Claude Fable 5.1 Cut Cache Reads 75%. Here's the Math.

Start with the cache-read line, not the eval charts.

Claude Fable 5.1 is that rare launch where the invoice moves more than the scores. Because the number doing the moving is a 75% cut to cached-read pricing. Anthropic shipped it September 1, 2026 as the successor to Claude Fable 5, holding sticker pricing flat at $10 per million input tokens and $50 per million output tokens while dropping cached prompt reads to $0.25 per million. A quarter of the old rate. The rest of the card: a 1M-token context window, 128k max output tokens, always-on adaptive thinking. And day-one access across the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and GitHub Copilot. Claude Mythos 5.1 arrived the same day, the same underlying model wearing different safeguards.

The shop I run builds automations, so launches reach me as invoices rather than press events. What follows is the arithmetic, plus one report worth holding at arm's length.

Claude Fable 5.1 Release: What Shipped September 1, 2026

The launch announcement introduces the pair as "the world's most advanced models for coding and knowledge work." The release notes describe Claude Fable 5.1 (model ID `claude-fable-5-1`) as the successor to Fable 5 for long-running agentic coding, knowledge work. And research, with Claude Mythos 5.1 (`claude-mythos-5-1`) succeeding Mythos 5 for Project Glasswing participants.

Per the "What's new" doc, the release "brings stronger long-running agentic coding, multistep research," rounded out by document, spreadsheet, and slide work. The capability that pays rent: agents staying alive across long sessions, research that survives multiple steps, output landing in files clients actually open. Spreadsheets, mostly, the one file format that has quietly outlived everything built to kill it.

On benchmarks, CryptoRank's coverage, citing Anthropic's own disclosures, reports Fable 5.1 more than doubled Fable 5's agentic science score to 52.6% and lifted coding to 55.8%. Treat the deltas as directional. The model docs set the knowledge cutoff at June 2026 with default effort `high`. That cutoff is the detail that decides whether your workflows reference recent events correctly.

Availability was genuinely day one: the Claude API for all customers, Claude in Amazon Bedrock, Claude Platform on AWS, Google Cloud. And Microsoft Foundry, with AWS confirming general availability. Anthropic's product page lists Pro, Max, Team, and Enterprise access, and GitHub announced general availability in Copilot.

Claude Fable 5.1 Pricing: The Cut That Beats the Benchmarks

Sticker price, unchanged. $10 in, $50 out, same as Fable 5.

What moved is everything around caching, which is where agentic workloads actually spend. Here is the full rate card from Anthropic's pricing table.

| Line item | Price per million tokens | |---|---| | Input | $10 | | Output | $50 | | Cache read | $0.25 | | 5-minute cache write | $12.50 | | 1-hour cache write | $20 |

The release notes do the ratio for you: cache reads run at 0.025x the base input price, versus 0.1x on other Claude models. At a $10 base, 0.1x is $1 per million. Reads on Fable 5.1 cost $0.25. That is the whole story.

The write side tells the other half, as a 5-minute cache write costs 50x a read and the 1-hour write costs 80x.

This is a rate card built for patience.

Long runs in my shop re-send the same system prompt, tool definitions, and client reference docs on every step. Precisely the traffic the new read price pays out on.

Stable prefix, and the model gets cheaper the longer the session runs.

Context that churns, and you pay write prices rebuilding the cache until the win shrinks to noise.

CryptoRank's coverage reports typical workload savings around 25%, climbing to 45% for complex agent tasks. Those percentages are the outlet's reporting, not a line on Anthropic's pricing page. So bank the ratio you can verify and treat the rest as directional.

Do you know what share of last month's bill was cache reads? Most operators do not, which is the actual reason launches like this get misjudged. The fix takes an hour: split spend into input, output, cache reads, and cache writes. The read number tells you whether this launch was built for you.

The decision rule falls out of the arithmetic. Long agentic runs with stable context should switch to `claude-fable-5-1` now, since the 75% read cut compounds at every step of every session. Short one-shot calls have no reason to move, since the sticker is unchanged.

Security-adjacent work gains for a other reason, covered next.

Claude Fable 5.1 vs Mythos 5.1: Same Model, Separate Safeguards

Anthropic says it plainly: the two are "the same model. But with other levels of safeguards." Per the Mythos page, Fable 5.1 ships with safeguards for cybersecurity and biology, while Mythos 5.1 relaxes them for vetted cybersecurity and biology/life-sciences users. And it is the version tied to Project Glasswing.

The precision claims deserve a second read. Fable 5.1 "can now be used to identify software vulnerabilities in source code," and Anthropic states its biology safeguards "intervene on benign requests 85% less often than the ones we launched with Fable 5." That 85% is the operator's number. Refusals on legitimate security work were a real tax on the previous generation. And a published, specific reduction is something you can check against your own logs in a way a vague promise of improved safety never is.

Gating capability through vetting rather than deleting it from the model strikes me as the more honest architecture. The line is visible. When your work sits near it, you know who to ask.

The Anti-Distillation Report Nobody Can Confirm

The spiciest claim in this launch is not on any official page. CryptoRank reports that Anthropic added an anti-distillation lock covering API accounts opened on or after August 31, after tracing roughly 16 million exchanges from about 24,000 fake accounts attempting to copy the model — about 3.4 million of those exchanges tied to Moonshot. One outlet. I cannot confirm any of it in Anthropic's documentation, so hold it loosely.

If the report holds, the signal outruns the mechanism.

Frontier labs defending their models through account-level controls means account age and terms of service become part of your infrastructure risk. Read the API agreement before assuming a new account behaves like your old one.

FAQ: Claude Fable 5.1 Pricing and Availability

How much does Claude Fable 5.1 cache read cost?

$0.25 per million tokens — 75% below the prior rate, running at 0.025x the base input price versus 0.1x on other Claude models. Cache writes cost $12.50 per million for the 5-minute tier and $20 for the 1-hour tier.

When was Claude Fable 5.1 released?

September 1, 2026, as the successor to Claude Fable 5, with Claude Mythos 5.1 shipping the same day.

What is the context window on Claude Fable 5.1?

A 1M-token context window, with 128k max output tokens and always-on adaptive thinking.

Where is Claude Fable 5.1 available?

Day one on the Claude API for all customers, Claude in Amazon Bedrock, Claude Platform on AWS, Google Cloud. And Microsoft Foundry, plus general availability in GitHub Copilot. Anthropic's product page lists Pro, Max, Team, and Enterprise access.

How much does Claude Fable 5.1 cost per token?

$10 per million input tokens and $50 per million output tokens — identical to Claude Fable 5.

If you want the invoice audit run against your real numbers instead of a vendor's math, that is the work this shop does.

Bring last month's bill and an hour. And you will leave knowing exactly how much of that 75% cut belongs to you.

Sources

- Anthropic launch announcement - Claude release notes - What's new in Claude Fable 5.1 - Claude Fable 5.1 model docs and pricing - AWS: Claude Fable 5.1 general availability - Anthropic Claude Fable product page - Anthropic Claude Mythos page - GitHub Copilot availability announcement - CryptoRank coverage