# Claude Fable 5.1 By Far Most Expensive Model On Artificial Analysis Intelligence Index, Costs 57% More Than Opus 5 > Source: > Published: 2026-09-02 08:25:06+00:00 It appears that incremental improvements to intelligence are now coming at a fairly big cost. Anthropic’s [Claude Fable 5.1](https://officechai.com/ai/fable-5-1-benchmarks/) has taken the top spot on the Artificial Analysis Intelligence Index with a score of 66, the highest the benchmarking firm has ever recorded. But that top spot comes at a price nobody else on the leaderboard is charging. At max effort, Fable 5.1 costs $3.69 per Intelligence Index task, making it by a wide margin the most expensive model Artificial Analysis tracks, and 57% pricier than [Claude Opus 5](https://officechai.com/ai/claude-opus-5-benchmarks/) (max), Anthropic’s own second-most capable model, which runs at $2.34 per task. ## The Pricing Picture Cost is where Fable 5.1’s launch gets complicated. Anthropic did cut cache read pricing by 75%, from $1 to $0.25 per million cached input tokens, while leaving standard rates unchanged at $10 per million input tokens and $50 per million output tokens. On paper, that should make Fable 5.1 cheaper to run, especially for agentic workloads that lean heavily on cached context. In practice, it seemingly doesn’t. Fable 5.1 at max effort still costs 20% more per task than its predecessor, Claude Fable 5 (max), which came in at $3.14. The reason is token usage: Fable 5.1 generates roughly 1.7 times the output tokens Fable 5 did to reach its answers, and output tokens are priced far higher than cached input, so the extra verbosity wipes out most of the savings from cheaper caching. Artificial Analysis estimates the cache cut alone saves about $1.40 per task; without it, Fable 5.1 (max) would cost closer to $5.16. Stepping down a notch to xhigh effort helps, but only somewhat. At that setting Fable 5.1 scores 65 on the Intelligence Index, just one point below its max-effort score, while costing $2.72 per task, $1.04 less than max. That’s still above Claude Opus 5 (max) at $2.34 for a lower Intelligence Index score of 63. To put the gap in context against the rest of the field, here’s how cost per task stacks up across leading models on the Artificial Analysis chart: GPT-5.6 Luna (max) comes in at just $0.05, DeepSeek V4 Pro 0813 (max) at $0.27, Muse Spark 1.2 (xhigh) and Gemini 3.7 Flash (high) both around $0.40, GLM-5.3 (max) at $0.68, Kimi K3 (max) at $0.84, and Grok 4.6 (high) and GPT-5.6 Sol (max) both approaching a dollar. Claude Opus 5 (max) at $2.34 was already the priciest model on the board before Fable 5.1 arrived at $3.69, nearly 74 times the cost of GPT-5.6 Luna for a task. Fable 5.1’s effort settings span an enormous range in token usage: from 13.1 million output tokens at low effort up to 143.7 million at max, an 11x spread, with Intelligence Index scores moving from 58 to 66 across that range. Even at its most frugal setting, Fable 5.1 uses more tokens than most rival flagships need at their most capable — OpenAI’s GPT-5.6 Sol at medium effort uses marginally fewer output tokens (12 million) than Fable 5.1 does at its lowest setting (13.1 million). Artificial Analysis notes that Fable 5.1 does sit on the Pareto frontier of intelligence versus output tokens used, meaning every model variant that scores higher than GPT-5.6 Sol (medium) is matched or beaten by some Fable 5.1 effort level on both intelligence and token efficiency. That’s a genuine efficiency claim, but it doesn’t change the fact that on raw dollar cost per task, nothing else comes close to what Fable 5.1 charges at its higher effort settings. ## Benchmark Gains Behind The Price The benchmark improvements are real. Fable 5.1 gains four points on the Intelligence Index over Fable 5, moving from 62 to 66, and edges out [Claude Opus 5](https://officechai.com/ai/claude-opus-5-becomes-top-model-in-the-world-on-artificial-analysis-intelligence-index-beats-fable-5/) (max, 63), Fable 5 (max, 62), GPT-5.6 Sol (max, 61) and Grok 4.6 (high, 61). On Humanity’s Last Exam, it scores 59.1%, ahead of the previous best of 55.5% set by Fable 5 itself. It also posts the highest scores Artificial Analysis has measured on Terminal-Bench v2.1 (91.4%) and SciCode (62.0%), and gains nine points over Fable 5 on τ³-Banking. On agentic work, Fable 5.1 (max) sets new highs on GDPval-AA v2 at 1,853 Elo, up 130 points over Fable 5, and on AA-Briefcase at 1,694 Elo, up 122 points. But against its own sibling, Claude Opus 5, the picture is closer than the headline suggests. The GDPval-AA v2 lead over Opus 5’s 1,824 Elo sits within the confidence interval, and on AA-Briefcase the two models are effectively tied, 1,694 to 1,685. Digging into the AA-Briefcase sub-scores shows Fable 5.1 ahead on analytical quality (2,025 versus 1,980) but behind Opus 5 on presentation (1,495 versus 1,572). Fable 5.1 also attempts more questions on AA-Omniscience, Artificial Analysis’s knowledge benchmark, going after 93.4% of questions compared to Opus 5’s 87.8%, and posts the highest accuracy the firm has measured at 67.2%, ahead of Fable 5’s 65.4%. That comes with a catch: of the questions it got wrong, it still attempted to answer 72.6% of the time, well above Fable 5’s 63.6% rate, indicating more willingness to guess rather than abstain. The higher accuracy and higher hallucination rate roughly cancel out, leaving Fable 5.1’s overall AA-Omniscience Index score level with Fable 5’s. Fable 5.1 retains a 1 million token context window with image and text input support, in line with Anthropic’s other recent releases, and keeps the same $10/$50 per million token input/output pricing and $12.5 cache write price as Fable 5, with only the cache read price cut applying as new. Evaluations were run with Anthropic’s default server-side fallback active, which routes safety-flagged requests to either Claude Opus 4.8 or Claude Opus 5; fallback handled about 4% of output tokens across the Intelligence Index run. That’s a similar setup to how Artificial Analysis evaluated the original [Claude Fable 5](https://officechai.com/ai/anthropic-extends-claude-fable-5s-access-on-paid-plans-until-19th-july/), whose access was briefly suspended and later restored following US export control action. For teams deciding whether the jump to Fable 5.1 is worth it, the numbers suggest a fairly narrow case: meaningfully better on raw intelligence and on tasks like coding and knowledge work, but priced well ahead of Claude Opus 5 for gains that are, on some agentic benchmarks, within the margin of error.