Fable 5.1 Cuts Cache Cost While Taking WebDev

Claude Fable 5.1 did not get cheaper on the sticker. Anthropic made it generally available on September 1, 2026 at the same $10 input / $50 output per million tokens as Fable 5, with the same 1M context and 128K max output. The real invoice change is cache reads: $1 down to $0.25 per million tokens, 0.025x base input versus 0.1x on other Claude models. Anthropic estimates typical workloads about 25% cheaper than Fable 5, and highly agentic work up to about 45% cheaper. The API id is claude-fable-5-1.

On arena.ai's September 2, 2026 text overall snapshot, claude-fable-5.1-max sits at 1504±11 Elo on 2,906 votes and is not flagged Preliminary. Predecessor claude-fable-5 is still 1507±5 on 27,189 votes; helloai now tracks 5.1 at 1504. Code Arena WebDev is the bigger shift: claude-fable-5.1-max is 1765, rank 1, while grok-4.6-high is 1629 Preliminary. That ends Grok 4.6's brief WebDev edge over Fable 5, which sat at 1626.

Vendor numbers are noisier and should be treated as vendor numbers. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6% against Fable 5's 24.7% and Opus 5's 29.0%. On CursorBench 3.2.0, SpaceXAI's Sualeh Asif said Fable 5.1 scored 73.4% at max effort, the highest they had run on that bench. Cognition's Walden Yan said they are moving Devin's Opus 5 traffic to Claude Fable 5.1 on launch day because it matched or edged Fable 5 at lower cost per task, and cheaper cache reads make a Fable-class model economical for workloads they had kept on Opus, starting with code review.

The next day, Meta shipped Muse Spark 1.3 on Muse Code and the Meta Model API at the same $1.25/$4.25 and 1M context as 1.2. Meta says about 20% fewer tool calls and about 25% fewer tokens versus 1.2 in internal comparisons. Max reasoning is not out yet — safety testing. Arena has no 1.3 slug yet, so helloai keeps 1.2 xHigh's 1499 Elo as the fallback.

helloai still tracks six frontier APIs. Fable 5.1 replaces Fable 5; Muse Spark 1.3 replaces 1.2. Gemini 3.1 Pro, Claude Opus 5, Qwen3.8-Max, and Grok 4.6 are unchanged this week, and Grok 4.6's text overall remains Preliminary at 1461 on 3,453 votes. A same-price successor with a 75% cheaper cache read rewrites routing tables before it rewrites leaderboards. If cache-heavy agent loops keep migrating off Opus-class prices, the interesting split is no longer Fable versus Grok on WebDev — it is which workloads still justify paying Opus once Fable-class cache is this cheap.