Claude Fable 5: Anthropic's First Public Mythos Model

A Mythos-Class Model, Finally in the Open

For most of 2026, Anthropic's most powerful technology has been something the public could read about but not touch. Mythos-class models, the tier that sits a rung above the Opus line, were introduced in April with Claude Mythos Preview and locked behind Project Glasswing, a restricted cybersecurity initiative run with a handful of approved organizations. The reason given was blunt: these models are good enough at finding and exploiting software vulnerabilities that releasing them widely without new guardrails was considered too dangerous.

On June 9, 2026, that changed. Anthropic launched Claude Fable 5, the first Mythos-class model made generally available, alongside Claude Mythos 5, an unrestricted sibling that stays inside Glasswing. Fable 5 and Mythos 5 are the same underlying model. The only thing separating them is a set of safety classifiers bolted onto Fable that intercept queries in a few high-risk domains. That single difference is why they carry different names, and it is the entire story of how Anthropic squared "most capable model we have ever shipped broadly" with "we just warned the industry that AI is advancing dangerously fast."

This post covers Fable 5 end to end: where it sits in the lineup, what it is measurably good at, how the safeguards and fallback behavior actually work, the API changes developers need to plan for, the full cost picture, the staged subscription rollout and what to do during it, and how it stacks up against Opus 4.8 and the rest of the frontier field.

What Claude Fable 5 Actually Is

The cleanest way to understand Fable 5 is by its place in Anthropic's tier structure. Below it sit the familiar production models: Haiku 4.5 for speed and volume, Sonnet 4.6 for balanced work, and Opus 4.8 as the previous flagship. Above all of them is the Mythos class, a designation Anthropic created specifically for models whose capabilities cross into territory the company considers genuinely risky.

There are now three Mythos-class models, and the distinction between them is purely about access and safeguards rather than capability:

Claude Mythos Preview — the original, released in April through Project Glasswing. Still the strongest cyber model in some narrow respects, and still restricted.
Claude Mythos 5 — the new unrestricted model, available only to approved Glasswing partners and, soon, a small set of vetted biology researchers. No safety classifiers.
Claude Fable 5 — the same model as Mythos 5, with safety classifiers added so it can be released to everyone.

The naming is deliberate. "Fable" comes from the Latin fabula, meaning roughly "that which is told," a cousin of the Greek mythos. The two names exist to mark the same model in two states: Mythos is the raw capability, Fable is that capability wrapped in safeguards for public use. If you are a normal developer, business user, or subscriber, Fable 5 is the version you get, and for the overwhelming majority of work it behaves identically to the unrestricted model.

Fable 5 is also the opening model of Anthropic's fifth generation. Its headline pitch is not single-turn cleverness but endurance: the ability to sustain coherent, useful work across very long, complex, asynchronous tasks that earlier models lost the thread on. Anthropic's framing throughout the launch was consistent on this point. The longer and more involved the task, the wider Fable 5's lead over Opus 4.8 grows.

Benchmark and Capability Breakdown

Anthropic describes Fable 5 as state-of-the-art on nearly every capability benchmark it tested, with particular strength in software engineering, knowledge work, vision, scientific research, and long-running agentic tasks. The interesting detail is that the previous Mythos Preview still edges it out on a few cyber-specific tests, which makes sense given Fable's classifiers are designed to actively block progress on exactly those tasks. For everything else, Fable 5 is the strongest generally available Claude to date.

Software Engineering

This is where the long-horizon story lands hardest. During early testing, Stripe reported that Fable 5 collapsed work measured in months down to days. The standout example: a codebase-wide migration in a roughly 50-million-line Ruby codebase that the model completed in a single day, work Anthropic estimates would have taken a full engineering team more than two months by hand.

Fable 5 is also more token-efficient than prior Claude models, which matters a great deal when you are paying frontier prices per token. On Cognition's FrontierCode evaluation, which checks whether a model can solve hard coding problems while meeting the standards of a real production codebase, Fable 5 scored highest among frontier models even at medium effort rather than maximum effort. Early partners were unusually direct about it. Cursor called it the state-of-the-art model on their CursorBench evaluation and credited it with opening up a class of long-horizon problems earlier models could not reach. GitHub's product leadership described it taking on complex, multi-step coding tasks with a level of autonomy that exceeded their previous benchmarks. Cognition reported it as the highest-scoring model on their frontier coding eval, noting it generalizes to unfamiliar tools out of the box.

Knowledge Work

Outside of code, Fable 5's gains show up most clearly in dense analytical work. On Hebbia's finance benchmark for senior-level reasoning, it posted the highest score of any model, with the biggest improvements in document-based reasoning, chart and table interpretation, and multi-step problem solving. Trading firm IMC reported it cleared their trading-analysis evaluations almost across the board, covering factual lookup, conceptual reasoning, root-cause analysis, and expected-value calculation. A legal tooling partner found that, in blind review, the model's contract redlines matched or beat their existing model every time. One analytics partner reported it as the first model to break 90% on their core benchmark of complex, long-running analytical tasks, a roughly ten-point jump over Opus.

Vision

Fable 5 is the new vision state-of-the-art for Anthropic. It can pull precise numbers out of detailed scientific figures and rebuild a working web app's source code from screenshots alone. The most telling demonstration is how little scaffolding it needs: earlier Claude models could not reliably play Pokémon FireRed even when handed elaborate helper tools, while Fable 5 finished the game using only raw screenshots and a minimal vision-only harness, with no maps or navigation aids.

Memory and Long Context

Fable 5 ships with a one-million-token context window by default and can hold focus across millions of tokens in a long-running task, improving its own outputs by referring back to its notes. In one illustrative test using the deck-builder Slay the Spire, giving the model persistent file-based memory improved its performance three times more than the same memory helped Opus 4.8, and it reached the game's final act three times as often. The point is not the game. It is that Fable 5 makes substantially better use of persistent memory in extended autonomous sessions, which is the core requirement for agents that run for hours or days.

Scientific Research

The scientific results were demonstrated mostly through Mythos 5 (the unrestricted twin), but the underlying capability is shared. Anthropic's internal protein-design experts reported accelerating parts of the drug-design process roughly tenfold, with the model autonomously choosing binding sites, selecting and running design tools, and recovering from its own failures, matching or beating skilled human operators on several targets. It is also Anthropic's first model to consistently produce novel scientific hypotheses that researchers rated as compelling, and it ran more than a week of largely autonomous genomics research that produced a custom model outperforming a recently published one despite being a hundred times smaller. These are the dual-use capabilities that explain why the safeguards exist.

The Safeguards and the Opus 4.8 Fallback

The defining engineering decision behind Fable 5 is that it does not refuse dangerous requests in the usual sense. Instead, when its classifiers flag a query, the response is quietly handled by Claude Opus 4.8, Anthropic's next-most-capable model. Users are told when this happens. The reasoning is sound: an Opus 4.8 answer is a far better experience than a flat refusal, and Opus is itself a highly capable model.

There are three covered domains:

Cybersecurity. Mythos-class models are exceptional at discovering and exploiting software vulnerabilities and at agentic hacking — chaining reconnaissance, discovery, lateral movement, and exploitation. The cyber classifiers are tuned to block both narrow exploit work and broader offensive cyber tasks. Anthropic reports that in a blocking configuration, Fable makes essentially no progress on these evaluations.
Biology and chemistry. Anthropic previously blocked only a narrow band of bioweapons-related queries, but given the model's real ability to perform meaningful scientific tasks, it has widened this for now, falling back to Opus on most biology and chemistry requests. The company is explicit that this is broader than ideal and that it intends to narrow it as quickly as it safely can.
Distillation. Requests flagged as attempts to extract Fable 5's capabilities to train competing models also fall back to Opus 4.8.

The numbers worth remembering: the safeguards are tuned conservatively and trigger, on average, in under 5% of sessions. More than 95% of Fable sessions involve no fallback at all, and in those sessions Fable 5 behaves identically to the unrestricted Mythos 5. Anthropic is candid that the classifiers are stricter than it would like and will sometimes catch harmless requests, and it has committed to reducing those false positives after launch. It also ran an external bug bounty that produced no universal jailbreaks in over 1,000 hours of testing, though it notes the UK's AI Safety Institute made progress toward one in early testing — a useful reminder that no classifier is perfect.

Technical Specifications and API Behavior

For developers, Fable 5 introduces real behavioral changes, not just a new model string. The essentials:

Model IDs: claude-fable-5 for the public model, claude-mythos-5 for the restricted one.
Context window: 1M tokens by default.
Max output: up to 128k tokens per request.
Pricing: $10 per million input tokens, $50 per million output tokens.

A basic call looks like any other Claude request, with the new model string:

python

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=8000,
    messages=[
        {"role": "user", "content": "Migrate this module from callbacks to async/await."}
    ],
)

Two Messages API behaviors are specific to Fable 5 and Mythos 5 and will surprise teams porting from Opus:

Adaptive thinking is always on. Adaptive thinking is the only thinking mode available. There is no way to disable thinking — thinking: {"type": "disabled"} is not supported. Instead, you control reasoning depth with the effort parameter, which is also the main lever for managing cost.

Raw chain-of-thought is never returned. The model's raw reasoning is never exposed. The thinking.display setting decides what you get instead: "summarized" returns a readable summary of the reasoning, while "omitted" (the default) returns thinking blocks with an empty field. In multi-turn conversations on the same model, you pass thinking blocks back unchanged.

At launch, Fable 5 supports the effort parameter, task budgets (a beta feature gated behind a header), the memory tool, code execution, programmatic tool calling, context editing with tool-result clearing, compaction, and vision. That feature set is what makes it viable as the engine behind genuinely long-running agents rather than just a smarter chat model.

Refusals, Fallback, and Billing for Developers

Because Fable 5 can decline requests, any integration that calls it needs to plan for three changes: refusal handling, a fallback path, and the billing rules that go with both.

Refusals are successful responses, not errors. When Fable 5 declines a request, the Messages API returns stop_reason: "refusal" as a normal HTTP 200, and the response reports which classifier triggered. You handle it in your success path, not your error handler:

python

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=8000,
    messages=messages,
)

if response.stop_reason == "refusal":
    # Route this request to a fallback model instead of failing.
    handle_fallback(messages)

Fallback has three paths. A request Fable refuses can usually be served by another Claude model. Anthropic offers server-side fallback via a fallbacks parameter (in beta on the Claude API and the Claude Platform on AWS), client-side fallback through official SDK middleware in TypeScript, Python, Go, Java, and C#, and a fully manual path you build yourself.

Billing is designed not to punish refusals. You are not billed for a request that is refused before any output is generated. When you retry on another model, "fallback credit" refunds the prompt-cache cost of switching, so you do not pay that caching cost twice. In practice this means a defensive integration that retries on Opus 4.8 after a Fable refusal does not carry a hidden cost penalty for the retry's cached context.

One non-negotiable constraint: Fable 5 and Mythos 5 carry mandatory 30-day data retention and are not available under zero-data-retention agreements. Both are designated Covered Models. If your compliance posture currently depends on ZDR, that is the single biggest blocker to adopting Fable 5, and it needs to be settled before anything else.

Use Cases: Where Fable 5 Earns Its Price

Fable 5 is expensive, so the useful question is not "what can it do" but "what is it worth paying frontier prices for." The honest answer is: tasks where its long-horizon endurance and reasoning depth save far more human time than the token bill costs.

Large-scale code migrations and refactors. This is the flagship use case and the one with the clearest return. A codebase-wide migration that would occupy a team for weeks, compressed into a day, pays for an enormous number of tokens. The same logic applies to framework upgrades, dependency overhauls, and dead-code removal across sprawling repositories.

Autonomous, multi-step agents. Fable 5's combination of persistent memory, sustained focus across millions of tokens, programmatic tool calling, and self-validation at high effort is exactly the substrate that long-running agents need. This is where the model stops being a chat assistant and becomes an execution engine. Teams building agent-orchestration layers — open dashboards like OpenClaw Mission Control exist precisely to supervise fleets of agents running tasks like this — are the natural early adopters, because the model's value compounds the longer and more autonomous the run.

Senior-level analytical and knowledge work. Financial document analysis, due diligence, trading-strategy reasoning, contract redlining, and complex multi-source research are areas where partners reported Fable 5 matching or beating their existing tools. The model's strength at chart, table, and figure interpretation makes it especially strong on the document-heavy work that mid-tier models tend to fumble.

Vision-heavy reconstruction and extraction. Rebuilding application code from screenshots, extracting precise data from scientific figures, and understanding diagrams nested inside PDFs are all areas where Fable 5 sets the new bar. For teams digitizing legacy interfaces or processing visually dense documents, this is a meaningful capability jump.

Vibe coding and one-shot app building. Several partners noted that apps which took a hundred prompts a year ago, Fable 5 now produces in one. For rapid prototyping and for getting unstuck on a hard build, it is the model people described reaching for when they hit a wall.

Frontier research. With the appropriate trusted-access program, the scientific applications — protein design, hypothesis generation, genomics — point toward genuine research acceleration. For the general public this is gated behind the biology classifiers, but it signals where the capability is heading.

What to Do When Fable 5 Lands in Your Subscription

The subscription rollout is the most confusing part of this launch, and it is worth understanding precisely because the timeline matters.

On the API and consumption-based Enterprise plans, Fable 5 was fully available immediately on June 9. For subscription users, Anthropic chose a staged approach because it expects demand to be very high and hard to predict:

June 9 through June 22: Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. This is a free window — you get the most capable Claude ever shipped, bundled into your existing plan.
June 23 onward: Fable 5 is removed from those plans. Continued use requires usage credits, the pay-as-you-go credit mechanism for paid Claude plans. If capacity allows, Anthropic says it may extend the included window.
Later, when capacity allows: Anthropic intends to restore Fable 5 as a standard included part of subscription plans, as quickly as it can.

So the practical advice for a subscriber is straightforward. Use the free window deliberately. Between now and June 22, route your hardest, highest-leverage work to Fable 5 — the migration you have been dreading, the analysis that defeated lighter models, the agent run you wanted to attempt but worried would drift. This is the cheapest access to Mythos-class capability you will get for a while.

After June 23, switch your default back to Opus 4.8 for everyday work and reserve Fable 5 for the specific tasks where its endurance genuinely pays off, funding those through usage credits. Do not leave Fable 5 as your default model for routine prompting; at frontier prices, that is the fastest way to burn credits on work a cheaper model would have handled fine. Watch Anthropic's communications, since they have committed to announcing changes ahead of time, and re-evaluate once Fable 5 returns as a standard plan feature.

For teams, the calculus is similar but with more at stake. Use the free window to run a proper internal evaluation — give it your real workloads, measure where it beats Opus 4.8 and where it does not, and build a credible cost model before June 23 so you are not surprised by a usage-credit bill.

The Full Cost Breakdown

Fable 5 is the most expensive generally available flagship from a major lab, and pretending otherwise would do readers a disservice. Here is the complete picture.

API token pricing. $10 per million input tokens and $50 per million output tokens, identical for Fable 5 and Mythos 5. That is notable on two fronts: it is less than half the price of the earlier Claude Mythos Preview, and it is still roughly double Opus-class pricing on both input and output.

Prompt caching. The existing 90% input-token discount for prompt caching applies, which materially changes the math for any workload with a large, stable prefix — long system prompts, big retrieved documents, persistent agent context. A cached input prefix effectively drops toward $1 per million rather than $10, so caching is not optional for cost-sensitive Fable workloads; it is the primary lever.

Effort as a cost dial. Because adaptive thinking is always on, the effort parameter is your second major cost control. Lower effort means fewer reasoning tokens and a smaller bill; higher effort means deeper self-validation. Several partners found Fable 5 strong even at medium effort, so maximum effort is not always the right default.

Fallback economics. Refused requests are not billed before output, and fallback credit refunds the prompt-cache cost of switching models, so a well-built fallback path does not double-charge you for the same context.

Where Fable 5 sits versus the field. For context, current flagship API pricing in mid-2026 runs roughly:

Claude Fable 5 / Mythos 5: $10 / $50
Claude Opus 4.8: about $5 / $25
OpenAI GPT-5.5: in the premium band (its Pro tier runs dramatically higher on output)
Google Gemini 3.1 Pro: about $2 / $12 up to 200K context, doubling above that
xAI Grok 4.1: about $0.20 / $0.50

Read that table and the strategy writes itself. Fable 5 is not a value play and was never meant to be one. It is a capability play. You reach for it when the task is hard enough and long enough that no cheaper model finishes it correctly, and where a correct, autonomous completion is worth far more than the token cost.

Subscription versus API. For individuals, the cheapest path to Fable 5 right now is simply a paid subscription during the free window: Pro at $20/month, Max 5x at $100, Max 20x at $200. After June 23, subscription access shifts to usage credits, at which point heavy programmatic users are often better served by the API directly, while light users may find the subscription-plus-credits model simpler. Team plans (Standard seats around $20–25/month, Premium seats around $100–125/month, minimum five seats) and Enterprise (custom, consumption-based) follow the same staged pattern, with consumption-based Enterprise already having full immediate access.

How Fable 5 Compares to Opus 4.8

The most important comparison for existing Anthropic users is Fable 5 against Opus 4.8, because Opus is both the fallback model and the sensible default for everyday work.

Opus 4.8 is cheaper, has no blocking classifiers, supports zero-data-retention, and is more than capable enough for the large majority of tasks. Fable 5 is stronger across nearly every benchmark, with its advantage widening as tasks get longer and more complex, but it costs roughly twice as much, mandates 30-day retention, and will occasionally bounce a benign query back to Opus anyway.

A reasonable decision rule: default to Opus 4.8. Escalate to Fable 5 specifically for long-horizon agentic runs, large-scale migrations, the hardest analytical work, and vision-heavy reconstruction — the tasks where its endurance and depth convert directly into saved human hours. For short prompts, routine generation, and anything inside a covered domain that would just fall back anyway, Opus 4.8 is the better and cheaper choice. Against the wider field, Gemini 3.1 Pro and the Grok line are far cheaper and appropriate for cost-sensitive, high-volume work, while GPT-5.5 occupies a similar premium tier; Fable 5's differentiator is specifically sustained autonomous performance on hard, long tasks.

The 30-Day Data Retention Change

This deserves its own section because it is a genuine policy shift, not a footnote. For Fable 5, Mythos 5, and future models at similar or higher capability levels, Anthropic now requires 30-day data retention on all traffic, across both first-party and third-party surfaces. The company says it will not use this data to train models or for any non-safety purpose, and it has added privacy protections including logging all human access to the data and deleting it after 30 days in almost all cases.

The rationale is defense: retained data helps Anthropic detect and counter complex, novel attacks — including new jailbreaks and attacks that span many requests — and helps it identify and reduce the false positives in its classifiers. For most users this is invisible. For organizations whose compliance frameworks require zero data retention, it is a hard blocker, since these models are designated Covered Models and are explicitly not available under ZDR. Settle this question with your compliance and legal teams before building anything on Fable 5.

Mythos 5 and the Trusted Access Programs

Fable 5's restricted twin, Mythos 5, became available the same day to existing Claude Mythos Preview users, primarily the cyber-defense partners inside Project Glasswing, as a direct upgrade. Anthropic describes Mythos 5 as comparable to or somewhat stronger than Mythos Preview in most cases, at substantially lower cost. It plans to expand Mythos 5 access steadily, in consultation with the US government, and to stand up a more systematic trusted-access program that lets cybersecurity organizations apply.

A second program is coming for biology. Anthropic intends to open a trusted-access track that provides Fable 5 with the biology and chemistry safeguards removed (but cyber safeguards intact) to a small number of vetted life-science researchers, aimed at accelerating biomedical research and therapy discovery. Both programs reflect the same philosophy that runs through the whole launch: ship the broad, safeguarded model to everyone, and grant the unrestricted version only to vetted parties with a defensible reason to need it.

Practical Pitfalls to Watch

A few realities are worth internalizing before you commit to Fable 5:

Expect occasional false positives. The classifiers are deliberately over-tuned and will sometimes bounce harmless requests to Opus 4.8. If you work near cybersecurity, biology, or chemistry topics legitimately, you will hit this. Build for it rather than fighting it.
No zero-data-retention. Mandatory 30-day retention is a real constraint for regulated organizations.
Cost discipline is mandatory. At $10/$50, leaving Fable 5 as your default model is an expensive mistake. Cache aggressively, tune effort, and reserve the model for tasks that justify it.
Refusal handling is not optional for integrations. If you call Fable 5 programmatically without handling stop_reason: "refusal" and a fallback path, you will ship a brittle integration that breaks the first time a classifier fires.
The subscription situation will change. The free window closes June 23, and the usage-credit phase is explicitly temporary. Plan around announced changes rather than assuming today's access is permanent.

The Bigger Picture

The timing of this launch is hard to ignore. Anthropic released its most powerful broadly available model just days after publicly urging major AI labs to agree on a coordinated "brake pedal" for frontier development, warning that systems are advancing fast enough that recursive self-improvement may not be far off. It also arrives as the company prepares for a potential public listing. The tension between "this technology is dangerous enough to gate behind a government program" and "here it is for $20 a month" is real, and Anthropic's answer is the classifier-and-fallback architecture: release the capability, but intercept the narrow slice of uses that could cause serious harm.

Whether that architecture holds up under sustained adversarial pressure is the open question, and Anthropic itself acknowledges that completely preventing universal jailbreaks is likely impossible — the goal is to make them slow and costly enough to catch first. For now, Fable 5 is a genuine inflection point: the first time Mythos-class capability has been put in front of ordinary developers and businesses, with the safety machinery running quietly in the background.

Getting Started: A Practical Checklist

If you are deciding how to approach Fable 5, this is the short version:

Confirm your data-retention posture. If you require ZDR, Fable 5 is off the table until that changes; use Opus 4.8.
Use the free subscription window. Between now and June 22, route your hardest long-horizon work to Fable 5 at no extra cost.
Set a sensible default after June 23. Keep Opus 4.8 as your everyday model; escalate to Fable 5 only for migrations, long agentic runs, dense analysis, and vision reconstruction.
Build refusal handling and a fallback path into any programmatic integration before going to production.
Turn on prompt caching and tune effort to keep frontier pricing under control.
Evaluate on your real workloads. Measure where Fable 5 actually beats Opus 4.8 for you, and build a cost model before committing.

Claude Fable 5 is not the model you run for everything. It is the model you run when the task is hard enough that finishing it correctly and autonomously is worth paying for — and for the first time, that level of capability is something anyone can reach.