MODELS & COSTS

Fable 5.1: A Cache Discount Is Not the Whole Bill

Open Source Factory · Sep 01, 2026 · 3 min read

Fable 5.1's 75% reduction applies to cache reads. A task that reuses a large input can benefit much more than one with little cached material. The unchanged input/output rate therefore says less about the final bill than the traffic shape and reasoning setting.

In this article

1. The cache-read price reduction
2. Reported capability changes
3. Effort as a cost variable
4. Mythos 5.1 and safeguard conditions

Fable 5.1 and Mythos 5.1 — same model underneath, different safeguards. (image: Anthropic)

The launch video — Fable 5.1 on a sky-and-moon cover. (image: Anthropic / YouTube)

The cache-read price reduction

Agent turns often reuse a large context, which makes cache pricing relevant. Cache tokens are not necessarily most of the bill: output tokens, uncached input and tool charges can dominate. Use the run’s usage fields to calculate the share that actually receives the lower rate.

Same sticker price, much lower effective cost. Indexed cost from Anthropic's own 4-week August 2026 usage data. Vendor-reported, redrawn.

Reported capability changes

The benchmark table has two kinds of jumps. There are the "good, incremental" ones — CursorBench 73.4% vs Fable 5's 70.5%, OSWorld up ~5 points. And then there are the "wait, really?" ones. Terminal-Bench-Science more than doubled: 52.6% vs 24.7%. AutomationBench nearly doubled: 31.4% vs 17.1%. Those aren't polish improvements, they're a different capability class showing up in the eval.

Fable 5.1 vs Fable 5, Opus 5, GPT-5.6 Sol. The science and business-workflow jumps (2x) are the real story; the coding gains are solid but incremental. Vendor-reported, day-1, no independent verification yet.

The real-world examples back it up in a way benchmarks rarely do. Millennium — an investment firm with some of the sharpest quant engineers around — says Fable 5.1 finally found the cause of a rare crash in their internal systems that no engineer and no other model had explained after years of trying. It disassembled an external vendor library, matched it against the core dump, and traced the crash to a bug in that library. That's the kind of story that moves money.

And the science stuff is wild: a neural network Fable 5.1 trained produced a new elevation map of a third of Venus at 2-3km resolution (the old map was 10-20km), using 30-year-old Magellan radar data. Anthropic released it under a Creative Commons license for the upcoming NASA and ESA missions.

Effort as a cost variable

Fable 5.1 has adaptive thinking always on, with an effort dial (low / medium / high). The interesting part: at Low or Medium effort, it matches or beats Fable 5's results — at much lower cost. It defaults to High in Claude Code, Medium in Claude Cowork and on claude.ai. So the "downgrade" isn't a downgrade anymore; it's the same quality at a cheaper setting.

Author's reconstruction from Anthropic's effort-level description (approximate, illustrative). Low/Medium effort ≈ Fable 5 quality at a fraction of the cost.

Mythos 5.1 and safeguard conditions

Fable 5.1 and Mythos 5.1 are the same model with different guardrails. Mythos — the version with stronger cyber and biology safeguards — is only available through trusted-access programs (cyberdefenders, life scientists, US orgs first). The safeguards themselves improved: 60% fewer false positives in cybersecurity, and the model can now find vulnerabilities but not build exploits. Biology access runs through a program built with the US government.

Two things worth knowing before you migrate. First, forced tool use is gone — setting tool_choice to a specific tool now returns a 400 error; you use strict tool use instead. Second, thinking blocks are bound to the model that made them — if you route a conversation from Fable 5.1 down to an older model, it can't read those blocks, and editing earlier turns invalidates them. Treat the conversation as append-only.

And there's a new Enterprise Frontier Safeguards system coming this fall: zero-data-retention-equivalent privacy by storing data in the customer's own cloud, not Anthropic's. Until then, eligible enterprise customers get Fable 5.1 with ZDR.

For every million eligible cache-read tokens, the new rate saves $0.75 against the old one. A cache-heavy job may benefit substantially, while a short job with little reused input may see much less. That calculation is more informative than the unchanged headline price.

The reported science and automation gains need their benchmark conditions and independent checks. For an existing workflow, log the effort setting and completed-task bill before and after the change. Anthropic’s demonstration is embedded above for comparison with its own claims.

Sources: Anthropic announcement · Claude platform docs · OpenRouter pricing · benchmarks vendor-reported as of September 1, 2026. Price comparison context: my earlier post on $10 coding subscriptions.

Edited October 26, 2026. Original publication and evidence dates are retained; this edit is not a new product test.

Reading guideAbout & methodsCorrectionsPrivacy