The most capable AI model Anthropic has ever made public arrived June 9, 2026 — and so did a more powerful version most people will never get to use.
On June 9, 2026, Anthropic officially released Claude Fable 5 to the general public and quietly handed Claude Mythos 5 to a tightly controlled circle of cybersecurity professionals and life sciences researchers. Both share identical model weights. The only thing separating them is a set of safety classifiers — and what those classifiers decide you're allowed to ask.
This is not a typical product launch. It is Anthropic's clearest statement yet on how it plans to handle AI that has grown powerful enough to help design drugs and build cyberweapons, depending on who is asking.
What Exactly Got Released?
Two models. One foundation.
Claude Fable 5 is the public-facing release. It is the most capable model Anthropic has ever made generally available, leading on nearly every major benchmark it was tested on. It handles long, complex, autonomous tasks better than any previous Claude model. It has advanced vision capabilities, stronger reasoning, and deeper memory management.
Claude Mythos 5 is the same model, but with the safety restrictions partially lifted. It is only accessible today through Project Glasswing — Anthropic's restricted program for vetted cyber defenders and infrastructure providers. A small group of life sciences researchers will gain access soon. A broader trusted access program is planned but not yet open.
The naming is not random. The word fable comes from the Latin fabula, and mythos from the Greek — both meaning "that which is told." The distinction in capability between the two is the safeguard layer, which is why they carry different names despite sharing the same underlying architecture.
Pricing for both: $10 per million input tokens and $50 per million output tokens. That is less than half what Claude Mythos Preview cost when it launched in April 2026.
How Fable 5 and Mythos 5 Compare to Previous Claude Models
Understanding this launch requires knowing where it sits in Anthropic's model history.
ModelClassPublic?Est. Input PriceKey DistinctionClaude Opus 4.6OpusYesLowerNo blocking cyber safeguardsClaude Opus 4.7OpusYesLowerUsed in jailbreak comparison testingClaude Opus 4.8OpusYesLowerCurrent fallback when Fable classifiers triggerClaude Mythos PreviewMythosNo (Glasswing only)>$20/M (est.)First Mythos-class; April 2026 launchClaude Fable 5MythosYes$10/MMost capable public model; classifiers activeClaude Mythos 5MythosNo (restricted)$10/MSame as Fable; some safeguards lifted
Mythos-class is a new internal tier Anthropic has created that sits above Opus in raw capability. Fable 5 is the first Mythos-class model released to the general public. Mythos 5 is the second deployment of this tier, accessible only to vetted partners.
When Fable 5's classifiers flag a request, the response is automatically handled by Claude Opus 4.8 instead — and users are told when this happens. According to Anthropic's early data, this fallback occurs in fewer than 5% of all sessions.
What Fable 5 Can Actually Do
The capability claims are backed by documented tests with real enterprise partners.
Software Engineering
Stripe gave Fable 5 access to a 50-million-line Ruby codebase and asked it to perform a full codebase-wide migration. Fable completed it in a single day. Stripe's engineers estimated the same task would take an entire team more than two months working manually. That is not incremental improvement — it is a qualitative shift in what one AI instance can accomplish in a working day.
On Cognition's FrontierCode benchmark — which tests whether a model can write code that passes difficult technical challenges while also meeting production-quality standards — Fable 5 scored highest among all frontier models, even when operating at medium computational effort.
Financial and Knowledge Work
Hebbia's Finance Benchmark evaluates senior-level analytical reasoning across documents, charts, and data tables. Fable 5 achieved the top score among all models tested, with the most notable gains in document-based reasoning, chart interpretation, and multi-step problem solving.
IMC, a quantitative trading and technology firm, ran Fable 5 through their internal trading-analysis evaluation suite — covering factual lookup, conceptual reasoning, root-cause analysis, and expected-value calculations. Their assessment: Fable 5 "aced" the evaluations nearly across the board.
Vision
Fable 5 can extract precise numerical values from complex scientific figures — a task that typically requires a human analyst to carefully read and interpret visual data. It can also reconstruct a web application's full source code from screenshots alone, without any access to the original codebase.
The most striking demonstration is Pokémon FireRed. Previous Claude models required a complex helper harness — extra tools providing maps, navigation aids, and game-state information — just to make progress in the game. Fable 5 completed FireRed from start to finish using raw game screenshots and nothing else. No harness. No aids. Vision only.
Memory and Long-Context Tasks
In testing with the deck-building game Slay the Spire, Anthropic gave Fable 5 access to persistent file-based memory and compared the improvement against Opus 4.8 in the same setup. Fable's performance improved three times more than Opus 4.8 did from the same memory advantage. Fable also reached the game's final act three times more frequently. This points to a qualitative difference in how effectively Fable can leverage extended context and persistent state — not just a quantitative lead.
What Mythos 5 Can Do That Fable 5 Cannot
The capabilities reserved for Mythos 5 are the ones that make the safety architecture necessary in the first place.
Drug Design — 10× Faster
Using Mythos 5 alongside protein design and bioinformatics tools, Anthropic's internal research team accelerated drug design workflows by approximately ten times compared to standard human-led processes.
In one documented evaluation, Mythos 5 was given a drug design task with no human assistance beyond high-level direction. The model selected binding sites, chose and executed the appropriate protein design tools, and recovered from errors along the way — replicating the full workflow of a skilled scientist. Across 14 protein targets tested, 9 yielded strong candidates currently under investigation. Targets span immune checkpoints, growth-factor and receptor signaling, neurodegeneration, muscle disease, and complex structural targets.
Scientific Hypotheses — Preferred 80% of the Time
In blinded evaluations where Anthropic's scientists compared Mythos 5 hypotheses against those from Opus-class models in molecular biology, researchers preferred Mythos 5's hypotheses roughly 80% of the time. Several have been advanced to experimental evaluation. One hypothesis — proposing a novel mechanism for an E. coli protein — was independently corroborated by a separate research lab working on the same problem without knowledge of Mythos 5's output.
Autonomous Genomics Research
Over more than a week of largely autonomous work, Mythos 5 assembled single-cell data covering millions of cells across 138 animal species. It then designed and trained its own custom machine learning model to identify cells performing the same biological function across distantly related organisms.
The result, with only high-level human guidance, outperformed a model recently published in the journal Science — while being 100 times smaller in size. Anthropic has said it intends to publish these findings.
Viral Shell Prediction
In a test of dual-use capability, Mythos 5 was asked to predict how genetic modifications would affect the assembly of adeno-associated virus (AAV) shells — a skill relevant to gene therapy development. The candidates were developed by Dyno Therapeutics. Mythos-class models, using biological reasoning alone, outperformed dedicated protein language models that were specifically built and trained for this type of task. AAVs are a legitimate tool in gene therapy, but the same capability raises concerns about the potential for misuse in designing harmful viruses.
The Safety Architecture: Three Classifiers, One Fallback
Fable 5's safety layer does not refuse flagged requests. It silently routes them to Claude Opus 4.8 and tells the user what happened. This design choice is intentional — a response from a highly capable model is a meaningfully better experience than an outright refusal.
The three domains covered:
1. Cybersecurity Mythos-class models can find and exploit software vulnerabilities. They can also conduct multi-step offensive cyber operations — reconnaissance, lateral movement, exploit deployment — without needing human guidance at each step. Fable 5's classifiers block the full scope of these capabilities.
In testing:
Fable 5 complied with zero harmful single-turn cyberattack requests
This held even when requests used any of 30 known public jailbreak techniques
An external partner confirmed Fable 5 had the most robust cyber safeguards of any model tested, including Opus 4.8 and 4.7
A bug bounty program yielded zero universal jailbreaks across 1,000+ hours of testing
The UK AI Safety Institute made early progress toward one during brief initial testing — Anthropic disclosed this openly
2. Biology and Chemistry Anthropic previously blocked only a narrow set of bioweapons-related queries. That approach is no longer considered sufficient. The AAV experiment above illustrates why: models that weren't explicitly trained for virology tasks are now outperforming dedicated specialist models on virology tasks. The same capability that accelerates gene therapy research could, in the wrong hands, inform the design of dangerous pathogens. For now, most biology and chemistry queries fall back to Opus 4.8 while more precise safeguards are developed.
3. Model Distillation Anthropic has detected large-scale organized attempts to query Claude in patterns designed to extract and replicate its capabilities — essentially training a competing model on Claude's outputs. These extracted models could then be released without safety infrastructure. Queries that match distillation patterns fall back to Opus 4.8.
The New Data Retention Policy
Starting with this release, Anthropic requires 30-day data retention for all traffic on Mythos-class models — including both its own products and third-party deployments. Key details:
Data is not used for model training
All human access to retained data is logged
Data is deleted after 30 days in almost all cases
Purpose: detect novel jailbreaks, identify multi-request attack patterns, and reduce classifier false positives over time
Project Glasswing and the Trusted Access Path
Project Glasswing launched in April 2026 as Anthropic's controlled deployment of its first Mythos-class model. The program was built around a simple premise: some AI capabilities are too sensitive for open release, but that does not mean they should go unused. Vetted organizations — cyber defenders, critical infrastructure providers — can access them under oversight.
With today's launch, all existing Glasswing partners can upgrade from Mythos Preview to Mythos 5. The model is comparable to or stronger than Mythos Preview on most tasks, at less than half the price.
The next phases:
Cybersecurity trusted access program: Anthropic plans a systematic application process for cybersecurity organizations
Biology trusted access program: A small group of life science researchers will gain access to Fable 5 with biology/chemistry safeguards removed (but cyber safeguards remaining)
Broader expansion: Both programs will grow progressively as safeguards improve
Availability and Subscription Rollout
Claude Fable 5 is available globally today via claude-fable-5 on the Claude API.
Access TypeAvailabilityClaude APIFully available todayConsumption-based EnterpriseFully available todayPro, Max, Team, seat-based EnterpriseIncluded free: June 9–22 onlyAfter June 23Usage credits requiredFuture subscription inclusionWhen capacity allows
Anthropic has flagged that demand will be exceptionally high and difficult to predict. The staged rollout for subscription plans is a capacity management decision, not a pricing decision.
Why This Moment Is Different
AI companies have launched powerful models before. What makes this launch structurally different is the explicit acknowledgment that the model being released publicly is not the most powerful version that exists — and that the gap is there on purpose.
The question Anthropic is implicitly answering is one the entire AI industry will eventually face: when a model becomes capable enough to cause serious harm at scale, what do you do? You can delay release indefinitely. You can release and hope for the best. Or you can build the infrastructure to release responsibly — classifiers, fallbacks, trusted programs, data retention policies, external red-teaming — and accept that the safeguards will be imperfect at launch.
Anthropic chose the third option. Whether the safeguards hold, and whether the trusted access programs scale without compromising their integrity, will define whether this approach becomes a template or a cautionary example.
What to Watch Next
How quickly Anthropic reduces false positives in the biology and chemistry classifiers
Whether the UK AI Safety Institute's progress on a universal jailbreak leads to a disclosed finding
Publication of the Mythos 5 autonomous genomics research
Timeline for the broader Mythos 5 trusted access program
Whether Fable 5 returns to standard subscription plans before or after June 23
Source: Anthropic Official Announcement — anthropic.com/news/claude-fable-5-mythos-5 — Published June 9, 2026
