Claude Fable 5 Explained: Anthropic's Most Powerful Public AI — Benchmarks, $10/$50 Pricing and The Refusal Backlash
Anthropic's Claude Fable 5 is the first Mythos-class AI the public can use. We break down the 80.3% SWE-bench Pro score, the Stripe 50-million-line migration, the $10/$50 pricing, and why developers are split on the safety filters.
On June 9, 2026, Anthropic quietly flipped a switch that the AI industry has been waiting on for two years: it released Claude Fable 5, the first "Mythos-class" model the general public can actually rent. Within 72 hours it became the most-searched AI product in America, the top trending term on Hacker News, and the subject of a viral 50-million-line Stripe case study that engineering leaders are still arguing about.
If you only have thirty seconds, here is the short version: Fable 5 scores 80.3% on SWE-bench Pro versus 58.6% for GPT-5.5, costs $10 per million input tokens and $50 per million output tokens, and refuses roughly 1 in 20 prompts. Everything else below is the detail behind those four numbers — what the model can do, what it costs in the real world, and why a vocal slice of developers is furious about the guardrails.
Why this matters: Fable 5 is the first time a frontier lab has shipped a "too dangerous to release unfiltered" model to anyone with a credit card. The trade-off it forces — capability versus caution — is the story that will define the Claude 5 family's first year.
What exactly is Claude Fable 5?
Fable 5 and Claude Mythos 5 are the same underlying model. The only difference is the guardrails.
- Fable 5 (public) ships with safety classifiers that detect queries touching cybersecurity, biology and chemistry, and model-distillation. Flagged prompts are silently routed to the less capable Claude Opus 4.8 instead of the frontier model.
- Mythos 5 (restricted) is the unrestricted version, available only to approved organizations under Anthropic's Project Glasswing partner program — pharma, biosecurity, defense, and a handful of academic labs.
Anthropic says the safeguards are tuned conservatively and trigger in under 5% of sessions on average. The release came just days after the company publicly warned that frontier AI capabilities were becoming dangerous — a juxtaposition TechCrunch called out directly in its launch coverage.
The benchmarks: how Fable 5 actually compares
The headline number is SWE-bench Pro, the eval that measures whether a model can complete difficult real-world software-engineering tasks end-to-end.
| Model | SWE-bench Pro | FrontierBench | Context window |
|---|---|---|---|
| Claude Fable 5 | 80.3% | State-of-the-art | 1,000,000 tokens |
| Claude Mythos 5 | ~82% (1–3 pts higher) | SOTA | 1,000,000 tokens |
| OpenAI GPT-5.5 | 58.6% | Trails Fable 5 | 400,000 tokens |
| Claude Opus 4.8 | ~61% | Trails Fable 5 | 200,000 tokens |
Fable 5 also tops Cognition's FrontierBench coding eval and posts state-of-the-art results across knowledge work, vision, and scientific reasoning. The 1-million-token context window reads diagrams, charts, and tables nested inside PDFs — a big deal for finance, legal, and analytics teams that have spent eighteen months stitching together hacky chunking pipelines.
On most benchmarks Fable 5 and Mythos 5 score within 1–3 points of each other. The gap widens only in the restricted domains, where Fable 5's filters pull its effective score down toward Opus 4.8.
Real-world case studies from launch week
The launch-week evidence is unusually concrete. Three case studies stand out:
1. Stripe's 50-million-line migration
Stripe reported that Fable 5 compressed months of engineering into a single day, running a code migration across a 50-million-line Ruby codebase without human babysitting. The internal engineering post called it "the first migration we didn't dread."
2. Autonomous research without being asked
One developer had Fable 5 build a fully researched isochrone travel-time map via Claude Code. The model spun up cheaper sub-agents on its own initiative, pulling data on more than 2,200 flight routes plus train schedules and road speeds from academic papers — without being told to. The agentic behavior, not the map, is what made the demo go viral.
3. Drug design (the Mythos side)
Anthropic's internal protein-design experts reported roughly a 10x acceleration in parts of the drug-design pipeline using Mythos 5, with the model choosing binding sites, running design tools, and recovering from its own failures unassisted. Nine of 14 protein targets yielded strong drug-design candidates.
Veteran developer Simon Willison summed up the early consensus in one line: the model is "something of a beast" — slow and expensive, but handling everything thrown at it, including problems users said they had been avoiding for months.
The price tag: $10 in, $50 out
Fable 5 costs $10 per million input tokens and $50 per million output tokens — less than half of what Mythos Preview cost, but still the most expensive major AI model on the market.
| Plan | Cost | What you get |
|---|---|---|
| API — input | $10 / 1M tokens | Cheapest "frontier" tier on the market for the capability |
| API — output | $50 / 1M tokens | Highest sticker price of any major API |
| Prompt cache | 90% discount on cached input | Unchanged from Claude 4 family |
| Pro / Max / Team | Included | No extra cost through June 22, 2026 |
| Enterprise | Seat-based | Also live on AWS Bedrock |
After the June 22 cutoff, consumer plans start drawing on usage credits. If you want to evaluate Fable 5 on a real workload, the window between now and then is effectively a free trial — and most engineering teams we've spoken to are using it that way.
The backlash: refusals, price, and "is it really that much better?"
Not all the launch coverage is glowing. Three critiques keep surfacing:
- Innocuous refusals. The Register reported cases of Fable 5 refusing benign prompts — collateral damage from safety classifiers that err toward caution. A security researcher asking about a public CVE got routed to Opus 4.8; so did a high-school student asking about photosynthesis.
- Sticker shock. The Decoder's verdict captured the skeptics' framing in five words: "powerful, expensive, and heavily filtered." At $50 per million output tokens, a single long agentic run can cost more than a developer's daily salary.
- Incremental, not revolutionary. A contingent of developers argues the jump over Opus 4.8 is real but smaller than the benchmark deltas suggest, raising the perennial question of benchmark overfitting.
The trade-off is the story. Anthropic is betting that mainstream users will accept a sub-5% refusal rate as the price of accessing a model it considers too capable to release unfiltered. Whether the market agrees — at $50 per million output tokens — is the question that will define the next twelve months.
Should you use Fable 5? A decision guide
- Use Fable 5 if your work is long-horizon coding, document-heavy analysis, agentic workflows, or anything where the 1M-token context window unlocks a workflow you previously couldn't run. The Stripe case study alone justifies a trial before the June 22 credit cutoff.
- Skip Fable 5 if you mostly write emails, summaries, or short chat replies. Opus 4.8 or a $20 consumer chatbot does the job for a fraction of the cost — and won't randomly refuse a prompt because a classifier got nervous.
- Wait and watch if you're in regulated industries (healthcare, legal, defense). The refusal behavior is still being tuned week-to-week and the bills can be unpredictable.
Frequently asked questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic's most powerful publicly available AI model, released on June 9, 2026. It is the public-facing version of the "Mythos-class" tier that was previously restricted to research partners, with safety classifiers that route sensitive queries to the smaller Claude Opus 4.8 model.
What's the difference between Fable 5 and Mythos 5?
They are the same underlying model. Mythos 5 has no safety routing and is only available to vetted organizations under Project Glasswing. Fable 5 is identical but adds classifiers that catch cybersecurity, bio/chem, and distillation queries and downgrade them to Opus 4.8.
How much does Claude Fable 5 cost?
$10 per million input tokens and $50 per million output tokens via the API. Prompt caching keeps the existing 90% input discount. Consumer Pro, Max, Team, and Enterprise seat plans include Fable 5 at no extra cost through June 22, 2026, after which usage draws on credits.
How does Fable 5 compare to GPT-5.5?
On SWE-bench Pro, Fable 5 scores 80.3% versus 58.6% for GPT-5.5 — a roughly 22-point gap on real-world coding tasks. Fable 5 also has a larger context window (1M vs 400K tokens) and stronger PDF and vision reasoning, but it is more expensive per token.
Why does Fable 5 refuse some prompts?
Fable 5 uses safety classifiers that trigger on cybersecurity, biology/chemistry, and model-distillation queries. When triggered, the prompt is silently routed to the less capable Claude Opus 4.8. Anthropic reports the classifiers fire in under 5% of sessions, but The Register documented cases where benign prompts were flagged as false positives.
Is Claude Fable 5 worth it for developers?
For long-horizon coding, large-codebase migrations, and agentic workflows — yes, the Stripe 50M-line case study and an 80.3% SWE-bench Pro score make it the strongest public coding model available. For routine chat, summarization, or short-context tasks, Opus 4.8 or a $20 plan is the better economic choice.
Related reading: see our ranked breakdown of the best AI chatbots of 2026 for how Fable 5 stacks up against GPT-5.5, Gemini 3 Ultra, and the open-weight challengers.
Reporting by The Daily Federal Newsroom. Sources: Anthropic, TechCrunch, VentureBeat, CNBC, Simon Willison, Vellum, Finout, The Register, The Decoder, AWS.