Claude Fable 5: Anthropic Opens Its Mythos-Class Model to the Public

On June 9, 2026, Anthropic made the first member of its Mythos model family generally available. Claude Fable 5 is, by Anthropic's own description, "state-of-the-art on nearly all tested benchmarks of AI capability," and as of today it sits at the top of our AI model leaderboard — ahead of Claude Opus 4.8, which had held the #1 spot since May 28.

What makes this release notable is less any single benchmark than the safety architecture Anthropic used to ship it broadly. Here is what Fable 5 is, what the numbers say, where it lands against the field, and how to think about adopting it.

The Fable / Mythos Split: One Frontier, Two Products

Anthropic shipped its strongest model as two distinct products (Anthropic):

• Mythos 5 — the unrestricted model, available only to vetted customers under oversight. It retains the full cyber and biology capabilities Anthropic has flagged as high-risk.
• Fable 5 — the generally available product. It is the same underlying model with safeguards that block responses in specific high-risk domains (notably offensive cybersecurity and biology). When a prompt lands in one of those areas, Fable 5 doesn't refuse — it routes the answer to Opus 4.8, a less capable model in those domains (NBC News, CNBC).

This routing is why Fable 5's published scorecard carries asterisks on certain rows. Capabilities like ExploitBench (78.0%) and the hard split of BiologyMysteryBench (46.1%) reflect Mythos 5 — Fable 5 itself scores effectively zero on offensive cyber tasks by design (Digital Applied). When the materials say "state-of-the-art," read it as "on everything that isn't deliberately walled off."

What the routing means in practice

For most developers, the fallback is invisible. General coding, analysis, writing, vision, and agentic work all run on the full Fable 5 model. You only hit the Opus 4.8 downgrade if your prompt is classified into a restricted domain — which, for ordinary product and engineering workloads, should be rare. The practical implications worth planning for: outputs on borderline security or biology topics may be noticeably weaker than the headline benchmarks suggest, and behavior on those topics is effectively pinned to a different model with different latency and formatting. If your application operates near those domains — security tooling, certain research workflows — test against the fallback path explicitly rather than the Fable 5 scorecard.

TechCrunch noted the timing: the release came days after Anthropic publicly warned that AI capability is outpacing safety. The steer-to-a-weaker-model mechanism is the company's stated answer to that tension — a way to release frontier capability broadly while keeping the highest-risk capabilities behind a gate.

The Benchmarks

Anthropic led with agentic and knowledge-work benchmarks rather than classic knowledge tests. The figures below are Fable 5's own results (the non-asterisked rows), with Opus 4.8 and GPT-5.5 for context (Digital Applied):

The SWE-bench Pro result is the clearest gain: 69.2% to 80.3% over a model that topped the field twelve days earlier, and roughly double GPT-5.5's score. SWE-bench Pro measures whether a model can resolve real GitHub issues end-to-end inside a repository, so an 11-point jump translates fairly directly into how often an autonomous coding agent lands a working change without human correction.

FrontierCode Diamond is a deliberately hard coding eval where even strong models score in the single digits to low teens; Fable 5's 29.3% more than doubles Opus 4.8 and is several times GPT-5.5. GDPval-AA, meanwhile, is an Elo-style rating over economically valuable knowledge-work tasks rather than a percentage benchmark — the move from 1890 to 1932 is incremental on that scale, which is a useful reminder that the gains are largest on coding and agentic tasks and more modest on general knowledge work.

It is also worth reading the rows where Fable 5 is not impressive. On Anthropic's own scorecard the model scores 17.4% on AutomationBench (tool use), 13.3% on the Legal Agent Benchmark, and 38.6% on Blueprint-Bench 2 (spatial reasoning). Those are low absolute numbers, and they are a useful corrective to the "state-of-the-art on nearly all benchmarks" framing: the model leads the field on coding and computer-use, but several categories of agentic and domain-specific work remain genuinely unsolved. If your workload sits in one of those columns, the right baseline expectation is "better than the previous best, but still unreliable," not "solved."

Several other rows on Anthropic's chart — Terminal-Bench 2.1 at 88.0%, Humanity's Last Exam with tools at 64.5%, HealthBench Professional at 66.0% — are Mythos 5 figures. Fable 5 falls back to Opus 4.8 for the prompts that would exercise them, so treat those as the family ceiling, not the public model's output on restricted topics.

Pricing, Access, and Undisclosed Specs

Fable 5 is double the price of Opus 4.8 (VentureBeat):

• Input: $10 per million tokens
• Output: $50 per million tokens
• Prompt caching: 90% discount on cached input
• US-only inference: 1.1x multiplier

To put that in workload terms: a coding-agent task that consumes, say, 200K input and 40K output tokens costs roughly $4 on Fable 5 versus about $2 on Opus 4.8. With heavy prompt caching the input side narrows considerably, but the output multiple is the durable difference, and output tends to dominate cost on agentic runs.

It's available now on the Claude API (claude-fable-5), claude.ai, and across AWS Bedrock, Google Cloud, and Microsoft Foundry, with enterprise access on consumption-based plans.

Anthropic did not publish the context window or maximum output tokens at launch (Digital Applied). Our leaderboard entry assumes the 1M-token window standard across recent Claude flagships, but that is an inference until the model card lands. If your use case depends on a specific context limit, confirm it against the API rather than the marketing materials.

Where Fable 5 Lands on the Leaderboard

We've added Fable 5 to the AI model leaderboard at the top of the table:

*Arena Elo for Fable 5 is our estimate; LMArena had not published a verified rating at the time of writing. The model card omits context window and max-output specs.

One note on methodology, since it bears on how much weight to put on that #1. Our quality index blends classic benchmarks (knowledge, coding, math) with Arena Elo. Anthropic led the Fable 5 launch with agentic and knowledge-work benchmarks rather than the MMLU-style tests our index has historically used, and it did not publish a verified Arena rating. So the inputs behind Fable 5's index score lean more on our reading of the agentic results than on like-for-like classic-benchmark numbers. The top-of-board placement is well supported by the SWE-bench Pro, FrontierCode, and OSWorld results, but treat the exact index value as provisional until the standard benchmark suite and a verified Arena Elo are available.

Fable 5 takes the top spot on capability. But note the price column: at $10/$50 it is twice Opus 4.8 and several times Gemini 3.1 Pro on output. For most production workloads, Opus 4.8 at half the price — or a routed setup that reserves Fable 5 for genuinely hard agentic tasks — is the more sensible default. A top-ranked model is usually better treated as an escalation target than a baseline.

Should You Switch From Opus 4.8?

For most teams, not wholesale. The case for moving a workload to Fable 5 is strongest where the task is hard agentic coding or multi-step computer-use, the failure cost of a wrong answer is high, and the token volume is low enough that doubling the rate is tolerable. The case against is everything else: high-volume, latency-sensitive, or cost-sensitive traffic where Opus 4.8 — itself the best public model in the world less than two weeks ago — already clears the bar.

A reasonable migration path is to leave your default on Opus 4.8 (or a cheaper tier) and add a rule that escalates to Fable 5 only when a task is flagged as high-complexity or has already failed once. That captures most of the capability upside while keeping the cost increase proportional to where it actually helps.

It's also worth pausing on availability before you commit. Fable 5 is live on the Claude API, claude.ai, and three major clouds, but a brand-new frontier model in its first days typically sees tighter rate limits, occasional capacity throttling, and a model card that is still filling in. For a production cutover, that argues for a staged rollout — shadow-traffic the model against your current default, watch for latency and rate-limit behavior under real load, and only shift the default once the operational picture is stable. The capability is available today; the operational maturity that production depends on usually trails a launch by a few weeks.

The Competitive Picture

This release also tightens Anthropic's grip on the top of the board. With Fable 5 at #1 and Opus 4.8 at #2, Anthropic now holds the two highest capability slots on our index, with OpenAI's GPT-5.5 family and Google's Gemini 3.1 Pro clustered just below. The gap between those models is small enough that benchmark leadership has been changing hands every few weeks through the spring, so the more durable question for a buyer is not "who is #1 today" but "which provider's pricing, availability, and roadmap fit my workload." On that question, the $10/$50 rate is the relevant data point: Anthropic is pricing Fable 5 as a premium escalation tier, not as a replacement for the workhorse models that handle the bulk of production traffic.

What It Means

Three observations:

1. The safety mechanism is part of the product. Anthropic released frontier capability broadly by steering high-risk prompts down to a weaker model. How that holds up under adversarial pressure is an open question, and worth watching over the coming months.

2. The frontier is being measured on coding and agents. Fable 5 led with SWE-bench Pro, FrontierCode, OSWorld, and GDPval — software engineering, autonomous tool use, and knowledge work — rather than trivia benchmarks. That reflects what the model is positioned for, and where the gains are concentrated.

3. Price discipline still applies. With a model now scoring at the top of our index, the temptation is to route everything to it. A fast, cheap tier for routine traffic with Fable 5 reserved for hard tasks will usually be the better-performing architecture on a cost-adjusted basis.

Running the same prompt across Fable 5, Opus 4.8, and a cheaper open-weight model in the Swfte Playground is the fastest way to see whether the capability gap justifies the price gap for your workload — sometimes it does, sometimes the cheaper model is close enough.

---

The AI model leaderboard refreshes regularly with Arena Elo, provider pricing, and benchmark cross-validation. For the open-weight side of this week's news, see our breakdown of MiniMax M3.

Sources: Anthropic — Claude Fable 5 & Mythos 5 · NBC News · CNBC · TechCrunch · VentureBeat · AWS / About Amazon · Digital Applied — benchmark breakdown