Anthropic shipped a pair of models on Tuesday that tells you where this industry is headed better than any benchmark chart. Claude Fable 5.1 and Claude Mythos 5.1 are, by Anthropic’s own description, the same model wearing different guardrails. Fable 5.1 is generally available today through the Anthropic API and the big cloud platforms. Mythos 5.1 is locked down to vetted cybersecurity and life-science researchers. Same brain, two cages. That alone is a policy story worth watching.

The part that should matter to anyone running real workloads is the rest of the announcement: pricing that drops as much as 45 percent on agentic work, a new enterprise privacy tier, and safeguards that block less of the work you actually paid for. Here is what shipped, what it costs, and the questions to ask before migrating anything.

What Actually Shipped: One Model, Two Front Doors

Anthropic is explicit that Fable 5.1 and Mythos 5.1 are the same underlying model with different levels of safeguards. Fable 5.1 is the general release, available on cloud platforms and through the API starting September 1. Mythos 5.1 carries safeguards tuned for cybersecurity and life sciences and is restricted to trusted access programs. Right now that means a set of US organizations, with access widening over time.

That structure matters. Anthropic has accepted that its most capable model needs vetting, while selling a restricted version to everyone else. The two-flavor approach is becoming the standard template in frontier AI, and the gap between flavors is now a product decision as much as a safety one.

The Pricing Story Is the Real News

Forget the benchmark bragging for a second. The number that should grab any operator: Fable 5.1 costs an estimated 25 percent less than Fable 5 for typical token-billed workloads, and up to roughly 45 percent less for highly agentic work. Anthropic got there by cutting cache-read pricing on inputs the model has already processed and stored.

Reported pricing lands at 10 dollars per million input tokens, 50 dollars per million output tokens, and 0.25 dollars per million cached input tokens, with a 1-million-token context window and 128,000-token maximum output. I have not seen every regional price sheet, so treat exact figures as directional, but the shape is clear: Anthropic is attacking the cost of long-running, repeated-context work, which is exactly what agent workloads burn through.

The underrated feature is effort control: at Low or Medium effort, Fable 5.1 matches or beats Fable 5 at full effort for a fraction of the cost. It defaults to High in Claude Code and Medium in Claude Cowork and on Claude.ai. For agent fleets, that dial is a cost dial.

Zero Data Retention Without the Trust Fall

Enterprises have asked for zero data retention for years, and Anthropic kept saying safeguards could not survive it. Fable 5.1 ships with a new answer: Enterprise Frontier Safeguards, or EFS. Customer data lives in cloud infrastructure controlled entirely by the customer, not on Anthropic’s systems, and any human review of activity is done by the customer by default. Anthropic says that gives clients the privacy of a zero data retention policy while still catching adversarial use.

EFS was built with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, and the public sector, plus AWS, Google Cloud, and Microsoft Azure. It will run on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google’s Agent Platform, and Microsoft Foundry, rolling out in phases starting this fall. Until EFS is ready, eligible customers can run Fable 5.1 with zero data retention. For regulated industries, this is the big one: the privacy objection was never just about model quality, it was about where data sat and who could look at it.

Safeguards That Block Less of the Work You Paid For

Anyone who has watched Claude refuse a legitimate task knows false positives are a tax on security teams, medical researchers, and developers working near sensitive topics. Anthropic says Fable 5.1’s cybersecurity safeguards block 60 percent fewer false positives than before, in part because the model can now be used to discover software vulnerabilities, though not to develop exploits.

The biology side is similar, with 85 percent fewer false positives on benign requests about elementary biology and medical questions. Genuine dual-use work still gets routed: penetration testing, exploit generation, and binary-based vulnerability scanning are redirected to Opus models, and advanced life-science queries still go to Opus unless you are inside the Mythos access program. The guardrails are lighter, not gone. If your shop does serious security research, plan for those redirects.

The Science Output Is the Flex

Anthropic rarely releases a model without some demonstration of research chops, but this one shows what long-horizon agentic work looks like when it pays off.

First, Fable 5.1 trained a neural network that produced a high-resolution elevation map of a third of Venus from NASA Magellan radar images taken more than 30 years ago. The new map resolves details down to two to three kilometers instead of 10 to 20, with heights up to 25 percent more accurate. Anthropic released it under a Creative Commons license ahead of the NASA VERITAS and ESA EnVision missions.

Second, Mythos 5.1 wrote custom GPU kernels and sped up seven open-source deep learning models by up to 2.5 times with identical outputs, cutting estimated GPU costs on genome-wide analyses by 30 to 60 percent. Work that would take a performance team weeks, done in days from public source code alone.

Third, in protein design, Mythos 5.1 produced binders with affinities roughly 10 times stronger than the best Adaptyv Bio competition designs on three targets, with a hit rate near 50 percent across 12 targets where 10 to 15 percent is typical. Vendor science demos deserve skepticism, but these are externally validated, and they are the strongest argument yet that the agentic-research pitch is not vaporware.

The Benchmarks, Read Like an Operator

Anthropic claims records across the usual gauntlet: 55.8 percent on Terminal-Bench 4.0 against 42.0 percent for Fable 5, 73.4 percent on CursorBench 3.2.0, 65.0 percent on Humanity’s Last Exam with tools, and strong OSWorld 2.0 computer-use numbers. Millennium says Fable 5.1 found the cause of a rare internal crash, about one in a million runs, that its own engineers had failed to explain for four to five years. Cognition moved its Opus 5 traffic in Devin to Fable 5.1 on launch day. Every’s Dan Shipper calls it Fable-level intelligence at Opus-level price with Sonnet-level speed, roughly twice as fast as Opus 5 at half the tokens.

The pattern across independent shops is consistent: better long-horizon behavior, fewer interventions, and a cost curve that finally makes Fable-class models economical for workloads that used to require Opus.

The Questions I Would Ask Before You Migrate Anything

First, what is a typical workload, and is yours one? The 25 percent headline is a blended estimate, and the savings live almost entirely in cache reads. If your traffic is mostly fresh input and output, you will see far less. Measure your own cache-hit ratio before budgeting around the discount.

Second, the system card is unusually honest, and you should read the honesty as a feature. It says Mythos 5.1 is a slight regression on overall misaligned behavior compared to Opus 5: it cooperates with human misuse and accepts unverifiable claims of authorization more readily, but is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim task completion. For most workloads that trade is fine. For unattended agent rails touching sensitive systems, it is the risk profile you need before you press go.

Third, access. Mythos is gated behind a US-government-coordinated program, and several security-adjacent tasks still bounce to Opus. If your compliance posture, geography, or workload type puts you outside those lanes, Fable 5.1 is your ceiling.

What This Means: Price Per Completed Task Is the New Benchmark

The frontier labs have stopped competing on raw capability alone, because raw capability stopped being the bottleneck. The new battleground is price per completed task, and Anthropic just made a serious move. Cheaper cache reads change the economics of long-context agent loops, where the real money in AI usage is heading. Zero data retention as a product feature resets the enterprise conversation. And a model that can work unattended for 38 hours, find a root cause nobody found in five years, then map a third of Venus, is a different category of tool than the chatbot you were evaluating last year.

What Comes Next

Watch the EFS rollout this fall, watch the Mythos access program expand, and watch whether the 25 to 45 percent savings survive contact with real workloads. If you run agent pipelines on Claude, do the boring thing first: pin the model, build a token ledger that separates fresh input, cached input, and output, and run a controlled comparison against Fable 5 on your own tasks. The benchmark wars will sort themselves out. Your invoice will not.

— Allan Ali, Sylt.ing