Claude Fable 5 Review: Anthropic's Mythos-Class Flagship

Strip away the launch language and four capabilities carry Fable 5: agentic coding, long-horizon autonomy, persistent memory, and vision. Each is backed by a concrete claim rather than an adjective, which makes them testable and worth taking seriously.

Long-Horizon Autonomy and Memory

The most material change over Opus 4.8 is that Fable 5 holds a goal across very long tasks. Anthropic reports it can run for days, and that its lead over prior models grows as task length and complexity increase. The memory system is file-based and persistent, so the model writes state to disk and reads it back rather than relying solely on context. The cleanest demonstration is an odd one: on the game Slay the Spire, persistent memory improved Fable 5's performance three times more than it improved Opus 4.8, and Fable 5 reached the final act three times more often. That is a narrow result, but it points at something real about sustained, stateful problem-solving rather than one-shot answers.

Vision That Does Real Work

Anthropic claims a new state of the art on vision, and the examples are specific rather than abstract. Fable 5 beat Pokemon FireRed start to finish using only raw screenshots, with no map data and no game-specific harness. It rebuilds web-application source code from screenshots of the rendered page, and it extracts precise numbers from scientific figures. For anyone doing UI work, document processing, or data extraction from charts, that last capability is more useful day-to-day than beating a video game from 2004.

Agentic Coding and Knowledge Work

Coding is where Fable 5's headline benchmarks land, and where Anthropic stakes its leadership claim. Beyond raw code generation, Anthropic and independent partners report strong results on legal drafting and markup and on finance reasoning. Hex, an analytics platform, reports Fable 5 as the first model to break 90% on its complex long-running analytical benchmark, roughly a ten-point jump over Opus 4.8. Hebbia reports the highest score of any model on its senior-level finance reasoning benchmark. Both of those are independent partners, which matters, and both are covered in the benchmark section below with that attribution attached.

This is the part of Fable 5 you have to understand before you adopt it, because it changes how the model behaves on a measurable fraction of requests. Fable 5 ships under ASL-3 protections and is classified CB-1, meaning it is capable with non-novel chemical and biological weapons information but does not cross the CB-2 threshold for novel weapons. That classification is why the safeguards exist.

How the Classifier Works

Anthropic runs conservative classifiers that watch for cybersecurity, biology, chemistry, and model-distillation requests. Here is the important detail: when a classifier triggers, the query is not refused. It is routed to Claude Opus 4.8, the next-most-capable model, and you are told it happened. So the failure mode is not a hard wall. It is a silent-ish downgrade to a previous-generation model. For sensitive-but-legitimate work, you still get an answer, just from a weaker model than the one you are paying for.

Anthropic tuned the classifier conservatively to ship fast, and says it triggers in under 5% of sessions on average, a figure that explicitly includes false positives on benign prompts. Users have reported downgrades on requests as harmless as a pulled-pork shopping list and basic low-level software-engineering questions. That is the cost of conservative tuning, and it is a real cost: on roughly one in twenty sessions you are not getting Fable 5, you are getting Opus 4.8, sometimes for no good reason.

Mythos 5 and Project Glasswing

The same model with safeguards lifted in some areas is Claude Mythos 5, and it is restricted to vetted partners defending critical software infrastructure through Project Glasswing, a collaboration that includes the US government. Access is expanding through a trusted-access program into areas like biomedical research, with roughly 150 organizations being added. Named participants include Dragos, Tenable, Trend Micro, Netskope, BeyondTrust, Rubrik, BT, Intercontinental Exchange, and Hitachi.

The capability that justifies all of this is concrete. On a Firefox exploit-development evaluation, the unsafeguarded Mythos 5 scored 88.4% against Opus 4.8's 8.8%. That is not a typo and not a rounding gap. It is the clearest single illustration of why Anthropic split the release: a model that good at building exploits is dual-use by definition. Framed factually, the safeguards plus the restricted-access program are Anthropic's attempt to keep the offensive edge inside organizations that defend infrastructure rather than handing it to everyone with a credit card. You can reasonably debate whether that boundary holds. You cannot reasonably call the precaution unnecessary given an 88.4% exploit-development score.

A launch this strong invites uncritical praise. Resist it. These are the documented weaknesses, drawn from Anthropic's own system card and from independent partners, and they should carry real weight in your decision rather than sit in a footnote.

Fable 5 is the strongest generally-available model Anthropic has shipped, and the agentic-coding, long-horizon, and vision claims hold up well enough that they are not just marketing. The independent signals reinforce the vendor story rather than contradict it: Artificial Analysis puts a 152-point Elo gap on it over Opus 4.8 on real professional tasks, and Cursor, Hex, and Hebbia all report gains on their own benchmarks. That is a healthier evidence base than a launch built purely on self-reported numbers.

But this is a skeptic's review, and the position has to be specific rather than "it depends." Here it is.

Adopt Now If

You do agentic coding in Claude Code, long-horizon automation, vision-heavy document or UI work, or guided legal and finance analysis. For those workloads Fable 5 is a clear step up from Opus 4.8, the pricing is reasonable at $10/$50, and the free window on Pro, Max, and Team through June 22 lets you validate it against your real tasks at no extra cost. Run that test this week, because the window is two weeks wide.

Wait and Watch If

Your work is open-ended biology or chemistry ideation, heavy multi-step quantitative finance modeling, or anything that brushes against the safety classifier's topics often enough that the under-5% downgrade rate becomes your everyday experience. For those cases, the rough edges, the over-refusal tax, and the lack of independent replication on the headline SWE-bench Pro lead are real reasons to let the dust settle. Let the June 23 onward replication cycle tell you whether the cross-vendor gaps survive someone else's harness.

For Agent Builders Specifically

Adopt, but keep humans on irreversible actions. The documented prefill vulnerability and the occasional reckless or destructive agentic behavior mean you should sandbox aggressively and require approval for anything that deletes data or escapes its boundary. The capability is worth it; the autonomy is not yet worth trusting unsupervised.