Claude Fable 5.1 Benchmarks: 27.9 Points, All Agentic

Claude Fable 5.1 leads Claude Fable 5 and Opus 5 on all nine benchmark rows Anthropic reported — but the gains are bunched almost entirely in agentic work, and the sticker price did not move.

Here is where the numbers actually land, and what that means for deciding whether to switch.

The gains are lopsided on purpose

From Anthropic's September 2026 evaluation table, as compiled in a benchmark write-up published this week:

Benchmark	Fable 5.1	Fable 5	Opus 5
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%
Terminal-Bench 4.0	55.8%	42.0%	52.3%
AutomationBench	31.4%	17.1%	26.9%
CursorBench 3.2.0	73.4%	70.5%	70.0%
Humanity's Last Exam, with tools	65.0%	63.8%	63.6%

The spread is the story. Terminal-Bench-Science jumped 27.9 points over Fable 5, which the write-up notes is well outside the reported per-model standard error of roughly ±3.5 to 4.5 points. AutomationBench rose 14.3 points, an increase of about 84% in relative terms. CursorBench moved 2.9 points and tool-assisted Humanity's Last Exam moved 1.2 — both inside or near that error band.

Translated: if your workload is an agent that plans, calls tools, checks its own output and recovers from mistakes over many steps, this release is aimed at you. If it is short IDE coding turns or closed-ended reasoning, the measured difference is small.

Two caveats the write-up flags and that are worth carrying. Fable 5.1's edge over Opus 5 on Terminal-Bench 4.0 is only 3.5 points, which sits inside the same error range and should be treated cautiously. And Anthropic's separately reported 60.9% for Claude Mythos 5.1 on Terminal-Bench 4.0 is a different deployment with different safeguards — not a generally available Fable 5.1 score.

The price change is in the cache, not the sticker

Input stays at $10 per million tokens and output at $50 per million — unchanged from Fable 5. The one that moved is cache reads, down 75% from $1 to $0.25 per million tokens.

Anthropic estimates roughly 25% lower cost for typical workloads and up to roughly 45% lower for highly agentic ones. That shape is consistent with the benchmark story: repeatedly resending repository context, tool definitions and accumulated state is exactly what a cache-read discount rewards.

Other specs per Anthropic's documentation: a 1,000,000-token context window, 128,000-token maximum output, text and image input, adaptive thinking always on, a June 2026 knowledge cutoff, and model ID claude-fable-5-1. Comparative latency is listed as slower.

So is it worth switching?

Answer it with your own workload, not the leaderboard. Two questions decide it:

• Does your agent re-send the same large context on every turn? If yes, the cache-read cut is a real line-item reduction even before any quality change, because headline input and output rates did not move.
• Is your failure mode "gave up halfway" or "got the answer wrong"? The measured gains sit in long-horizon execution and recovery. For single-turn correctness, the gap over Fable 5 is small.

One number is worth keeping in view against the enthusiasm: AutomationBench nearly doubled and still sits at 31.4%, meaning roughly two thirds of those end-to-end business workflow tasks did not complete under Anthropic's reported configuration. This is a better agent, not a solved one.

Worth watching: a GPT-6 Astra comparison. The write-up notes it was not part of Anthropic's September 1 benchmark table, so any direct ranking needs a controlled run on the same harness rather than a cross-table read.