Anthropic
Claude Fable 5.1
Eight of ten benchmark scores tracked for Claude Fable 5.1 are independent runs — two exceptions being Anthropic's own claimed OSWorld 2.0 numbers — but the consensus is thinner than it looks: Terminal-Bench 2.1 has three different numbers for the identical model (91.4% from Artificial Analysis, 85.02% from vals.ai's own harness, none from the canonical board tbench.ai), and vals.ai's disclosed refusal-to-Opus fallback would cut its figure to 79.03%.
Claude Fable 5.1 benchmarks and pricing, every number sourced: 10 tracked Claude Fable 5.1 benchmark scores (8 independently run, 2 still resting on a vendor’s own claim), priced at $10.00 per million input tokens and $50.00 per million output.
- Released
- 2026-09-01
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- 2026-06
- Verified
- sources checked 2026-09-02–2026-10-08
- Succeeded
- Claude Fable 5 (2026-06-09)
- Parameters
- Not disclosed
- Architecture
- Not disclosed
Claude Fable 5.1’s verified record
Claude Fable 5.1’s most-compared rival is Kimi K3: 5 leads and 3 not callable across their 8 shared comparisons. Claude Fable 5.1 is priced at $10.00/$50.00 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.
Against the 204 head-to-head comparisons Claude Fable 5.1 shares with other tracked models:
A gap counts for Claude Fable 5.1 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Claude Fable 5.1 trails on 1 of them.
HLE · no tools Reasoning
±2 is noiseLiveBench Composite score across 7 domains
±2.7 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseTerminal-Bench 4.0 · max Terminal ops
±12.4 is noiseARC-AGI-2 · max Compositional visual reasoning
±9.2 is noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseNo verdict for Claude Fable 5.1 anywhere on GPQA Diamond, LiveCodeBench ( saturated); OSWorld 2.0 · v2 1 full, OSWorld 2.0 · v2026 08 08 full modified ( nothing independently confirmed on both sides).
- None of Claude Fable 5.1’s coding comparisons are independently confirmed on both sides yet.
What changed from Claude Fable 5 to Claude Fable 5.1
The 9 benchmarks both models have been scored on, using the same variant each time. A raw Claude Fable 5.1 gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.
| Benchmark | Claude Fable 5 | Claude Fable 5.1 | Change | Verdict |
|---|---|---|---|---|
| HLE | 55.5 | 59.1 | +3.6 | Real gap |
| Terminal-Bench 2.1 | 83.8 | 91.4 | +7.6 | Setup-dependent |
| GPQA Diamond | 93.18 | 93.43 | +0.3 | Tainted |
| LiveCodeBench | 89.78 | 90.52 | +0.7 | Tainted |
| ARC-AGI-2 | 89.2 | 90 | +0.8 | Tie |
| LiveBench | 83 | 83.4 | +0.4 | Tie |
| AnalystAgent | 48.75 | 57.5 | +8.8 | Tie |
| Terminal-Bench 4.0 | 41.41 | 58.08 | +16.7 | Real gap |
| OSWorld 2.0 | 72.9 | 77.9 | +5.0 | Unverified |
Claude Fable 5.1 API pricing
$10.00 in / $50.00 out per 1M tokens — official pricing source
What Claude Fable 5.1 costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $1.50 |
| A codebase review | 1,000K / 100K | $15.00 |
| A day of agent work | 10,000K / 1,000K | $150.00 |
Computed from Claude Fable 5.1’s list rates above — cache discounts and batch tiers are not applied.
Claude Fable 5.1 is one of 8 Anthropic models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| 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 | $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 5.1 benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| Terminal-Bench 2.1[2] Terminal ops · ±10.6 is noise | |
| GPQA Diamondsaturated[3] Expert science Q&A — not ranked at any gap size | |
| LiveCodeBenchsaturated[4] Contest coding — not ranked at any gap size | |
| ARC-AGI-2(max)[5] Compositional visual reasoning · ±9.2 is noise | |
| LiveBench[6] Composite score across 7 domains · ±2.7 is noise | |
| AnalystAgent[7] Spreadsheet & document analysis · ±11.2 is noise | |
| Terminal-Bench 4.0(max)[8] Terminal ops · ±12.4 is noise | |
| OSWorld 2.0(v2026 08_08_full_modified)[9] Long-horizon computer use · ±9.7 is noise | |
| OSWorld 2.0(v2 1_full)[10] Long-horizon computer use · ±9.7 is noise |
Who ran these numbers: 8 of 10 independent — artificialanalysis.ai (3), vals.ai (3), arcprize.org (1), livebench.ai (1); vendor self-reported (2).
- HLE: AA live leaderboard value, row labeled "Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback)" — the board's top-ranked row for this model. AA's label doesn't name which model the fallback substitutes (Fable 5's own AA row is labeled "Opus 4.8 Fallback" by contrast); Anthropic names both Opus 5 and Opus 4.8 as fallback options without saying which applies when. Two lower AA effort tiers for the same model, Xhigh (58.7%) and High (55.9%), are on the same page but not recorded as separate rows, matching this site's one-row-per-model convention.
- Terminal-Bench 2.1: AA Max Effort row, rank 1 of 28 shown, Terminus 2 harness in an e2b sandbox. This benchmark shows three disagreeing numbers for this model — see "Notes on the record" on this page for the full comparison and why Artificial Analysis is the tracked figure. Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists it at 85.02, with fallbacks on 27 of 267 tasks; counting those as failures gives 79.03.
- GPQA Diamond: vals.ai board, rank 8 of 137 systems, updated 2026-09-01. GPQA Diamond is graded saturated on this site, so this number never drives a verdict. Artificial Analysis separately reports 93.7% for the same model on its own harness — not recorded as a second row; both figures round to the same near-ceiling conclusion regardless of which is used.
- LiveCodeBench: vals.ai board, rank 1 of 142 systems, updated 2026-09-01.
- ARC-AGI-2: Top of five populated effort tiers (Max 90.0%, XHigh 90.0%, High 88.8%, Medium 86.3%, Low 78.3%); Max and XHigh tie on ARC-AGI-2, but ARC Prize's own listed inference cost is higher for Max ($4.49/task vs $3.12/task) — this is ARC Prize's compute cost to run the eval, not Anthropic's API price. Dated 2026-09-01 on arcprize.org. Companion score on the benchmark's earlier, easier predecessor, ARC-AGI-1 (not independently tracked on this site): 97.5%.
- LiveBench: Rank 1 of the entire board on the LiveBench-2026-06-25 release, Max Effort row. Category sub-scores: Reasoning 91.7, Coding 86.4, Agentic Coding 66.1, Mathematics 97.0, Data Analysis 80.3, Language 89.5, Instruction Following 73.0. LiveBench's own listed inference cost: $1.212 per successful task (its compute cost to run the eval, not Anthropic's API price).
- AnalystAgent: AA's own run, board row 'Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback)'. pass@1 65.75, pass@5 72.5.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 58.081 at effort max. Artificial Analysis same-conditions run (Max, Default Fallback): 52.02. Official board (grant-funded, Claude Code, max effort, 5-trial pass@1): 57.88 ± 3.76 CI — 0.2 points below this run. The official xhigh row ties max at 57.88; Artificial Analysis's own max row (52.02) sits below its xhigh (55.05).
- OSWorld 2.0: Anthropic's own evaluation (model launched 2026-09-01; system card §8.14.3): August 2026 task release plus Anthropic's own task and grading fixes, Claude Opus 4.8 as model grader, 1080p, 500 steps, pass@1 averaged over five runs. Strict pass rate 41.7. Anthropic evaluated with production safeguards enabled and scored a zero on tasks where they intervened. The card states these numbers supersede the Opus 5 system card figures and are not directly comparable to other releases; OpenAI's Astra footnote and Google's methodology both explicitly exclude this modified protocol. Verified on Anthropic's launch grid and in the Fable 5.1 system card text.
- OSWorld 2.0: Anthropic's rerun of Fable 5.1 under the same September 10 task files and modified harness as its Opus 5.5 row (2026-09-22 launch chart and system card §8.13.3, re-evaluated under identical conditions). Partial reward 80.7, strict pass rate 42.8. Note the 2.8-point rise from the same model's August-plus-fixes figure (77.9) purely from moving to the newer task files and harness. Verified on the launch page grid and in the system card text.
Notes on the record
Launched 2026-09-01 (Anthropic announcement). vals.ai's 2026-08-28 reads as an internal API cutover; the public date is used, as for Fable 5. Anthropic's models table calls it "Successor to Claude Fable 5" — the relationship this site's previous-version rule also reads. Pricing unchanged at $10 in / $50 out per 1M tokens — the highest Claude price tracked, tied with Fable 5.
Claude Mythos 5.1 is the identical model with different safety guardrails, limited to vetted organizations (Project Glasswing, invitation-only). Context window 1M in; max output 128K. Reliable knowledge cutoff moves from January 2026 to June 2026, one month past Claude Opus 5's (platform.claude.com models overview, checked 2026-09-02). Cached reads drop 75%, $1 to $0.25 per MTok; Anthropic estimates 25-45% savings depending on workload.
The provenance caveat from Fable 5 carries over: safety classifiers can reroute a refused prompt to Claude Opus 5 or Opus 4.8 as an automatic fallback, and vals.ai discloses running several Fable 5.1 evaluations with that fallback on. Where it publishes the correction, its Terminal-Bench 2.1 score falls from 85.02% to 79.03% once fallback-assisted answers count as failures. Anthropic claims ~60% fewer cybersecurity refusal false positives than Fable 5 and ~85% fewer on elementary biology questions (a safeguard update applied to both models); neither settles how often the fallback fires.
Terminal-Bench 2.1 has three pictures of the same model: Artificial Analysis (Terminus 2 harness in an e2b sandbox) 91.4%; vals.ai's own harness 85.02% (79.03% with the fallback correction); tbench.ai, the canonical board, no Fable 5.1 row at all. The AA figure is recorded, per this benchmark's source order (tbench.ai, then AA). GPQA Diamond also has two independent numbers (93.43% vals.ai, 93.7% AA), too close to matter.
All seven tracked scores — HLE, Terminal-Bench 2.1, GPQA Diamond, LiveCodeBench, ARC-AGI-2, LiveBench, AA-AnalystAgent — are independent (Artificial Analysis, vals.ai, ARC Prize, LiveBench). DeepSWE, Toolathlon-Verified, Agents' Last Exam, SWE-bench Verified and HMMT had no Fable 5.1 row as of 2026-09-02.
Compare with
FAQ
Has Claude Fable 5.1 been independently benchmarked?
Yes, on eight of the ten benchmarks this site has scored it against so far: Humanity's Last Exam, Terminal-Bench 2.1, GPQA Diamond, LiveCodeBench, ARC-AGI-2, LiveBench, and AA-AnalystAgent come from independent evaluators (Artificial Analysis, vals.ai, ARC Prize, and LiveBench's own board), and Terminal-Bench 4.0 joined them on 2026-10-08 (58.08, vals.ai). The other two rows are Anthropic's own claimed OSWorld 2.0 numbers (80.7 on the v2.1 slice, 77.9 on Anthropic's modified slice). Five other boards this site tracks — DeepSWE, Toolathlon-Verified, Agents' Last Exam, SWE-bench Verified, and HMMT Feb 2026 — had not added the model as of 2026-09-02, one day after its public launch.
Does Claude Fable 5.1 still substitute a different model's answer when it refuses a prompt?
Yes — the mechanism from Claude Fable 5 carries over: when Fable 5.1's safety classifiers refuse a prompt, Claude Opus 5 or Opus 4.8 answers instead, and that substituted answer can still count toward a published benchmark score. vals.ai discloses using the fallback for several of its Fable 5.1 evaluations; where it publishes the correction, its Terminal-Bench 2.1 score falls from 85.02% to 79.03% once fallback-assisted answers are recounted as failures. Anthropic claims roughly 60% fewer cybersecurity refusal false positives than Fable 5, and a biology-safeguard update applied to both models cuts false positives on elementary biology and medical questions by roughly 85% — both reduce how often the fallback fires without eliminating it.
How does Claude Fable 5.1 pricing compare to Claude Fable 5?
The headline rate is identical: $10 per million input tokens and $50 per million output tokens, the same price Fable 5 has carried since its June 2026 launch, and the two of them remain tied for the highest of the Claude models tracked on this site. The one change is on cached reads, which drop 75% — from $1/MTok under Fable 5 to $0.25/MTok under Fable 5.1. Batch API pricing (50% off standard rates) is unchanged (source: platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1).
Is Claude Fable 5.1 included with a Claude subscription plan?
For chat use, yes: Anthropic lists Claude Fable 5.1 as available to Pro, Max, Team, and Enterprise subscribers (the Free plan is not included). Anthropic's pages did not state specific per-plan usage caps for Fable 5.1 at launch, pointing users to its general pricing page instead — Claude Fable 5's own record on this site shows those specifics can arrive and change weeks after a model's launch, so treat any cap figure elsewhere as provisional until Anthropic publishes one for 5.1 specifically (source: anthropic.com/claude/fable, checked 2026-09-02).
Does Claude Fable 5.1 replace Claude Fable 5, or sit alongside it?
It replaces it. Anthropic's own models comparison table describes Claude Fable 5.1 as the "successor to Claude Fable 5," the same relationship this site uses to mark a previous version (as with Claude Opus 4.8 → Claude Opus 5 and GLM-5.2 → GLM-5.3) — Fable 5 stays on this site for comparison, not deleted, but is no longer the current recommendation. As with the earlier model, a less-restricted sibling exists: Claude Mythos 5.1 shares the same underlying weights and pricing but is limited to vetted organizations under Anthropic's invitation-only Project Glasswing. (For the full per-benchmark comparison against Fable 5, see the "What changed" table above.)
What is Claude Fable 5.1's knowledge cutoff?
June 2026, per Anthropic's disclosure. Like the rest of the current Claude generation it serves the full 1,000,000-token window at one flat rate, so the cutoff — not the window — is the spec that moved.
Further reading
- GPQA Diamond leaderboard 2026 — Claude Fable 5.1 is one of the 23 models it compares.
- Models with 10M token context windows 2026 — Claude Fable 5.1 is one of the 40 models it compares.
Benchmark guides
- Humanity's Last Exam — what Claude Fable 5.1’s reasoning score on it does and doesn’t prove.