Claude Fable 5 Explained: Anthropic's Mythos-Class Model vs Opus 4.8 and GPT-5.5

This morning, Anthropic released Claude Fable 5 — and full disclosure, this post was researched and written with it on launch day. Fable 5 isn't another incremental Opus bump. It's a Mythos-class frontier model — the same architecture Anthropic has been keeping behind locked doors for vetted cybersecurity researchers — re-engineered with safety classifiers so the rest of us can finally use it. The benchmarks are absurd, the price is double GPT-5.5's, and the story behind the release is more interesting than either.

What Is Claude Fable 5?

Claude Fable 5 launched on June 9, 2026 as half of a two-model release. The other half, Claude Mythos 5, shares the exact same underlying architecture but stays restricted — available only to Anthropic's Project Glasswing cybersecurity partners, with approved biology researchers joining through a trusted access program soon.

The difference between the two isn't capability. It's safeguards. Fable 5 ships with safety classifiers that monitor incoming requests, and Mythos 5 has those restrictions removed. That's it. Anthropic's own framing is that Fable 5 is "a Mythos-class model made safe for general use" — which makes this the first time a frontier lab has shipped its actual top-end internal model to the public and been explicit that the only thing separating you from the unrestricted version is a classifier layer.

How the Safety Fallback Actually Works

This is the clever part, and most launch-day coverage is burying it. Instead of refusing sensitive requests outright, Fable 5 runs three classifiers:

• Cybersecurity — blocks exploitation and offensive cyber tasks entirely. Anthropic ran an external bug bounty with over 1,000 hours of testing and reports no universal jailbreaks were found.
• Biology & chemistry — flags dual-use biomedical research. Anthropic admits this one is deliberately conservative for launch.
• Distillation prevention — blocks attempts to extract Fable 5's capabilities to train competing models.

When a request trips a classifier, it quietly falls back to Claude Opus 4.8 instead of dying with a refusal. Anthropic says more than 95% of sessions never trigger a fallback at all. For normal work — code, writing, research, analysis — you're getting the full Mythos-class model every time.

Fable 5 vs GPT-5.5

GPT-5.5 shipped in late April and has been the default frontier model for a huge share of engineering teams since. Fable 5 just made that default look shaky in three places.

Coding: Not Even Close

On SWE-bench Pro — the benchmark that tests end-to-end resolution of real GitHub issues — Fable 5 scores 80.3% against GPT-5.5's 58.6%. That's not a leaderboard squabble over decimal points like the Opus 4.7 vs GPT-5.4 race back in April. That's a 21.7-point gap on the hardest coding benchmark that exists.

Fable 5 also posts 95.0% on SWE-bench Verified, and Cognition reports it's the highest-scoring model ever on their FrontierBench evaluation. Cursor's team says it "opened up a class of long-horizon problems that were out of reach."

Hallucinations: The Stat Nobody's Talking About

On the independent AA-Omniscience hallucination benchmark, the Claude family posts a 36.18% hallucination rate against GPT-5.5's 85.53%, with Gemini sitting at 49.87%. Benchmarks measure hallucination differently and no model is close to perfect — but a gap that wide changes how much you can trust unverified output, which matters a lot more than a few SWE-bench points if you're using AI for client work, legal summaries, or anything where confidently wrong answers cost real money.

Pricing: Where GPT-5.5 Punches Back

Fable 5 costs $10 per million input tokens and $50 per million output tokens. GPT-5.5 is $5/$30. So OpenAI's model is roughly half the price — and for high-volume, low-stakes workloads, that math still favors GPT-5.5. Anthropic's counterpoint: Fable 5 is less than half the cost of the Claude Mythos Preview it replaces, and in at least one physics research case it matched four days of GPT-5.5 output in 36 hours using a third of the reasoning tokens. Cheaper per token isn't the same as cheaper per finished task.

Fable 5 vs Opus 4.8

Opus 4.8 was Anthropic's flagship until this morning, and it's still excellent — it's literally the fallback model inside Fable 5. So is the upgrade worth double the price?

On knowledge work the gap over Opus 4.8 looks modest — 1932 vs 1890. The differences show up at the edges:

• Document reasoning (GDPpdf, no tools): 29.8% vs Opus 4.8's 22.5% — a 32% relative improvement on pulling answers out of dense PDFs.
• Memory: with persistent file-based memory, Fable 5's improvement on long-horizon tasks was 3x larger than the same setup gave Opus 4.8. If you run agents that take notes and come back to them, this compounds.
• Vision: state-of-the-art across the board. Fable 5 completed Pokémon FireRed using vision alone — prior Claude models needed helper tools to get through it.
• Autonomy: it self-checks and validates its own work over extended runs, which is the difference between an agent you babysit and one you leave overnight.

The honest answer: if your work is short prompts and quick answers, Opus 4.8 at $5/$25 is still the value play. Fable 5 earns its premium on long, complex, multi-hour tasks where it finishes work the cheaper models can't.

Head-to-Head Benchmark Summary

Sources: Anthropic, Tom's Hardware, Digital Applied. Em dashes indicate published scores not available for direct comparison.

The Real-World Stories

Benchmarks are one thing. The launch-day case studies are the part that actually made me sit up:

• Stripe used Fable 5 to migrate a 50-million-line Ruby codebase in one day — work they estimated at two months of manual engineering. They described compressing months of engineering into days.
• Physics researchers got Fable 5 to reproduce in 36 hours what GPT-5.5 had reached after four days — on a third of the reasoning tokens.
• Life sciences teams had it design protein complexes for drug targets, and in blind comparisons its novel molecular biology hypotheses were preferred about 80% of the time over Opus-class output. It also ran autonomous genomics research across 138 species and outperformed published models from the journal Science.
• GitHub says performance "exceeded previous benchmarks," and it's already generally available in GitHub Copilot as of today.

The Catch

Per the house rules around here, the tradeoffs get their own section.

• It's expensive. $10/$50 is double GPT-5.5 and double Opus 4.8. For a lot of everyday work, the cheaper models are still the rational choice.
• The free window closes June 22. Fable 5 is included on Pro, Max, Team, and Enterprise plans through June 22 only. From June 23, it requires usage credits until Anthropic's capacity catches up with demand. If you want to evaluate it on your subscription, this is the two-week window.
• The bio/chem classifier is over-eager by design. Anthropic says so themselves. If you work in biomedical fields, expect legitimate requests to occasionally fall back to Opus 4.8 until the classifier is tuned.
• Business customers get a 30-day data retention requirement on Mythos-class models — used only for safety purposes with all human access logged, but it's a policy change worth knowing about before you route sensitive client data through it.
• The API surface is stricter. Fable 5 only supports adaptive thinking — manual thinking budgets, temperature, top_p, and even explicitly disabling thinking all return errors. If you're migrating code from older Claude models, budget a small cleanup pass.

Which Model Should You Actually Use?

Use Claude Fable 5 if…

• You're running long-horizon agents, large migrations, or overnight autonomous coding runs
• Hallucination rate matters — client deliverables, research, anything fact-sensitive
• You do heavy document analysis or vision-dependent work
• The cost of a failed or wrong result outweighs 2x token pricing

Use Claude Opus 4.8 if…

• Your tasks are short and interactive rather than long and autonomous
• You want most of the capability at half the price
• You're in a field that trips Fable 5's conservative classifiers — you'd be talking to Opus 4.8 anyway

Use GPT-5.5 if…

• Cost per token is your primary constraint at scale
• Your team is already deep in the Codex/OpenAI ecosystem and the switching cost is real
• Your workload doesn't hit the coding or hallucination gaps where Fable 5 dominates

Bottom Line

Back in April, the frontier race was being decided by single-digit margins — Opus 4.7 edged GPT-5.4 by a few points and we all argued about whether it mattered. Fable 5 ends that conversation. A 21.7-point coding gap and a hallucination rate less than half of GPT-5.5's isn't a leaderboard shuffle; it's a generation gap. The asterisks are real — double the price, a credit wall after June 22, and classifiers that occasionally hand you Opus 4.8 instead — but Anthropic just shipped the model it was previously only willing to give to security researchers, and it shows. If you build with AI for a living, you owe it to yourself to test it before the free window closes.

Sources: Anthropic, CNBC, Tom's Hardware, VentureBeat, GitHub Changelog, Digital Applied