCodeBench, SWE-bench Verified, ARC-AGI-2, LiveBench, AA-AnalystAgent, and Terminal-Bench 4.0 — are independent runs, drawn from seven evaluators: Artificial Analysis, tbench.ai, DeepSWE's own leaderboard, vals.ai, Snorkel AI, ARC Prize, and LiveBench's own board. That's 11 of 12, with the one self-reported exception — the OSWorld 2.0 figure — noted above. One caveat applies to any Fable 5 benchmark, including these: the model's safety classifiers can silently reroute a refused prompt to Claude Opus 4.8, and Anthropic has said some published Fable 5 scores already include those substituted answers without identifying which ones.

Why was Claude Fable 5 unavailable for part of June 2026?

On 2026-06-12 the US Commerce Department ordered Anthropic to suspend access to Claude Fable 5 and Claude Mythos 5 for non-US nationals. Amazon researchers had found a jailbreak that bypassed Fable 5's safety classifier well enough to get it to identify — and in one case write exploit code for — a known software vulnerability. Anthropic had no way to verify user nationality in real time, so it suspended both models for everyone rather than just foreign nationals. After retraining the safety classifier to block the specific technique (Anthropic reports over 99% of cases) and a review by Commerce's Center for AI Standards and Innovation, the order was lifted on 2026-06-30 and Fable 5 returned for all users on 2026-07-01 (sources: cnbc.com, 2026-06-30; anthropic.com/news/redeploying-fable-5).

How does Claude Fable 5 pricing compare to Claude Opus 5?

Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. That's exactly double Claude Opus 5's $5/$25 and, tied with its successor Claude Fable 5.1, the highest per-token price of the nine Claude models tracked on this site (Opus 4.8 and Opus 5 at $5/$25, Opus 5.5 at $4/$20, Sonnet 5 and Sonnet 5.5 at $2/$10). It is not the highest price Anthropic charges anywhere on the Claude API, though: legacy Claude Opus 4.1 and Opus 4 (still billable on Bedrock and Google Cloud) run $15/$75, and Anthropic's Fast Mode add-on for Opus 5/4.8 is also $10/$50 (source: platform.claude.com/docs/en/about-claude/pricing). Claude Mythos 5 matches Fable 5's $10/$50 price but is invitation-only, not generally available. The same discounts apply as elsewhere on the API: a 90% price cut on prompt-cache hits and a 50% cut on Batch API requests.

Is Claude Fable 5 included free with a Claude subscription?

It depends on the plan, and the terms changed several times before settling on 2026-07-20. Claude Max and Team Premium subscribers get Claude Fable 5 bundled into their plan at no extra token charge, up to 50% of their weekly usage limit; past that they need usage credits. Claude Pro and Team Standard subscribers get no bundled access at all — Fable 5 there runs entirely on pay-as-you-go usage credits at the standard API rate, though Anthropic gave eligible Pro/Team Standard users a one-time promotional credit when the change took effect (source: support.claude.com, 'Claude Fable 5 on your plan').

What is the 'Mythos-class' tier that Claude Fable 5 belongs to, and is it the successor to Opus?

'Mythos-class' is a new capability tier Anthropic created above Opus — Claude Fable 5 is not a successor to Opus 4.8, and Anthropic kept developing and selling Opus 4.8 alongside it. Claude Fable 5, launched 2026-06-09, is the first Mythos-class model released to the general public. A less-restricted sibling, Claude Mythos 5, is limited to vetted cybersecurity customers under Anthropic's invitation-only Project Glasswing and shares Fable 5's specs and pricing (source: platform.claude.com/docs/en/about-claude/models/overview).

What is Claude Fable 5's knowledge cutoff?

January 2026, disclosed by Anthropic at month precision rather than an exact date. The context window runs the full 1,000,000 tokens at a single rate — no long-context price tier on this model.

Further reading

Benchmark guides