Anthropic Just Opened a New Tier Above Opus — Meet Claude Fable 5 and Mythos 5

Anthropic Bumped Its Whole Ladder Up a Rung

Here's the deal: on June 9, 2026, Anthropic created a new class that sits above what used to be its top tier, Opus. They're calling it "Mythos." And they shipped the first two models in it on the same day — Claude Fable 5 and Claude Mythos 5. The short version: Fable 5 is the first Mythos-class model the public can actually use today, and Mythos 5 is the same underlying model with some safeguards lifted, handed only to government and cyber-defense agencies.

Why is this a big deal? Because just days earlier Anthropic was the company warning that AI is "getting too dangerous." Then it turned around and released the most powerful model it has ever made publicly available. Warning and shipping in the same week — which immediately kicked off the debate: so is it safety-first or shipping-first?

The numbers explain the move. Fable 5 scored 95% on SWE-bench Verified and 80% on SWE-bench Pro. It's not just coding — Anthropic says it's state-of-the-art across nearly every benchmark, from knowledge work to vision to scientific research. In plain terms, it's the strongest model on the market right now.

The Cast — Fable 5, Mythos 5, and Project Glasswing

First, the "Mythos class" itself. Claude used to be a three-tier lineup: Haiku (fast, cheap), Sonnet (mid), Opus (top). Anthropic just stacked one more rung on top called Mythos — the frontier of the frontier, capabilities strong enough that releasing them wholesale gives the company pause.

Second, Claude Fable 5. This is the Mythos-class model wrapped in guardrails so the public can use it safely. On certain high-risk topics, queries get routed to Claude Opus 4.8 instead of answered by Fable 5 directly. Anthropic says these safeguards trigger in under 5% of sessions on average — the other 95%+ get the full power.

Third, Claude Mythos 5. Same base model as Fable 5, but with some guardrails lifted. It isn't handed to everyone — it's deployed selectively through "Project Glasswing," in collaboration with the U.S. government, to cyber-defense agencies and critical-infrastructure providers, upgrading the earlier Claude Mythos Preview. Anthropic says that during the Preview stage it already found thousands of zero-day vulnerabilities across operating systems and browsers.

The last character is price. Both models cost $10 per million input tokens and $50 per million output — roughly double Opus 4.8. But from June 9 through June 22, they're included at no extra cost for Pro, Max, Team, and seat-based Enterprise subscribers. On the API and consumption-based Enterprise plans, normal billing starts right away. It's the classic free-trial-then-lock-in play.

What Actually Changed

Item	Claude Fable 5	Claude Mythos 5
Tier	Mythos (above Opus)	Mythos
Access	Public (Pro/Max/Team/Enterprise/API)	Government & cyber-defense only (Glasswing)
Guardrails	On (high-risk <5% → routed to Opus 4.8)	Partially lifted
Benchmarks	SWE-bench Verified 95%, Pro 80%	Same base model
Price	$10/M in · $50/M out	$10/M in · $50/M out
Free window	Through June 22 (paid subs)	N/A

The key point this table makes: Fable and Mythos aren't different models — they're the same model with different safety settings. Anthropic chose not to dumb the capability down, but to gate only the dangerous slice. So the public gets near-peak performance, and the guardrails only fire in that genuinely risky 5%.

AWS announced the same day that Fable 5 is available on Bedrock with built-in safeguards, widening distribution beyond Anthropic's own API. Day one, it was already wired into the broader developer ecosystem.

Who Gains What

For Anthropic, the win is clear: keep the "safety-first" brand while proving it can match OpenAI and Google on raw capability. Splitting Fable from Mythos is a clever compromise — delay a release over risk and you cede the market; ship everything and your safety story collapses. "Gated version for institutions, guardrailed version for the public" threads that needle, and the two-week free trial harvests usage and lock-in at once.

For developers and enterprises, the question is whether double the Opus price is worth it. 95% on SWE-bench Verified is essentially senior-engineer-level coding. If it can handle complex refactors, multi-file debugging, and agentic automation, the productivity easily covers a 2x token cost — and the free window is a no-risk way to find out.

For U.S. agencies, Mythos 5 is about arming the defenders before the attackers. In an era where AI can auto-discover zero-days, who holds that capability first matters. Anthropic finding thousands of vulnerabilities during Preview signals intent to put that power in defenders' hands first. That's the context for Dario Amodei's line about giving "researchers and defenders access to capabilities beyond what we've ever made publicly available."

Past Parallels — Capability vs. Safety

The "how do you release powerful-but-risky capability" problem isn't new. A success story: OpenAI's staged GPT-2 release, which started small out of caution and earned a "responsible" reputation. Anthropic's Fable/Mythos split is the same lineage — don't release the whole capability, carve off the risky slice and restrict it.

The failures cut both ways too: models jailbroken because guardrails were too loose, and models that lost users because guardrails were so tight they refused harmless questions. The trick isn't how much you gate — it's gating exactly the risky band. Anthropic stressing "<5% of sessions" is a direct nod to past over-blocking complaints.

The lesson: the stronger the model, the sharper the safety-vs-usefulness tradeoff. Block too much and nobody uses it; release too much and things break. Anthropic's bet is to solve the balance by splitting the model in two. Whether it works depends on whether Mythos 5 delivers defense wins without incidents and Fable 5 holds up without jailbreaks.

Competitor Counterplay

OpenAI is the most direct rival, pushing GPT-5.5 into the mass market with personalization and multimodal. If Anthropic goes "up" with a premium Mythos tier, OpenAI likely counters on benchmarks or on price — "comparable power, cheaper." With OpenAI now in the IPO race too, "owns the strongest model" is a key listing-story card.

Google plays a different game, putting Gemini 3.5 Flash into Search as the global default — winning on speed and distribution. Anthropic chases premium performance; Google runs a volume play with a good-enough model already in front of billions. Same market, different angle: the "premium vs. ubiquity" split just got sharper.

The interesting wildcard is safety-as-marketing. If Anthropic leads with "Project Glasswing arms cyber defense," rivals may roll out their own security programs aimed at government and enterprise. As raw capability commoditizes, "how safely, how responsibly" becomes the differentiator — and this launch just set a new baseline.

So What Changes — By Reader

For developers, the impact is biggest. A 95%-on-SWE-bench model, free for two weeks, means you can bolt an "AI senior pair programmer" onto your workflow at zero cost. Throw a hard task at it before June 22 and decide for yourself if the 2x token price pays off. Just mind the date — billing starts after.

For enterprise and security teams, treat this as the signal that "AI auto-finds vulnerabilities" is now real. Mythos 5 going to defenders first means attackers get the same capability soon. Reason enough to audit your zero-day exposure and accelerate AI-based security tooling. It's early to be certain, but the center of gravity in cyber offense/defense is shifting fast toward automation.

For everyone else, the big picture: the AI-safety debate just moved from abstract argument to product design. Anthropic warning "this is dangerous" and shipping "the strongest model" in the same week shows safety is no longer a brake on shipping but a variable in how you ship. Whether this balancing act succeeds will shape every AI company's release strategy.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? If you're a paid Claude subscriber, you can use the strongest model available right now at no extra cost through June 22. Throw real coding, writing, and research work at it and judge the value yourself. Just remember the token price is double Opus once the free window closes.

— What's the actual difference between Fable and Mythos? Same base model. The only difference is the safety settings. Fable 5 gates the risky band (<5% of sessions) and ships to the public; Mythos 5 lifts some of those guardrails and goes only to government and cyber-defense agencies. It's not a capability gap — it's a "who's allowed how far" gap.

— If it's "dangerous," why ship it? Anthropic's logic is "defenders should be armed first." AI that auto-finds vulnerabilities is coming regardless, so it's better that defenders hold it before bad actors. Whether that's genuinely safer or a rationalization under competitive pressure is too early to call — it'll be settled by whether incidents actually happen.

References

• Claude Fable 5 and Claude Mythos 5 — Anthropic
• Anthropic's Claude Fable 5 is a version of Mythos the public can access today — TechCrunch
• Anthropic releases Mythos-like AI model to the public, Claude Fable 5 — CNBC
• Anthropic Releases Claude Fable 5, Its Most Powerful AI Yet, With Cyber Safeguards — The Hacker News
• Initial impressions of Claude Fable 5 — Simon Willison

Numbers and criteria are as of announcement and may change.