Claude Fable 5 Is Here: The Mythos-Class Model, Made Safe for Everyone

From Too Dangerous to Release, to Available Everywhere

To understand why Fable 5 matters, you have to start in April 2026. That is when Anthropic revealed it had built a model called Mythos — a tier above its Opus class — so capable at autonomously finding and chaining cybersecurity exploits that the company decided it was too dangerous to release publicly. Instead, it locked Mythos behind Project Glasswing, granting access only to a small circle of cyber defenders and critical infrastructure partners, including names like Apple and Microsoft. The promise at the time was that, once safeguards were strong enough, Mythos-level capability would eventually reach everyone. June 9, 2026 is the day that promise was kept. Claude Fable 5 is that same Mythos-class model, wrapped in safeguards robust enough for general release. The naming is deliberate and tells the whole story: Fable comes from the Latin fabula, the linguistic sibling of the Greek mythos. Same root, same underlying power — the safeguards are the only thing that separates the public Fable from the restricted Mythos. Alongside it, Anthropic also released Mythos 5 to its trusted Glasswing partners: the identical model with the cybersecurity safeguards removed, described as having the strongest cyber capabilities of any model in the world.

The Benchmarks: It Beats the Model Anthropic Called Too Dangerous

The single most striking number is on SWE-Bench Pro, the standard benchmark for agentic coding. Fable 5 scores 80.3%. To put that in context: Opus 4.8 — which until this morning was the strongest model Anthropic offered to the public — scores 69.2%. GPT-5.5 sits at 58.6%, and Gemini 3.1 Pro at 54.2%. So Fable 5 is roughly eleven points above the best public model, and more than twenty points above its two biggest competitors. But the genuinely remarkable part is the comparison to Mythos itself: the original Mythos Preview, the model Anthropic judged too dangerous to release back in April, scored 77.8% — which means the public, safeguarded Fable 5 actually edges out the restricted model that started this entire story. The lead holds across the board. On Terminal-Bench 2.1, another agentic coding test, Fable 5 reaches 88.0% against Opus 4.8 at 82.7%, OpenAI's Codex CLI at 83.4%, and Gemini CLI at 70.7%. On GDPval-AA for knowledge work it scores 1932 versus Opus at 1890, GPT-5.5 at 1769, and Gemini at 1314. And on Humanity's Last Exam for multidisciplinary reasoning it leads both with and without tools. One honest caveat keeps the picture accurate: on the benchmarks marked with an asterisk — the cybersecurity-heavy and biology-heavy ones — the public Fable 5 performs closer to Opus 4.8 rather than the top number, because those are precisely the queries its safeguards route to Opus. The published table shows the higher of Mythos 5 and Fable 5; on the everyday coding, knowledge, and reasoning benchmarks most people actually use, that higher number is the one you get.

What It Can Actually Do — Beyond the Marketing

The capability claims around Fable 5 are unusually concrete, because they come from named customers running real work. In software engineering, Stripe reported that the model compressed months of effort into days: in a Ruby codebase of fifty million lines, it carried out a codebase-wide migration in a single day — work estimated at more than two months for a full team working by hand. On Cognition's FrontierCode evaluation, which judges whether code meets production-quality standards, Fable 5 scored highest among frontier models even at medium effort. In vision, it set a new state of the art: it can extract precise numbers from dense scientific figures and rebuild a web app's entire source code from screenshots alone. Most strikingly, it beat the game Pokémon FireRed from start to finish using nothing but raw game screenshots — no maps, no navigation aids — where earlier Claude models needed elaborate helper tools just to function. It also holds focus across millions of tokens in long-running tasks and improves its own output by keeping notes for itself. And across early-access benchmarks in finance, law, and analytics, customer after customer reported it clearly surpassing Opus 4.8 — one noting it was the first model to break ninety percent on their core analytics benchmark, a ten-point jump.

How the Safeguards Work — and Why You Will Barely Notice Them

Releasing a model this capable safely required a different approach from a simple refusal. Fable 5 ships with a layer of separate AI classifiers that watch for three categories of sensitive request: cybersecurity, biology and chemistry, and attempts to distill the model's capabilities to train competitors. When a request falls into one of these areas, Fable 5 does not refuse — it quietly hands the response over to Opus 4.8 instead, and tells you that it has done so. This is a meaningful design choice: a fallback to Opus 4.8, itself a highly capable model, is a far better experience than a flat refusal. Anthropic deliberately tuned these classifiers to be cautious, which means they will sometimes catch harmless requests — an acknowledged trade-off the company says it will narrow over time. But the practical impact is small: the safeguards trigger in fewer than five percent of sessions, meaning that for more than ninety-five percent of real use, you are interacting with the full, unrestricted power of Fable 5. Anthropic also ran an external bug bounty of over a thousand hours that found no universal jailbreak, and one partner reported Fable 5's cyber safeguards were the most robust of any model tested. There is one policy change worth knowing: business traffic on Mythos-class models carries a mandatory thirty-day data retention for safety monitoring, after which it is deleted and never used for training.

Price, Access, and Whether It Is for You

Fable 5 is priced at ten dollars per million input tokens and fifty per million output — notably, less than half the price of the original Mythos Preview, and roughly double the current Opus tier. On the API and consumption-based Enterprise plans, it is fully available starting today as claude-fable-5. For subscription plans, Anthropic chose a staged rollout because it expects very high and hard-to-predict demand: from today through June 22, Fable 5 is included on Pro, Max, Team, and Enterprise plans at no extra cost; on June 23, it will be removed from those plans and continued use will require usage credits until capacity expands; and eventually, once there is enough capacity, Anthropic aims to restore it as a standard part of subscription plans. The honest practical guidance is this: Fable 5's advantage grows with the length and difficulty of the task, so it is the model to reach for on large migrations, deep research, senior-level analysis, and vision-heavy work — the problems where its lead is widest. For quick questions, simple edits, and everyday tasks, Opus 4.8 or Sonnet remains the more economical choice, especially once usage credits enter the picture after June 22. The bigger picture is what stands out: a tier of capability that Anthropic considered too dangerous to release just two months ago is now, with safeguards, in the hands of anyone with a paid plan.

Prompt

# CLAUDE FABLE 5 — WHAT LAUNCHED JUNE 9, 2026

# ─── THE MODEL ───
# API string: claude-fable-5
# Tier: Mythos-class — sits ABOVE Opus (a new top tier)
# Claim: capabilities exceed any model Anthropic has made generally available
#        state-of-the-art on nearly all tested benchmarks
#        the longer/harder the task, the bigger its lead

# ─── WHERE IT CAME FROM ───
# April 2026: Anthropic built "Mythos" — too capable at finding security
#             vulnerabilities to release publicly. Held back via Project Glasswing.
# June 9 2026: Fable 5 = that same Mythos-class model + safeguards for public use.
# Name: "Fable" (Latin fabula) = safe sibling of "Mythos" (Greek). Same model,
#        the safeguards are the only difference.
# Mythos 5 = the same model with cyber safeguards LIFTED — restricted to
#            Glasswing cyber-defense partners + (soon) select biology researchers.

# ─── VERIFIED CAPABILITY HIGHLIGHTS ───
# Coding:  Stripe — codebase-wide migration in a 50M-line Ruby codebase in ONE
#          day (would take a team 2+ months by hand). Highest on FrontierCode.
# Vision:  new SOTA — rebuilds a web app's source from screenshots alone;
#          beat Pokémon FireRed with a vision-only harness (no maps/tools)
# Memory:  stays focused across millions of tokens; uses its own notes;
#          file-based memory helped it 3x more than Opus 4.8 on Slay the Spire
# Knowledge work: top score on Hebbia finance benchmark; aced IMC trading evals
# Science: (via Mythos 5) ~10x faster protein design; novel biology hypotheses
#          preferred ~80% vs Opus-class; novel genomics beating a Science paper

# ─── HOW THE SAFEGUARDS WORK (important for daily use) ───
# When a request touches: 1) cybersecurity  2) biology/chemistry  3) distillation
#   → Fable 5 hands the answer to Opus 4.8 instead, and TELLS you it did
# Tuned conservatively → sometimes catches harmless requests
# BUT: triggers in <5% of sessions → 95%+ of use gets full Fable 5 power
# A fallback to Opus 4.8 is still a strong answer, not a refusal

# ─── PRICING & ACCESS ───
# Price: $10 / M input · $50 / M output (less than half of Mythos Preview)
# API + consumption Enterprise: fully available today
# Subscriptions (staged rollout):
#   - Now → June 22: included on Pro, Max, Team, Enterprise at NO extra cost
#   - June 23 onward: may require usage credits until capacity expands
#   - Later: aim to restore as a standard part of subscription plans
# Business data: mandatory 30-day retention for safety monitoring
#   (not used for training; deleted after 30 days)

# ─── WHEN TO REACH FOR FABLE 5 ───
# Long-horizon, complex work where its lead is largest:
#   large migrations, deep research, senior-level analysis, vision-heavy tasks
# For quick/simple work, Opus 4.8 or Sonnet is more economical.