Anthropic: Claude Fable 5
Anthropic: Claude Fable 5 is anthropic's model. On the un-leaked meo benchmark it ranks #6 of 82 with 88.7% accuracy and an Effective Value (π) of 1.1643 β earning a correct answer for about $0.0573 in 14s each.
as of 2026-10-18
Analysis: why Anthropic: Claude Fable 5 ranks #6
Analysis by Jason Davis Β· updated 2026-06-09The Effective-Value champion: #3 on raw accuracy, but the #1 model on the board for real-world intelligence-per-cost-and-speed.
As of June 2026, Anthropic's Claude Fable 5 ranks #3 of 75 fully-scored models on the un-leaked meo benchmark, with 88.7% micro accuracy (92.0% macro) and a 100% answer rate β it never declines a question. On raw accuracy it trails only OpenAI's GPT-5.5 (94.1%) and GPT-5.5 Pro (91.6%).
On Effective Value (π) β meoβs fused measure of accuracy, speed, cost and error-cascade resistance β Claude Fable 5 is #1 of all 75 models (π β 1.16), ahead of GPT-5.5 (β 0.98). It earns a correct answer in about 14 seconds, far faster than GPT-5.5 (β 22s) or GPT-5.5 Pro (β 56s), on a 1M-token context with native text, image and file input. In short: the best intelligence-per-cost-and-second on the board.
Fable 5 is a remarkably well-rounded reasoner. It scores a perfect 100% on six of meoβs eleven contamination-resistant domains β theory of mind, unsatisfiable-constraint detection (CSP), multi-step state tracking, regex/finite-automata, long-form exact arithmetic, and Watson-Glaser critical thinking β and lands strong on perceptual illusions (~89%), tape-machine simulation (~93%), and logic/math/CS (~88%).
What separates it from the top two spots is narrow. GPT-5.5 and GPT-5.5 Pro edge ahead on raw micro-accuracy, helped by a couple of domains where Fable 5 is merely good rather than perfect β most notably faithful framework-application (applying a named rule under adversarial pull), where it scores about one-third, and base-rate-bias statistical reasoning (~73%). Its 92.0% macro versus 88.7% micro gap shows the remaining weakness is concentrated in one or two domains, not spread across the board.
This profile changed materially in June 2026: meo discovered its own logic/math/CS ground truth was broken, culled it, and rebuilt the domain with solver-verified answers. On the corrected domain Fable 5 scores ~88% β so the βformal-math weaknessβ reported under the old grading was largely a benchmark artifact, not a model limitation.
The headline, though, is efficiency. Fable 5 answers in about 14 seconds per correct item β roughly 1.6Γ faster than GPT-5.5 and ~4Γ faster than GPT-5.5 Pro β while holding a 100% answer rate, so it is not buying speed by abstaining on hard questions. In meoβs Effective Value model, where the cost of latency compounds across multi-step agentic chains (N=10 task depth), that speed advantage makes it the single best-value model on the board: #3 on accuracy, #1 on π.
Practically, the choice inverts with task depth. For one-shot questions where squeezing out the last accuracy points matters most, the GPT-5.5 family wins. For long, latency-sensitive agent loops β many short reasoning steps where speed, cost and consistency compound β Fable 5 delivers more usable work per dollar and per second than anything else tested.
Strengths
- #1 of 75 models on Effective Value (π) β the best accuracy-per-cost-and-speed on the board.
- Perfect or near-perfect across most reasoning domains: theory of mind, constraint satisfaction, state tracking, regex/automata, long arithmetic and Watson-Glaser β plus a strong ~88% on logic/math/CS.
- Fastest top-tier latency (~14s per correct answer) β about 1.6Γ faster than GPT-5.5 and ~4Γ faster than GPT-5.5 Pro.
- 100% answer rate β it does not hedge or abstain, so its accuracy is earned on the full question set.
- 1M-token context with native multimodal input (text, image, file) β suited to long-document, tool-rich agentic workloads.
Weaknesses
- Faithful framework-application (applying a named rule or rubric under adversarial pull) is its standout soft spot, at about one-third.
- Base-rate-bias statistical reasoning (~73%) trails its otherwise near-perfect domains.
- Slightly behind the GPT-5.5 family on raw micro-accuracy (88.7% vs 94.1% / 91.6%).
- Higher cost per correct answer than GPT-5.5 (~$0.057 vs ~$0.034) β though far cheaper than GPT-5.5 Pro (~$0.32); closed weights, no published knowledge cutoff.
Best for
- Latency- and cost-sensitive agentic workflows with many reasoning steps, where Fable 5's speed compounds into large wall-clock and cost savings β its π leadership is built for exactly this.
- Broad reasoning workloads β including formal logic and math, now that meoβs corrected grading shows ~88% there β with strong social/pragmatic reasoning, constraint checking and stateful tracking.
- Long-context and multimodal pipelines (1M tokens; text + image + file input).
- One caveat: tasks that hinge on faithfully applying a named framework or rubric under pressure, where it is currently weakest β verify those outputs or route them elsewhere.
Why an un-leaked reasoning score is the signal that matters
Public AI benchmarks have a contamination problem β their questions leak into model training data and inflate scores. A 2024 EACL best-paper survey (Balloccu et al., "Leak, Cheat, Repeat") estimated that roughly 4.7 million benchmark samples across 263 benchmarks had been exposed to closed models. When Scale AI built GSM1k β a held-out set matched to the public GSM8K on difficulty, solution length and answer magnitude (NeurIPS 2024) β some model families scored up to 8% lower on the unseen version, and a modelβs tendency to reproduce GSM8K text correlated with its inflated public score (Spearman rΒ²β0.32). meo grades an un-leaked private holdout for exactly this reason: a high score reflects capability, not recall.
The skills meo measures donβt saturate. On LR2Bench, a 2025 constraint-satisfaction benchmark (ACL 2025), even dedicated reasoning models such as DeepSeek-R1 and o1-preview managed only 20.0% and 23.6% exact-match, because constraint problems require multi-step verification and backtracking that single-pass pattern matching cannot fake. Claude Fable 5βs perfect marks on unsatisfiable-constraint detection and multi-step state tracking are a genuine signal, not a memorized answer key.
Theory-of-mind results are famously fragile: Ullman (2023) showed that small, principle-preserving edits to belief-attribution problems can flip a modelβs answer, so apparent success is often shortcut behavior. Holding up across meoβs theory-of-mind items is therefore harder to game than a public score suggests.
Formal logic and math are hard for models to do β and hard to even grade correctly. Appleβs GSM-Symbolic work (ICLR 2025) argues current models replicate training-data reasoning steps rather than reasoning symbolically, a process that stays brittle under variation. meo treats its own ground truth as fallible: in June 2026 it found its logic/math/CS grading was broken, culled the domain, and rebuilt it with solver-verified answers β which lifted Fable 5βs score there from the low double digits to about 88%. The lesson cuts both ways: a single bad domain can distort a leaderboard, which is why transparent, corrected, contamination-resistant grading matters.
Why can a #3-accuracy model lead the entire board on Effective Value? Because π weighs speed and cost alongside accuracy: across long agentic chains, a faster, cheaper model that stays correct yields more usable work per dollar and per second. That speed-and-cost compounding is meoβs own methodology rather than a settled external finding β see the methodology for how π is computed and why the ranking inverts with task depth.
Per-domain accuracy
221 correct of 250 questions asked across 11 domains Β· answer rate 100% Β· macro accuracy 92.0%. Headline accuracy is micro-averaged per meoβs methodology; per-domain question counts (correct/asked) are shown so you can judge significance.
Efficiency
- Cost per correct
- $0.0573
- Seconds per correct
- 14
- Tokens per correct
- 1,053
- Correct answers
- 221 / 250
- Total benchmark cost
- $12.66
- Avg tokens / item
- 930
Model details
- Lab
- Anthropic
- Released
- 2026-06-09
- Weights
- β
- Context
- 1M
- Max output
- 128K
- Knowledge cutoff
- β
- Price in
- $10.00/M
- Price out
- $50.00/M
- Modality
- text+image+file->text
Sources: openrouter Β· meo
FAQ
Why does Claude Fable 5 rank #3 instead of #1?
On raw micro-accuracy it trails OpenAI's GPT-5.5 (94.1%) and GPT-5.5 Pro (91.6%) with 88.7%, mainly because of a weak result on faithful framework-application (~33%) and base-rate-bias reasoning (~73%). It still scores a perfect 100% on six reasoning domains and ~88% on logic/math, and it ranks #1 of all 75 models on Effective Value (π).
Is Claude Fable 5 better than GPT-5.5?
It depends on the workload. GPT-5.5 has higher raw accuracy (94.1% vs 88.7%), so it wins one-shot tasks where correctness is everything. Claude Fable 5 leads on Effective Value (π β 1.16 vs 0.98) because it answers about 1.6Γ faster (~14s vs ~22s per correct answer) at the same 100% answer rate, so on deep, latency-sensitive agentic chains it delivers more intelligence per dollar and per second.
Didnβt Claude Fable 5 score badly on math?
That was a benchmark bug, not the model. In June 2026 meo found its logic/math/CS ground truth was broken, culled it, and rebuilt the domain with solver-verified answers. On the corrected domain Fable 5 scores about 88% β a strength, not a weakness. Its current soft spot is faithful framework-application (~33%).
What is Claude Fable 5 best at?
Latency- and cost-sensitive agentic work. It is the #1 Effective-Value model on the meo benchmark β the fastest of the top tier (~14s per correct answer) at a 100% answer rate β with perfect scores on theory of mind, constraint satisfaction, state tracking, regex/automata, long arithmetic and Watson-Glaser reasoning.
How accurate is Anthropic: Claude Fable 5?
Anthropic: Claude Fable 5 scores 88.7% overall on the meo benchmark (rank #6 of 82), an un-leaked private holdout graded objective-first.
What is Anthropic: Claude Fable 5's Effective Value?
Its Effective Value (π) is 1.1643 at N=10, Ο=1, Ξ΄=1.5 β a fused metric of accuracy, speed, cost, and error-cascade. π rankings shift with task depth.
How much does Anthropic: Claude Fable 5 cost per correct answer?
About $0.0573 per correct answer (1,053 tokens, 14 seconds each) at benchmark-time pricing via OpenRouter.
Is Anthropic: Claude Fable 5 open weights?
Open-weights status for Anthropic: Claude Fable 5 is not recorded.
Benchmark updates affecting this model
- 2026-06-09majorlogic_math_cs ground truth was broken β culled and rebuilt with solver-verified items
A 5-lab frontier jury revealed that 29 of 37 LLM-authored logic_math_cs items had broken ground truth: on most, five independent frontier models unanimously concluded the clue sets were contradictory while the stored answer claimed a number, and on eight, all five agreed on the SAME different number (author arithmetic errors). Models were being marked wrong for being right. The 29 were retired (full audit published in the repo) and the domain rebuilt with generator-as-oracle items: an embedded constraint solver PROVES each puzzle has exactly one solution (or provably none), verified by an independent exhaustive enumerator. A 3-round difficulty study found frontier models saturate consistent human-scale logic puzzles, so items use anonymous global constraints and carry a measured difficulty gradient.
- 2026-06-09Added anthropic/claude-fable-5 β debuts at #2
Benchmarked Anthropic Claude Fable 5 (released 2026-06-09, vision + reasoning, 1M context) over the full holdout the day it shipped, and re-run at maximum (xhigh) reasoning effort with the model dictating its own token + temperature defaults. Concise even at max effort, so the full battery cost only ~$12.
- 2026-06-08majorCorrected base_bias grading (LLM-equivalence judge fix)
The LLM-equivalence grader uses a reasoning model whose verdict token budget was too small, so its YES/NO answer was emitted as empty and silently counted as a failure. This understated the base_bias domain to ~7% for every model when the true accuracy is ~85-90%. We raised the judge budget, excluded its reasoning tokens from the verdict, switched to verbosity-robust conclusion matching, and re-graded the whole board.
Embed this ranking
Show Anthropic: Claude Fable 5βs meo rank on your site β the badge links back here.
Sources & further reading
- meo benchmark methodology & Effective Value (π) β How accuracy, speed, cost and error-cascade are fused; micro vs macro scoring.
- meo benchmark domains β What each contamination-resistant reasoning domain tests and why it resists leakage.
- Effective Value research note β Why π rankings invert with task depth (N).
- meo benchmark changelog β Including the June 2026 logic/math/CS ground-truth rebuild (solver-verified).
- Claude Fable 5 on OpenRouter (pricing & specs) β β Source of metadata, context window and per-token pricing.
- Anthropic β β Model developer.
- Zhang et al., "A Careful Examination of LLM Performance on Grade-School Arithmetic" (GSM1k, NeurIPS 2024) β β Held-out mirror benchmark; up to 8% inflation and rΒ²β0.32 memorization correlation.
- Balloccu et al., "Leak, Cheat, Repeat" (EACL 2024) β β ~4.7M benchmark samples across 263 benchmarks leaked to closed models.
- Chen et al., "LR2Bench" β constraint-satisfaction reasoning (ACL 2025) β β Even DeepSeek-R1 / o1-preview score ~20β24% exact-match; verification + backtracking are hard.
- Ullman, "Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks" (2023) β β Theory-of-mind success is brittle under principle-preserving perturbation.
- Mirzadeh et al., "GSM-Symbolic" (Apple, ICLR 2025) β β Formal logic/math reasoning is fragile pattern-matching, not genuine symbolic reasoning.
- White et al., "LiveBench: A Challenging, Contamination-Free LLM Benchmark" (ICLR 2025) β β Test-set contamination is a well-documented obstacle; continual refresh resists it.
Related: all Anthropic models Β· Unsatisfiable constraints domain Β· best Effective Value Β· biggest movers Β· methodology Β· full leaderboard.