Anthropic Ships Fable 5.1 at Same Price, Cuts Cache Reads 75% to $0.25 per Million

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1 at Fable 5’s unchanged base price, while reducing the price of cache reads. The discount takes reused context to $0.25 per million tokens. Artificial Analysis measured Fable 5.1 at maximum effort as costing more per completed task than Fable 5 because it produces about 1.7 times as many output tokens.

The price change

Base pricing remains $10 per million input tokens and $50 per million output tokens, twice Claude Opus 5’s rates as of September 2026. Fable 5.1 is generally available. Mythos 5.1 is the same model with more permissive safeguards, restricted to trusted-access programs that cover U.S. organizations.

Anthropic’s claimed savings of about 25% on typical workloads and as much as 45% on highly agentic work came from four weeks of its own August 2026 usage at default effort. A workload without reusable context receives no benefit from the 75% cache cut.

Two measures of cost

Artificial Analysis, which performed pre-release evaluation for Anthropic, measured Fable 5.1 at maximum effort at $3.76 per Intelligence Index task on September 1. That was 20% above Fable 5’s $3.14 and 1.6 times Claude Opus 5’s $2.34, driven by about 1.7 times as many output tokens as Fable 5. The cheaper cache saved about $1.40 per task.

Fable 5.1 led the index at 66, against 63 for Claude Opus 5. At extra-high effort, it scored 65 at $2.72 per task, still above Claude Opus 5’s $2.34 for a score of 63. It attempted more questions than Claude Opus 5 and, among the questions it got wrong, answered rather than declining more often than Fable 5.

Cognition found the opposite cost result on FrontierCode 1.1 Extended, Cognition’s coding benchmark that scores whether a model’s patch is mergeable, at medium effort. Fable 5.1 cut the September 2026 cost per task from Fable 5’s $5.84 to $2.68, below Claude Opus 5’s $3.51. More than 95% of tokens were cache reads, and Fable 5.1 used 33% fewer tokens than Claude Opus 5. Its score peaked at medium effort and fell below Fable 5 at higher settings because it made unrequested changes outside the task.

Walden Yan, Cognition’s co-founder and chief product officer, said: “We’re moving our Opus 5 traffic in Devin to Claude Fable 5.1 on launch day.”

Vendor benchmarks

Anthropic measured Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1, a test of agentic scientific research tasks run in a terminal, compared with 24.7% for Fable 5 and 29.0% for Claude Opus 5 in its September 2026 harness. The benchmark was first announced on August 27, five days before the launch, and Anthropic reported a standard error of 3.5 to 4.5 points per model.

The benchmark and scientific figures, including protein-binder and GPU-kernel results, are Anthropic’s measurements and had not been independently reproduced at publication. Artificial Analysis’s test used Anthropic’s default fallback, which routed about 4% of output tokens to Claude Opus 4.8 or Claude Opus 5.

On Terminal-Bench 4.0, Fable 5.1 scored 55.8% and Mythos 5.1 scored 60.9% in September 2026. Anthropic said the gap between the identical models came from safeguard interventions. Tasks where Fable’s production safeguards fired received zeroes on OSWorld 2.0.

The fine print

Anthropic’s documentation tells customers to start with Claude Opus 5 for most workloads and move to Fable 5.1 when Opus falls short at higher effort. Anthropic’s model documentation says Fable 5.1 and Mythos 5.1 carry 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it, while the announcement says eligible customers can use Fable 5.1 with zero data retention until Enterprise Frontier Safeguards is available. Enterprise Frontier Safeguards, presented as equivalent to zero retention, begins a phased rollout this fall and is not available today.

At launch, forced tool use returns an error, earlier models cannot read Fable 5.1 thinking blocks, and editing earlier turns invalidates thinking blocks.

The system card assesses alignment-related risk as “low rather than very low.” It reports that during external testing a partner observed Mythos 5.1 “exploiting a sandbox vulnerability to read files outside its environment,” an incident Anthropic rated low severity and said it reports for transparency. It describes Mythos 5.1 as “less honest under pressure than recent Claude models.”

All text output carries an EU AI Act watermark, while its detection API is in private preview.

Digital Applied wrote: “The 25% figure is real and it is measured, but it is a measurement of Anthropic's workloads, not a discount applied to yours.”

Anthropic's Opus 5 Cut Tokens 17% but Cost More to Benchmark Than Opus 4.8

Anthropic released Claude Opus 5 on July 24 with per-token API rates unchanged from Opus 4.8, the company announced. Artificial Analysis reported that its independent Intelligence Index run cost $3,83

The Implicator

GPU Price Trackers Disagree by 8x as CME Prepares Compute Futures

Rival GPU price trackers put the cost of the same H100 chip roughly eight times apart, based on a July 29 table from Gpuinstances and a Silicon Data index published April 27. The eightfold gap compar

The Implicator

Sonnet 5 Closes Most of the Gap to Opus 4.8 on Agent Work

Anthropic released Claude Sonnet 5 on June 30 with a simple pitch: most of Opus 4.8's capability at well under half the cost. On the company's own agent benchmarks, the model trails its flagship by a

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai