Claude Fable 5 and Mythos 5: The Definitive Deep Dive, and What We Are Seeing Inside Intueo

On June 9, 2026, Anthropic launched Claude Fable 5 and Claude Mythos 5, and with them a new tier of AI capability the company calls "Mythos-class," sitting above the familiar Opus class. Anthropic's own framing is unusually direct: Fable 5's capabilities exceed those of any model it has ever made generally available, it is state-of-the-art on nearly all tested benchmarks, and the longer and more complex the task, the larger its lead over every other Claude model. [1]

That last sentence is the one that matters most to us, because long, complex tasks are exactly what the Intueo platform runs all day. We have been putting Fable 5 to work inside our own stack since launch, and we can say it plainly: the claims hold up. What Anthropic describes on paper is what we are seeing in production-style agentic runs on our platform.

This post is our complete breakdown of the release. We cover what shipped, the capability evidence across software engineering, knowledge work, vision, memory, and science, the new safety architecture, the pricing and rollout, the broader industry context that makes this moment significant, and a detailed look at how Fable 5 performs inside Intueo.

Click to expand

The Short Version, If You Only Have Two Minutes

There are two models, and they share the same underlying brain. The difference between them is safety configuration, not capability. [1]

• Claude Fable 5 is the version made safe for general use. When a request touches sensitive areas, specifically cybersecurity, biology and chemistry, or model distillation, dedicated classifier systems hand the response to the next-most-capable model, Claude Opus 4.8, instead. Anthropic reports this fallback triggers in fewer than 5% of sessions, which means for more than 95% of sessions, Fable 5's performance is effectively identical to Mythos 5. [1]
• Claude Mythos 5 is the same model with the cyber safeguards lifted, available initially to a vetted group of cyber defenders and critical infrastructure providers through Project Glasswing, in collaboration with the US government. Anthropic says it has the strongest cybersecurity capabilities of any model in the world. [1]

Both models are priced at $10 per million input tokens and $50 per million output tokens, less than half the price of the earlier Claude Mythos Preview. Fable 5 is available today through the Claude API as `claude-fable-5`. [1]

Even the names tell the story. As Anthropic notes, "Fable" comes from the Latin fabula, "that which is told," a cousin of the Greek mythos. Same story, two tellings: one with guardrails for everyone, one with fewer guardrails for the defenders who need them. [1]

Why "Mythos-Class" Is a Real Tier, Not Marketing

It is worth pausing on the class structure, because it explains a lot about how this launch was handled.

In April 2026, Anthropic released the first Mythos-class model, Claude Mythos Preview, exclusively through Project Glasswing to a limited group of cyber defenders and critical software infrastructure providers. The reasoning was straightforward: the model had crossed a capability threshold where open release without new safeguards would be irresponsible, particularly in offensive cybersecurity. At the time, Anthropic said it hoped to eventually bring Mythos-level capabilities to all users once safeguards were strong enough. [1]

That is exactly what this launch is. Fable 5 is the fulfillment of that promise: the full Mythos-class capability, wrapped in a new safeguard architecture robust enough for general release. The two months between Mythos Preview and this launch were spent hardening classifiers, running red teams, and building the fallback system. We will dig into all of that below, because the safety engineering here is genuinely interesting in its own right.

The Trend Line Behind the Launch: Task Horizons Are Compounding

Before diving into the individual claims, it helps to place this release on the curve the research community has been tracking for years, because Fable 5 is not an isolated event. It is the latest data point on one of the most consequential trend lines in technology.

METR, the independent research organization that measures AI autonomy, popularized the "time horizon" metric: the length of task, measured in how long it would take a skilled human, that a model can complete at a given success rate. Their landmark finding is that frontier model time horizons have been doubling roughly every seven months since 2019, and more recent analyses suggest the doubling pace has accelerated since 2023, with estimates ranging from roughly 90 to 130 days depending on methodology. [2]

That curve is exactly why Anthropic's emphasis on "the longer the task, the larger the lead" matters so much. A model that extends the reliable task horizon does not just get a better benchmark score; it crosses thresholds where entire categories of work flip from "human project with AI assistance" to "agent project with human oversight." Stripe's reported one-day migration of a 50-million-line codebase, which we cover below, is what one of those threshold crossings looks like in the wild. If METR's doubling pattern holds even approximately, the strange results in this launch are a preview of the ordinary results of the next one.

There is a second trend line worth holding in mind: the gap between AI enthusiasm and AI deployment. McKinsey's 2026 research finds that while 88% of organizations now use AI in at least one business function, only about 23% are scaling agentic systems anywhere in the business, and no more than 10% are scaling them within any single function. [3] Gartner's CIO data tells the same story from another angle: roughly 17% actual agent deployment, and a forecast that over 40% of agentic AI projects will be cancelled by 2027 due to unclear ROI and rising costs. [4]

Read those two trend lines together and the picture sharpens. Capability is compounding fast, but most enterprises are still failing to convert it into production value. The bottleneck is not model intelligence; it is reliability over time, integration, memory, and operational trust, which is precisely the territory this release targets and precisely the gap platforms like ours exist to close.

The Capability Evidence, Area by Area

Anthropic backed the launch with an unusually concrete set of capability demonstrations, many of them from third parties with early access. Independent third-party validation matters here: benchmark claims from a model's maker always deserve scrutiny, which is why the most convincing evidence in this release comes from outside firms, Stripe, Cognition, Hebbia, IMC, Cursor, and GitHub among them, who tested the model against their own private evaluations before launch. Here is the full picture, area by area, with our read on what each one actually means.

Click to expand

Software Engineering: Months Into Days

The headline engineering result came from Stripe. During early testing, Stripe reported that Fable 5 compressed months of engineering into days: in a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a single day that would otherwise have taken an entire team more than two months by hand. [1]

Stop and consider what that requires. A codebase-wide migration in a repository that size is not a clever one-shot answer. It is thousands of coordinated edits, each one needing to respect conventions, avoid breaking call sites, and stay consistent with decisions made hours earlier in the run. It is the purest possible test of long-horizon coherence, and it is precisely the kind of task where every previous generation of models eventually lost the thread.

Fable 5 is also more token-efficient than past Claude models. On Cognition's FrontierCode evaluation, which tests whether models can pass difficult coding tasks while meeting the standards of high-quality production codebases, Fable 5 scores highest among frontier models, even at medium effort. [1] Cognition separately reported that Fable 5 is the highest-scoring model on FrontierBench, their frontier coding eval, and that it "excels at long-horizon reasoning and generalizes to unfamiliar tools out of the box." [1]

The early-access chorus from the developer tools world was consistent:

• Cursor called Fable 5 "the state of the art model on CursorBench," saying it has "opened up a class of long-horizon problems that were out of reach for earlier models." [1]
• GitHub described it as "a real step forward," noting it took on complex, long-horizon coding tasks "with a level of autonomy and reliability that exceeded previous benchmarks," and pointed at a future where developers hand increasingly ambitious work to agents and trust the results. [1]
• Another partner reported that apps that took a hundred prompts a year ago, Fable 5 "now one-shots." [1]

Knowledge Work: Senior-Level Reasoning

The knowledge work results are just as notable, because they come from evaluators whose entire business is measuring exactly this.

On Hebbia's Finance Benchmark for senior-level reasoning, Fable 5 posted the highest score of any model, with substantial gains in document-based reasoning, chart and table interpretation, and problem solving. Hebbia called it "the strongest finance-first model we've tested." [1]

Trading firm IMC reported that Fable 5 "aced their trading-analysis evaluations nearly across the board," spanning factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis. [1] One analytics partner said Fable 5 is "the first to break 90% on our core analytics benchmark of complex, long-running analytical tasks," a 10-point jump over Opus, showing "strong judgment and attention to nuance" on the hardest questions. [1]

In the legal domain, one early customer ran blind reviews and found Fable 5's contract redlines "matched or beat our current model every time." [1] A physics research group reported it was the strongest model they had tested on frontier physics research while using a third of the reasoning tokens, getting in 36 hours nearly to where a competing frontier model landed after four days. [1]

Click to expand

Vision: Less Scaffolding, More Capability

Fable 5 is the new state-of-the-art model for vision tasks. It can extract precise numbers from detailed scientific figures and perform complex vision-based work like rebuilding a web app's source code from screenshots alone. [1]

The most memorable demonstration is also the most playful: Pokémon FireRed. Previous Claude models struggled to play the game even with elaborate helper harnesses providing maps, navigation aids, and extra game-state information. Fable 5 beat FireRed start to finish with a minimal, vision-only harness, working from raw game screenshots with no maps or extra game state at all. [1]

That sounds like a toy, but the implication is serious. "Needs less scaffolding" is one of the most valuable properties a model can have in production. Every piece of scaffolding you build around a model is code you must maintain, a failure surface you must monitor, and an assumption that breaks when the world changes. A model that can operate from raw visual input alone collapses entire categories of glue work.

Memory and Long-Context: The Three-Times Result

This is the section we care most about, and we will return to it when we discuss our own platform experience.

Anthropic reports that Fable 5 stays focused across millions of tokens in long-running tasks and improves its outputs using its own notes. In a controlled demonstration, they had the model play the deck-building game Slay the Spire. When given access to persistent, file-based memory, Fable 5's performance improved three times more than Opus 4.8's did under the same conditions, and Fable reached the game's final act three times more often. [1]

Click to expand

Why a deck-building game? Because it is a near-perfect proxy for long-horizon agentic work. Success requires accumulating a strategy over many decisions, remembering why earlier choices were made, adapting when circumstances change, and not abandoning a coherent plan halfway through. Swap "deck" for "customer workflow" and you have described the daily life of a production AI agent.

The deeper finding inside that result: it is not only that Fable 5 has memory, it is that Fable 5 knows how to use memory. Giving a model a notebook only helps if the model writes useful notes and actually consults them. The three-times multiplier over Opus 4.8 suggests the model has internalized note-taking as a working strategy, not a checkbox feature.

Science: Where Mythos 5 Gets Astonishing

The scientific results, mostly demonstrated with Mythos 5, are arguably the most consequential part of the announcement, even if they got fewer headlines than the coding numbers.

Click to expand

Drug design, ten times faster. Anthropic's internal protein design experts used Mythos 5 to accelerate aspects of the drug design process by around ten times. In one study, Mythos 5, equipped with protein design and bioinformatics tools but no human assistance, matched or beat skilled human operators, executing every task a scientist normally performs: choosing binding sites, selecting and running protein design tools, and recovering from failures along the way. Nine of the 14 protein targets in the study yielded strong candidates for drug design that Anthropic is now investigating, with targets spanning immune checkpoints, growth-factor and receptor signaling, neurodegeneration, and muscle disease. [1]

Novel scientific hypotheses. Mythos 5 is Anthropic's first model to consistently produce novel, compelling scientific hypotheses. In blinded head-to-head comparisons against Opus-class models, Anthropic's scientists preferred Mythos's molecular biology hypotheses roughly 80% of the time, and several have advanced to experimental evaluation. One Mythos hypothesis, a novel mechanism for an E. coli protein, was independently corroborated by a lab working on the same problem. [1] Read that again: a model proposed a new biological mechanism, and independent human science confirmed it.

Autonomous genomics research. Over more than a week of largely autonomous work, Mythos 5 assembled single-cell data for millions of cells spanning 138 animal species, then designed and trained a custom machine learning model to identify cells performing the same role across even distantly related organisms. With only high-level human input, the model Mythos 5 trained outperformed a recent model published in the journal Science, despite being 100 times smaller. Anthropic intends to publish these results in the coming months. [1]

One early-access research partner summed up the shift: Fable 5 "works at senior research scientist grade — picking directions, allocating resources, killing its incorrect beliefs, and producing novel first-principles outputs." [1] A public demonstration in the same spirit: Fable 5 built a simulation of the solar system, deriving the planets' orbital motion from physics first principles and using it to predict solar eclipses. [1]

Alignment

On safety behavior, Anthropic's automated alignment assessment found Mythos 5's level of misaligned behavior, including deception and cooperation with misuse, was low and similar to Opus 4.8. Since Fable 5 is the same underlying model, its alignment profile is similar. The full assessment, along with a detailed suite of safety and capability tests, is published in the model's system card. [1]

The Evidence at a Glance

The Safeguards: How You Ship a Model This Capable

Here is the part of the announcement that we think deserves far more attention than it will get, because it is a preview of how every frontier release will work from now on.

Click to expand

The core problem is dual use. The same capabilities that let Mythos 5 help defenders find vulnerabilities in critical software would, without safeguards, let attackers find them too. The same biological reasoning that accelerates gene therapy research could, in the wrong hands, assist with dangerous virus design. Anthropic demonstrated this duality explicitly: in a test on predicting how genetic modifications affect viral shell assembly in adeno-associated viruses (a key component for delivering gene therapies, on unpublished candidates from Dyno Therapeutics), Mythos-class models outperformed specialized protein language models using biological reasoning alone, without ever being trained for the task. [1] That is wonderful for medicine and sobering for biosecurity, simultaneously.

The Classifier-Plus-Fallback Design

Fable 5 ships with a new set of classifiers: separate AI systems that watch for potential misuse, including jailbreak attempts, across three areas: [1]

1. Cybersecurity. Mythos-class models excel at discovering and exploiting software vulnerabilities, and at agentic hacking more broadly, covering reconnaissance, discovery, and lateral movement, not just exploit-finding. The classifiers are designed to block progress on both exploitation and offensive cyber tasks in the broader sense. [1]
2. Biology and chemistry. Coverage here is deliberately broad for now, falling back on most biology and chemistry requests, because Anthropic prioritized releasing safely and quickly over tuning precision. They explicitly state they want to narrow these safeguards so that false positives do not block legitimate science. [1]
3. Distillation. Anthropic has previously identified large-scale attempts to extract Claude's capabilities to train competing models in authoritarian countries. Flagged distillation attempts also trigger the fallback. [1]

The elegant part is what happens on a trigger. Instead of a flat refusal, the response is automatically handled by Claude Opus 4.8, and the user is told when this happens. A fallback to a still-excellent model is a far better experience than a brick wall, and with more than 95% of sessions involving no fallback at all, the typical user simply experiences Mythos-level capability. [1]

How Hard Is It to Break?

Anthropic published an unusual amount of red-team detail: [1]

• An external bug bounty produced no universal jailbreaks in over 1,000 hours of testing.
• External red-teaming organizations failed to find universal jailbreaks on long-form agentic tasks, although the UK AI Safety Institute made progress toward one within a brief initial testing window, which Anthropic disclosed openly.
• In an internal evaluation where an automated red-teamer attacked across 400 turns, restarting and rewinding when blocked, the production system held on complex, realistic tasks.
• One external partner found Fable 5's safeguards against harmful cyber queries were the most robust of any model tested, complying with zero harmful single-turn requests about cyberattack planning, exploit development, or defense evasion, whether or not the request used any of 30 public jailbreak techniques.

Anthropic is candid that completely preventing universal jailbreaks is likely impossible. Their stated goal is to make remaining jailbreaks slow and costly enough to detect and stop before they are used at scale. [1] That is the honest version of the claim, and we respect it more than a pretense of perfection.

The Data Retention Tradeoff

One change businesses should note: all traffic on Mythos-class models now requires 30-day retention, on both first- and third-party surfaces. Anthropic commits that this data will not train new Claude models or serve any non-safety purpose, with new privacy protections including logging of all human access and deletion after 30 days in almost all cases. The purpose is defending against complex, multi-request attacks and reducing false positives. [1] If your compliance posture depends on zero retention, factor this in before adopting Mythos-class models for sensitive workloads.

Pricing, Availability, and the Rollout Schedule

The economics are arguably the most disruptive part of the launch. Frontier-exceeding capability at $10 per million input tokens and $50 per million output tokens, less than half the Mythos Preview price, resets expectations for what top-tier intelligence costs. [1]

Anthropic is being transparent that demand will be very high and hard to predict, so subscription access rolls out in stages, with changes communicated ahead of time. [1] The forward plan for Mythos 5 includes a systematic trusted access program for cybersecurity organizations, developed in consultation with the US government, plus a parallel trusted access program for biology that will give selected life science researchers access with biology and chemistry safeguards removed while keeping cyber safeguards in place. [1]

What We Are Seeing Inside Intueo

Now for the part you cannot get from the announcement: what this model is actually like to operate.

We have been running Fable 5 inside the Intueo platform on the long-horizon, multi-step agentic workloads our platform exists for, and our experience confirms what Anthropic published. The long-horizon steadiness is real. On extended workflows that require an agent to hold context, plan across many steps, call tools, and follow through without losing the thread, Fable 5 stays sharp deep into runs where weaker models drift, re-derive their goals, or quietly abandon constraints set early on.

The memory result resonated with us most, because persistent memory is foundational to how Intueo works. Our agents maintain memory blocks and a file system precisely so they can carry intent across sessions and pick up exactly where they left off. When we pair Fable 5 with that persistent memory on long-running tasks, the model does what Anthropic's Slay the Spire experiment suggests: it writes notes worth keeping, consults them at the right moments, and carries decisions forward instead of relitigating them. The difference between an agent that finishes work and one that needs constant supervision lives exactly here, and Fable 5 lands firmly on the right side of it.

For context on how we deploy models: our Sage 1.5 architecture is model-flexible by design. It routes across several models, including Kimi, GLM, and a ChatGPT model it can switch to on its own, and that blend already performs well across the majority of agentic tasks we run. We support a wide range of LLMs on purpose, so our customers are never locked to a single provider, and every job runs on the engine best suited for it. Within that mix, we lean on Anthropic's Claude models heavily, and Fable 5 has immediately earned a place at the demanding end of our routing: the long-running, high-stakes, many-step jobs where sustained coherence is worth paying for.

A few operational observations from our runs that go beyond the press release:

• Fewer dead ends in tool orchestration. On multi-tool workflows, Fable 5 recovers from failed calls more gracefully, adjusting its approach rather than repeating the same failing action. Anthropic's drug design study described the model "recovering from failures along the way," and that behavior generalizes to ordinary business tooling. [1]
• Self-checking pays for itself. At higher effort settings the model reflects on and validates its own work before moving on, which is what one early-access customer meant by "the extra thinking pays for itself." [1] For autonomous operations, output you do not have to re-check is the entire value proposition.
• Token efficiency is not hypothetical. Partners reported runs finishing 25–30% faster with fewer turns, and physics research using a third of the reasoning tokens. [1] In our experience the model gets to the point. For a platform billing real workloads, fewer wasted tokens directly improves the economics of every automation.
• The fallback is a non-event in practice. Our workloads are business operations, not exploit development, so the sub-5% classifier fallback essentially never intersects with what our agents do.

What This Unlocks for Real Businesses

Step back from the benchmarks and ask what actually changes for a company automating its operations.

The constraint on AI automation has never really been intelligence-per-turn. Models have been smart enough for individual steps for a while. The constraint has been trust over time: how many steps can run unattended before a human has to check in? Every checkpoint a workflow requires is a place where automation stalls, queues behind a busy person, and loses its speed advantage.

Fable 5 moves that frontier. A model that holds focus across millions of tokens, uses persistent memory the way a careful employee uses notes, validates its own work, and recovers from failures changes the unit economics of delegation:

Migrations, research synthesis, multi-system operations, compliance reviews, financial analysis spanning dozens of documents: these stop being brittle chains of small AI calls stitched together with hope, and start being single coherent agent runs. That is the territory Intueo was built for, and it is why this release is more than a headline to us. It directly raises the ceiling on what our agents can take off a customer's plate.

It is also worth being clear-eyed about why so many agentic projects fail, because better models alone will not fix it. Gartner's projection that over 40% of agentic AI projects will be cancelled by 2027 is not a verdict on the technology; it is a verdict on deployments that bolt a model onto a workflow with no memory architecture, no tool governance, no evaluation loop, and no clear owner of the outcome. [4] McKinsey's finding that only a quarter of organizations are scaling agents anywhere points the same direction: the winners are not the companies with API keys to the best model, they are the companies that built the operational scaffolding to use it. [3] A Mythos-class model raises the ceiling, but the floor, reliable production operations, is still built from platform engineering. That division of labor is exactly where we focus.

Our Honest Take on the Caveats

A launch review is only useful if it also tells you what to watch. Four things temper the excitement:

1. The safeguards are deliberately overcautious for now. Anthropic admits benign requests will sometimes trigger the classifiers, especially in biology and chemistry, and is working to reduce false positives. If your domain is life sciences, expect some friction until the trusted access program for biology spins up. [1]
2. Subscription access is staged. The included window on Pro, Max, Team, and seat-based Enterprise runs through June 22, after which usage credits apply until capacity allows otherwise. Plan accordingly if your workflows depend on subscription access. [1]
3. The 30-day retention requirement is new. For most businesses it is a reasonable trade for the safety monitoring it enables, but regulated industries should review it against their data policies before adopting. [1]
4. Vendor-reported results deserve independent replication. Most of the headline numbers come from Anthropic and its early-access partners, who had commercial relationships with the launch. The pattern is consistent and the third-party voices are credible, but the field's history says wait for independent evaluations, METR-style autonomy measurements, and a few months of production mileage before treating any single claim as settled. Our own platform experience supports the long-horizon and memory claims specifically; we have not independently verified the science results. [2]

None of these change our overall assessment. They are the visible cost of shipping a Mythos-class model to everyone, quickly, without pretending the risks away, and Anthropic has been unusually transparent about every one of them.

Where We Go From Here

We are expanding Fable 5's role inside Sage for the demanding, long-running tasks where it demonstrably shines, while keeping our multi-model routing so every job continues to run on the best engine for it. A frontier model that genuinely improves long-horizon autonomy, memory use, tool reliability, and token efficiency is exactly the kind of advance that compounds inside a platform like ours: every one of those gains flows directly into what our agents can do for the businesses that rely on them.

The bigger picture is just as encouraging. A year ago, "AI agent" mostly meant a demo that worked until it didn't. Today, the most capable model in the world is being judged, by its maker and its earliest customers alike, on how long it can work alone and how well it remembers. The industry has converged on the same definition of useful AI that we built Intueo around: agents that hold context, carry memory, use tools, and finish the job.

Fable 5 delivers on the dimensions we care about most, we have verified that on our own platform, and we are already putting it to work. If you want to see what Mythos-class capability looks like applied to your business operations, that is exactly the conversation we love to have.