Claude Fable 5.1 and Mythos 5.1: Pricing and Access

Anthropic shipped a paired model release on Monday, September 1, 2026: Claude Fable 5.1, generally available to all API customers, and Claude Mythos 5.1, restricted to a vetted set of cybersecurity and life-sciences partners. The two are, in Anthropic’s framing, the same underlying model carrying different safeguards — a structure the company introduced with the Fable 5 generation and is now extending into a point release aimed squarely at agentic and security-adjacent work.

The headline for most developers is not a new capability but a price cut. Fable 5.1 keeps the same headline token pricing as Fable 5 — $10 per million input tokens and $50 per million output tokens — but slashes the cost of cached input reads by 75%, from $1.00 to $0.25 per million tokens. For workloads that lean heavily on prompt caching, the effective savings are far larger than the sticker price suggests.

Why the cache cut matters

Prompt caching stores the unchanging front of a request — system prompts, tool definitions, retrieved documents, prior conversation turns — so the model does not re-ingest it on every call. Agent loops are the extreme case: a coding agent working through a repository may replay a large, stable context hundreds of times while changing only the last few hundred tokens. In that regime, cache reads dominate the bill.

By cutting cache reads to a quarter of their prior cost, Anthropic estimates Fable 5.1 will run about 25% cheaper than Fable 5 for typical workloads and, for highly agentic work, deliver savings of up to roughly 45%. That is a meaningful move in a market where the cost per completed task — not the cost per token — increasingly decides which model teams deploy at scale. It also lands in the middle of a broader AI price war, as frontier labs compete on the economics of long-running agents rather than raw benchmark scores.

The pricing structure is a tell about where Anthropic thinks the money is. Input and output rates held flat while the caching lever moved, a change that disproportionately benefits exactly the long-horizon agent workloads the Fable line was built for.

Fable versus Mythos: same model, different gates

The split between the two models is not about capability — it is about safeguards. Fable ships with classifiers that screen sensitive requests, routing certain cybersecurity and biology queries away from the frontier model and toward a more conservative fallback. Mythos is the same model with those classifiers lifted in specific domains for partners who have been vetted through Anthropic’s trusted-access program.

Access to Mythos 5.1 is limited to participants in Anthropic’s Project Glasswing, admitted through either the Cyber Verification Program (CVP) or the Life Sciences Verification Program (LSVP). The logic is the dual-use problem in miniature: the same model that helps a defender find and patch a vulnerability can help an attacker discover and weaponize one, and the same model that accelerates drug discovery touches knowledge with obvious misuse potential. Anthropic’s answer is to keep the unrestricted version behind vetting and monitoring while shipping the classifier-gated version to everyone else.

That approach mirrors the industry’s converging playbook on frontier cyber capability. OpenAI took a comparable path with its next flagship, restricting Astra’s most advanced cyber skills to vetted testers after classifying it at the “Critical” threshold, and the same trusted-access philosophy runs through the 116-company AI cyber-defense letter that Anthropic co-signed in August. The premise across all of it is that powerful security capability should reach defenders first, under supervision, rather than ship unconditionally or not at all.

The benchmarks tell on the safeguards

The capability gap between the two variants shows up in the numbers — and it runs the direction the safeguard design predicts. On Terminal-Bench 4.0, a test of agents operating in a real command-line environment, Mythos 5.1 scores 60.9% against Fable 5.1’s 55.8%. The unrestricted model does better precisely because the restricted classifiers on Fable intercept or reroute some of the security-flavored tasks that make up part of the benchmark.

That five-point spread is a concrete illustration of what the safeguards cost in raw agentic performance, and why vetted partners doing legitimate offensive-security and life-sciences research want the ungated version. For general software work, the gap narrows: the Fable 5 generation posted 95.0% on SWE-bench Verified and 80.0% on the harder SWE-bench Pro, and the 5.1 point release is tuned to hold those coding gains while improving tool use and long-horizon reliability.

Anthropic also moved on transparency tooling alongside the model. The company opened a private preview of a watermark-detection API, part of a broader push to let enterprises and regulators verify AI-generated content — a nod to compliance regimes taking shape under frameworks like the EU AI Act. The release notes frame the point update as an incremental step: better economics, steadier agentic behavior, and a cleaner separation between the gated and ungated safeguard tiers.

How to get it, and what to weigh

Fable 5.1 is available now through Anthropic’s API, the Claude Developer Platform, and the major cloud model catalogs, with the same request shape as Fable 5 — for most teams the migration is a model-string swap, not a rewrite. Because pricing on input and output tokens is unchanged, there is no cost penalty to moving up, and the cache-read reduction accrues automatically to anyone already using prompt caching. Anthropic is positioning 5.1 as a drop-in upgrade rather than a model that demands re-tuning.

The one real decision point is caching discipline. The savings are only as large as a workload’s cache-hit rate, which means the teams that benefit most are those that have already structured their prompts to keep the stable prefix — system instructions, tool schemas, retrieved context — ahead of the variable tail. Shops that rebuild their context on every call, or that shuffle the order of cached and uncached content, will see a smaller share of the advertised 45%. In practice, the release rewards engineering that was already good hygiene: stable, front-loaded context and tight, incremental turns.

For organizations weighing Mythos, the calculus is different and slower. Admission runs through vetting, monitoring commitments, and use-case review under the CVP or LSVP tracks, so it is a procurement-and-compliance decision as much as a technical one. The payoff — a measurably more capable ungated model on security and life-sciences tasks — is real, but it comes with obligations that most general developers neither need nor want.

What it means

This is a release about deployment economics and safety posture, not a leap in raw intelligence — and that is the point. Anthropic is competing for the workloads that actually run in production: agents that loop for hours, replay large contexts, and increasingly touch domains where an unguarded model is a liability.

Who wins. Teams running agent-heavy pipelines. A 75% cut to cache reads meaningfully lowers the cost of the exact usage pattern — stable context, many iterations — that defines coding agents and long-horizon automation. For those buyers, Fable 5.1 is a straightforward cost reduction with no capability trade-off. Vetted security and life-sciences partners also win, gaining a more capable ungated model through Project Glasswing.

The tension to watch. The two-tier model is now the industry norm, and it concentrates the most powerful capability in the hands of a vetted few while everyone else runs a classifier-gated version. That is a defensible answer to the dual-use problem — the same one the sector reached for after incidents like the Hugging Face agent breach — but it also formalizes a capability gap between approved partners and the general public, and puts the labs in the position of deciding who clears the bar.

What to watch next. Three signals over the coming weeks: whether the cache-read cut pressures rivals into matching moves on caching economics; how quickly the watermark-detection API graduates from private preview as compliance deadlines approach; and whether the Terminal-Bench gap between Fable and Mythos widens or narrows as Anthropic tunes its classifiers. The models are converging on a shape where price and safeguards, not headline benchmarks, are the real product.