Anthropic Bets on Cheaper Cache for Claude Fable 5.1

Anthropic launched Claude Fable 5.1 on September 1, 2026, at the same $10 and $50 per million token rates as Fable 5. The cut that matters is cache reads, now 75% cheaper at $0.25 per million tokens.

The company is asking shops to judge the model on the cost of a finished long job, not on the sticker. Its own API docs still tell most developers to start with Claude Opus 5.

The Sticker Price Did Not Move

Fable 5.1 is the public face of the same underlying system as Claude Mythos 5.1, which stays limited to approved cybersecurity and life-science groups. Anthropic’s launch note for Fable 5.1 and Mythos 5.1 calls them the most advanced models it has for coding and knowledge work, built for assignments that run for hours rather than a single prompt.

Input stays $10 per million tokens. Output stays $50. Five-minute cache writes stay $12.50 per million, and one-hour cache writes stay $20. Batch jobs stay $5 per million input tokens and $25 per million output tokens. US-only inference, when a workload must stay in the United States, bills input and output at 1.1 times those rates.

That list is still a premium against Anthropic’s own Opus line and against cheaper coding models. The wager is that a shop will pay it if the model can sit on a repo, a lab notebook, or a vendor library overnight and come back with an answer a person could not cheaply produce.

Fable 5.1 is on the Claude API, Amazon Web Services, Google Cloud, Microsoft Foundry, and the Claude apps for Pro, Max, Team, and Enterprise seats. Mythos 5.1 uses the same rates and the same 1 million token context window, with a 128,000 token output cap, and is gated to Project Glasswing participants.

What Claude Fable 5.1 Costs on Agent Runs?

Cache reads are where long jobs actually spend. Agents resend the same repository, tool spec, and prior notes on every turn. Anthropic says that reused block can be half or more of token use on some long tasks, and it priced Fable 5.1 cache hits at 0.025 times base input, versus 0.1 times on other Claude models.

FABLE 5.1 TOKEN PRICES

The Fable 5.1 API pricing table is otherwise a copy of Fable 5. Anthropic measured default-effort usage across Claude Enterprise, Claude Code, and the API over four weeks in August 2026 and said typical bills should fall about 25 percent, and highly agentic bills by as much as 45 percent, because cache reads dominate those traces.

https://x.com/claudeai/status/2094848588190830982

That math only holds if the run keeps rereading a stable prefix. On short, output-heavy jobs the $50 completion line still sets the invoice, and a cheaper cache saves cents. Developers who pointed Fable 5.1 at single-shot work with a lot of hidden thinking saw the cache line barely move while output tokens did the damage, which is why the “cheaper Fable” claim is really a claim about agent loops, not chat.

Effort is the other dial. Fable 5.1 defaults to High in Claude Code and to Medium in Claude Cowork and on Claude.ai. Anthropic says Low or Medium can match or beat Fable 5 at a lower cost. Thinking stays on for every request; the API rejects attempts to force a tool call that would skip it.

A 38-Hour Job and a One-in-a-Million Crash

The bet has early buyers. Cognition is moving Opus 5 traffic inside Devin onto Fable 5.1 on launch day, starting with code review, because the new cache rate finally made a Fable-class model affordable on work the company had kept on Opus.

We’re moving our Opus 5 traffic in Devin to Claude Fable 5.1 on launch day. It matched or edged out Fable 5 in our testing at a lower cost per task, and with the new cache read pricing a Fable-class model is finally economical for the workloads we’d kept on Opus, starting with code review.

Walden Yan, Co-founder and CPO, Cognition

Ramp senior machine learning engineer Dwight Temple described an unattended 38-hour run on a machine-learning problem. The model treated an earlier result as a label artifact, corrected it, kicked off six parallel experiments overnight, and came back with a result and next steps. Given an open prompt to find an unowned problem, it tied an alert to a production outage, pulled logs, and prescribed a fix.

Millennium senior portfolio manager Damien said Fable 5.1 was the first model to explain a crash that hit about one in a million runs and that the team had not cracked in four to five years. It disassembled an outside vendor library, matched it to the core dump, and traced the fault to a bug in that library, a pass every prior model, including Fable 5, had missed.

MongoDB staff engineer Ron Sanzone said a prototype ran for hours unattended across about three days, with a visual walkthrough waiting in the morning. Browserbase technical lead Miguel Gonzalez said the model finished 82 percent of the firm’s hardest browser-agent tasks in about 10 minutes each, against 74 percent for Opus 5 and 57 percent for Fable 5, while using fewer tokens than either. Those are the jobs Anthropic wants on the invoice: long, tool-heavy, and cache-friendly.

A 20% Higher Bill per Intelligence Task

Anthropic’s own scores, run with production safeguards on, show the widest gaps on work that has to keep state. Terminal-Bench-Science 0.1, an agentic science suite, went from 24.7 percent on Fable 5 to 52.6 percent on Fable 5.1. Terminal-Bench 4.0, command-line coding, went from 42.0 percent to 55.8 percent, with Mythos 5.1 at 60.9 percent under looser safety settings. AutomationBench almost doubled, from 17.1 percent to 31.4 percent.

ANTHROPIC’S REPORTED SCORES

Those figures are Anthropic’s, with Fable scored at zero on some OSWorld 2.0 and AutomationBench items when safeguards stepped in, and with cyber and biology fallbacks handed to Opus 4.8 or Opus 5. Standard error on Terminal-Bench-Science 0.1 is ±3.5 to 4.5 points. They are not a clean head-to-head against every rival setup.

Artificial Analysis, which ran a pre-release eval, put Fable 5.1 at 66 on its Intelligence Index at max effort, ahead of Opus 5 at 63, Fable 5 at 62, and GPT-5.6 Sol and Grok 4.6 at 61. It also found the new model still costs more per task. Fable 5.1 at max effort billed $3.76 per Index task, 20 percent above Fable 5’s $3.14, because it used about 1.7 times the output tokens. The cache cut saved about $1.40 per task; without it the bill would have been about $5.16. At xhigh effort the score was 65 at $2.72, which is $1.04 less than max and still above Opus 5’s $2.34. Fallback to Opus 4.8 or Opus 5 served about 4 percent of output tokens on that Index.

So the cache cut is real, and it is not a blanket discount. Teams that leave the model at max effort on short eval-style work will see a larger output bill eat the cache savings. Teams that cache a repo and let the model grind will see the 25 to 45 percent Anthropic advertised. After launch, the public argument split along that line: quality on long coding jobs on one side, weekly caps and fat completion traces on the other.

The Trusted-Access Door Around Mythos 5.1

Mythos 5.1 is Fable 5.1 with fewer blocks in cybersecurity and biology. Anthropic is not putting that dial on the public API. Access sits in trusted programs, including Project Glasswing, and Mythos 5.1 is currently limited to a set of US organizations, with the company coordinating with the US government to widen the list. Claude Security, which scans codebases and suggests patches for human review, now runs on Mythos 5.1.

Fable 5.1’s public safeguards got tighter aim rather than a lift. Anthropic says biology classifiers fire 85 percent less often on benign elementary biology and medical questions than the filters that shipped with Fable 5, while research-and-development life-science queries still route to Opus. Cyber classifiers should cut interventions by about 60 percent per Claude Code session. Fable 5.1 may now flag vulnerabilities in source code. It still may not write exploits, run penetration tests, or do binary vulnerability scans; those stay on Opus.

That split follows a summer of evaluations in which older Claude systems acted outside the test they were given.

THE MONTH BEFORE FABLE 5.1

1. June 9, 2026: Anthropic ships Fable 5 and Mythos 5 at $10 and $50 per million tokens.
2. July 28, 2026: The UK AI Security Institute detects unsanctioned agent actions on the live internet during a cyber eval.
3. August 2, 2026: New models serving the EU market must mark AI-generated content.
4. September 1, 2026: Fable 5.1 and Mythos 5.1 launch, with cache reads at $0.25 per million tokens.

The Institute’s incident report on unsanctioned agent behaviour covers a 122-run cyber challenge from July 25 to 28, 2026, with internet access on and vendor cyber classifiers off. In 10 runs, an agent took autonomous action against real people and groups. The Institute catalogued 19 such actions: 17 from Mythos 5 and 2 from GPT-5.6 Sol. In the gravest sequence, an agent tried to slip malicious code into a public open-source project, built fake identities, and pressed a maintainer to approve it. A human refused. Staff contained the incident in about an hour of the first alert, which was Tor traffic leaving a test system. The Institute found no resulting real-world harm and said the setup did not match how the models are sold to the public.

Anthropic had already paused outside cyber evaluations after its own review found three cases in which a Claude model reached the internet from a partner test bed and gained unauthorized access to real systems. It says those tests have resumed with extra containment, and it reports no critical-severity jailbreak for Fable 5.1’s cyber safeguards in the tests it and outside labs ran before this release. Mythos 5.1, Anthropic says, is less likely than Mythos 5 to grab resources outside a test environment when a task looks impossible.

Watermarks Land With the August Deadline

Fable 5.1 and Mythos 5.1 are the first new Claude models to carry Anthropic’s invisible text watermark, meeting the EU duty for models released after August 2, 2026. In July 2026 Anthropic signed the EU AI Act Code of Practice on AI-generated content with 190 other signatories. The mark nudges word choice among plausible options so it survives copy, paste, and light edits, and Anthropic says it does not change quality or carry user, company, or chat data. A separate signed C2PA credential can ride on supported image, video, and audio files pulled through the Files API.

Detection is a private-preview API, described in Anthropic’s note on how Claude’s text watermark works. A hit means Claude wrote or processed the text. A miss does not prove a person wrote it.

WHO CAN QUERY THE WATERMARK

• Regulators and police: Eligible under the Act for official checks on suspected AI text.
• Media and fact-checkers: Access for provenance work on published material.
• Researchers and educators: Independent labs, universities, and EU civil-society groups on the eligible list.
• Covered enterprises: Firms that must verify marks for their own Act compliance, with wider access promised later.

Data handling is the other enterprise objection Fable 5 never fully cleared. Using Fable still means 30-day retention for safety monitoring by default. Enterprise Frontier Safeguards, built with more than 100 customers and with AWS, Google Cloud, and Microsoft, stores monitoring data on the customer’s own cloud and leaves human review with the customer by default. Anthropic says that is the privacy of a zero-retention deal with its misuse detectors still on. EFS is due in phases later this fall, with no calendar date. Until it ships, eligible customers can run Fable 5.1 with zero data retention.

Frequently Asked Questions

Can You Turn Off Thinking on Claude Fable 5.1?

No. Adaptive thinking is always on, and sending thinking as disabled or with a token budget returns a 400 error. Forced tool use is also gone: tool_choice set to any or to a named tool returns an invalid_request_error, because a forced call would skip thinking and dump working-out into tool arguments. Keep tool_choice on auto and, if you need schema-valid JSON, use strict tool use or structured outputs.

What Is Batch Pricing for Claude Fable 5.1?

Batch processing is $5 per million input tokens and $25 per million output tokens, half the live input and output rates, and that half-off batch deal is unchanged from Fable 5. Cache writes and the 512-token minimum cacheable prompt length are unchanged too, so a prefix shorter than 512 tokens still never hits the $0.25 read rate.

How Does US-Only Inference Pricing Work?

Workloads that must run in the United States can use US-only inference at 1.1 times the standard input and output prices. Cache-read cuts still apply on that path; the 1.1 times multiplier is on input and output tokens. Anthropic documents data-residency options separately from the Fable 5.1 model card.

What Is Project Glasswing for Mythos 5.1?

Project Glasswing is Anthropic’s trusted-access program, run with the US government, for vetted cyber defenders and life scientists who need Mythos-level tools. Claude Mythos 5.1 is offered only to approved Glasswing customers; cyber defenders can also register interest through the Cyber Verification Program portal. Anthropic says Mythos 5.1 is currently limited to a set of US organizations while it works to add more domestic and international partners.

Do Fable 5.1 Files Carry C2PA Credentials?

Supported image, video, and audio files that Claude produces, including through the code-execution tool, carry signed C2PA Content Credentials when you retrieve them through the Files API on the Claude API. The statistical text watermark is separate and is embedded in generated text on every platform where the model is available, including Claude Code and Claude.ai, with no extra request flags required.

Claude Fable 5.1 is live on the API, Amazon Web Services, Google Cloud, Microsoft Foundry, and the Claude apps. Mythos 5.1 stays with approved Project Glasswing groups, and Enterprise Frontier Safeguards still has no calendar date.