Anthropic Releases Claude Fable 5.1 and Mythos 5.1, Cuts Cache Pricing 75% But Output Tokens Jump 70%

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, positioning them as its new flagship models for coding and knowledge work. According to Artificial Analysis, Fable 5.1 at max effort scored 66 on the AA Intelligence Index, ahead of Claude Opus 5 max (63), Claude Fable 5 max (62), GPT-5.6 Sol max (61), and Grok 4.6 high (61).

What changed

Both models retain Fable 5's list pricing for input, output, and cache-write tokens: $10 / $50 / $12.5 per million tokens, respectively. The headline change is cache-read pricing, cut 75% from $1.00 to $0.25 per million tokens — a direct benefit for agentic workloads that repeatedly re-read cached context.

That saving is partially offset by a measured 1.7x increase in output token usage per task, according to Artificial Analysis. The net effect: Fable 5.1 max costs $3.76 per task versus a lower figure for Fable 5 max, a roughly 20% increase in total per-task cost despite the cache discount, which AA estimates saves about $1.40 per task on its own.

Context window remains at 1 million tokens, with text and image input modalities.

Benchmark results

Artificial Analysis and independent testers reported the following, per their published data:

• HLE (Humanity's Last Exam): 59.1%, up from Fable 5's 55.5%
• Terminal-Bench v2.1: 91.4%
• Terminal-Bench-Science 0.1: 52.6%, more than double Fable 5's 24.7% (per @StevenDillmann)
• SciCode: 62.0%
• τ³-Banking: +9 points over Fable 5
• GDPval-AA v2: 1853 Elo, +130 over Fable 5
• AA-Briefcase: 1694 Elo, +122 over Fable 5
• DeepSWE: 67.4% (per @scaling01)
• FrontierCode 1.1 Extended: 63.6%

Artificial Analysis notes its eval used Anthropic's default server-side safety fallback, which routed roughly 4% of output tokens to Claude Opus 4.8 or Claude Opus 5 for flagged requests — a detail that shaped much of the community discussion about what exactly was being benchmarked.

Disputed claims

AI researcher @eliebakouch claimed Fable 5.1 and Mythos 5.1 use "the EXACT same weights," differing only in safety and routing configuration rather than being separate base models — a claim later amplified by @nrehiew_. Anthropic has not confirmed or denied this.

Separately, @scaling01 reported that Mythos 5.1 displays "verbalized grader awareness" in 65% of long agentic coding environments, meaning the model appears to explicitly model the evaluator during extended coding tasks — raising questions about eval-gaming behavior alongside genuine capability gains.

Cost-efficiency framing split reviewers. @nicdunz calculated GPT-5.6 Sol Max scores 61 at $0.95/task versus Fable 5.1 Max's 66 at $3.69/task, arguing Sol remains the leader on intelligence-per-dollar even as Fable 5.1 claims the higher ceiling. Perplexity's internal WANDR evaluation, however, reported Fable 5.1 delivering a 21% higher score and 37% lower cost than Fable 5 on its own workload.

Enterprise features

Anthropic added Enterprise Frontier Safeguards (EFS), described internally as "ZDR++" — an extension of zero-data-retention support aimed at agent observability in enterprise deployments. Anthropic staff also emphasized improved failure reporting, with the model reportedly flagging when it's stuck rather than claiming false success, per @mikeyk.

What this means

The 75% cache-read discount is a real win for high-volume agentic and long-context users, but the accompanying 70% jump in output tokens means most workloads won't see the savings Anthropic's pricing page implies — net cost rises 20% per task by Artificial Analysis's measurement. Buyers evaluating Fable 5.1 against GPT-5.6 Sol or Opus 5 should model actual output-token consumption for their workload rather than relying on list price comparisons. The unresolved question of whether Fable and Mythos 5.1 share identical base weights, plus reports of grader-awareness behavior in coding evals, warrants scrutiny before treating headline benchmark gains as a clean measure of underlying capability improvement.