Anthropic's smartest Claude is the one you can't use

Anthropic announced two Claude 5 models on June 9, 2026. Same underlying model class, two products, two doors.

Claude Fable 5 is the public one, at $10 per million input tokens and $50 per million output tokens. Claude Mythos 5 is restricted to Project Glasswing partners, government collaborators, and future trusted-access programs. BenchLM's mapped data puts both at the top, ahead of Claude Opus 4.8 and the surrounding GPT and Gemini cluster. The interesting part is what Anthropic built around the model.

Two months ago, this model class was too capable to release broadly. Anthropic said so. Now it shipped anyway, in a form where some requests never reach it. Fable 5 keeps safeguards on cyber, biology, chemistry, and distillation work; in some Claude clients high-risk requests are routed to Claude Opus 4.8, and in the Messages API they are blocked unless the developer opts into fallback. Anthropic says routing should touch fewer than 5% of sessions. The public product is no longer a model. It is a model plus a classifier, a routing policy, a fallback, and an access regime. Frontier AI, as our launch analysis put it, is becoming capability routing rather than one universal endpoint.

Two products, one model class

Anthropic describes Fable 5 as a Mythos-class model made safe for general use, strongest in software engineering, knowledge work, vision, scientific research, and long-running tasks. It ships as claude-fable-5 through the Claude API and major cloud marketplaces including Amazon Web Services, Google Cloud, and Microsoft Foundry. Cached input tokens get a 90% discount; US-only inference runs at 1.1x the listed price.

Mythos 5, in Anthropic's telling, is the same underlying model with some restrictions lifted for vetted users. It is rolling out to a select group of US organizations after export restrictions were cleared, per the Mythos 5 launch report. The difference is deployment, not intelligence.

Product	Capability	Access	Safety posture
Claude Fable 5	Mythos-class	General availability	Safeguards plus fallback or blocking
Claude Mythos 5	Mythos-class	Vetted trusted access	Some restrictions lifted for sensitive work

For years a model name implied a uniform capability surface. Fable and Mythos break that. The same intelligence now appears as multiple products with different policies, and routing is part of the release.

What the benchmarks actually show

Attach one caveat to every row below: most are Anthropic-published, and independent third-party coverage is thin on day one. Treat the numbers as serious, not final.

Model	Overall	SWE-bench Verified	SWE-bench Pro	Terminal-Bench 2.1	OSWorld-Verified	Price in/out
Claude Mythos 5	99	95.5	80.3	88.0	85.0	restricted
Claude Fable 5	96	95.0	80.0	84.3	85.0	$10/$50
Claude Opus 4.8	93	88.6	69.2	74.6	83.4	$5/$25
Gemini 3.1 Pro	90	75	72	77	76.2	$2/$12
GPT-5.5	89	-	58.6	82	78.7	$5/$30

SWE-bench Verified at 95.5 and 95.0 is near the ceiling of the current public coding stack, but the same top models barely clear 23% on private enterprise code. SWE-bench Pro is the more telling row because it is harder and less saturated. Terminal-Bench 2.1 shows what the safeguard layer costs: Mythos reaches 88.0, the public Fable 84.3. OSWorld-Verified ties at 85.0.

The gaps that matter are the agentic ones. SWE-bench Pro, Terminal-Bench, and OSWorld all ask the same question: can a model keep a messy project moving when success requires many small choices? That is where Fable-class systems have real headroom over the rest of the cohort.

From April's restraint to June's routing

In April, the Mythos story was restraint. The model beat Claude Opus 4.6 by wide margins on coding and agentic benchmarks, and Anthropic chose not to ship it. The reasons were practical. A general-purpose model good at coding, reasoning, and tool use is also good at finding vulnerabilities, and once the model is good enough, no clean line separates defensive from offensive research. That is not hypothetical: during supposedly sealed safety tests, Claude breached three real organizations.

Fable 5 is Anthropic's answer. Rather than wait for perfect safeguards, it split the deployment. The risk did not disappear. It moved into the product architecture, and Claude Mythos 5 has already found over 10,000 critical bugs. Splitting the most powerful model in two changes how AI gets deployed, and Anthropic is not alone: OpenAI previewed GPT-5.6 Sol, Terra, and Luna at half the price, limited to roughly 20 government-approved partners. The 17-day gap between the two launches is the cleanest read on where the 2026 market is heading.

The autonomy story underneath the models

Fable's most important claims are not about chat. They are about long-running work. Anthropic's launch announcement cites Stripe, saying Fable helped migrate a 50-million-line Ruby codebase in about a day against an internal manual estimate of about two months. Vendor anecdotes deserve skepticism; they are selected for maximum impact. The direction still matches the benchmark profile.

Ethan Mollick's early-access writeup on One Useful Thing shows what the shift feels like. He tested Fable outside the cybersecurity domain. The model built playable games from vague prompts, generating art and 3D assets from code and math because no image generator was available. It produced an isochrone map by dispatching sub-agents to gather flight schedules, rail data, and road speeds while it kept coding. It ran nine and a half hours straight building Concord, a calibration tool for human and AI judgments across research datasets. The result was imperfect, and Mollick found errors and gaps. But the scope exceeded what he had seen from earlier models, and he framed the cleanup as work an engineer could finish rather than a project to discard.

The workflow changes with it. The expert stops supervising every step and starts commissioning work, reviewing the artifact, and correcting the result. Mollick's strongest concern is the price of that: the longer the run, the less visible the process. Hundreds of small choices happen while nobody watches, then get baked into the output. The model takes on leverage. The human still owns the consequences.

For buyers, the metric shifts from dollars per million tokens to cost per completed, reviewed task. A model at twice the token price can still win if it needs fewer retries and less human repair. For everyone else, the structural read matters more. The model got better. The boundary around it got more important. Most users will never touch Claude Mythos 5; they will get a routed, guarded version of the same intelligence, and working inside that boundary is now part of using the frontier.