Something significant happened yesterday — and most founders still haven’t processed it.
Anthropic launched Claude Fable 5 on June 9, 2026. Not another model update. Not a fine-tuned variant. The first publicly available Mythos-class AI — the same tier of intelligence that the U.S. government quietly restricted two months ago because it was deemed too powerful for general release. It’s now available to any developer with a credit card.
This is not hype. Here’s everything you actually need to know.
1. What “Mythos-Class” Actually Means
To understand why this launch is different, you need a two-minute history lesson.
In April 2026, Anthropic unveiled Claude Mythos Preview — a model so capable in cybersecurity that they refused to release it publicly. Instead, it went into a classified program called Project Glasswing, made available only to a vetted circle of cyber defenders, critical infrastructure operators, and government contractors. U.S. intelligence agencies were watching.
Claude Fable 5 is that same model — same training, same weights — deployed with surgical classifiers that redirect the highest-risk capabilities (offensive cybersecurity, biological synthesis, certain chemistry queries) to Claude Opus 4.8 instead. Everything else? Full Mythos power, open to everyone.
Anthropic created a brand new tier above Opus to house it. The old hierarchy was Haiku → Sonnet → Opus. Now there’s a fourth level: Mythos-class. Fable 5 sits there — and it’s not close.
2. What Fable 5 Can Actually Do
The demos are real. The third-party testing is credible. Here’s what matters for builders.
Software engineering: Stripe fed Fable 5 a 50-million-line Ruby codebase and asked it to perform a full codebase-wide migration. It finished in one day. Stripe estimated the same job would take a full engineering team over two months by hand. That’s not a marginal efficiency gain — that’s a structural change in how you think about your engineering headcount.
Knowledge work: IMC Trading — one of the most analytically rigorous firms on the planet — reported that Fable 5 aced their trading-analysis evaluations nearly across the board. Factual lookup, conceptual reasoning, root-cause analysis, expected-value analysis. All of it. Hebbia’s Finance Benchmark for senior-analyst-level reasoning? Fable 5 scored highest of any model tested.
Vision: Fable 5 can rebuild a web app’s complete source code from screenshots alone — no API docs, no schema files. It also beat Pokémon FireRed using only raw game screenshots with a minimal harness, something previous Claude models couldn’t do even with heavy scaffolding. The implication for production AI workflows: Fable 5 needs dramatically less infrastructure to get vision tasks done.
Memory and long-context: With a 1 million token context window and persistent file-based memory, Fable 5 can operate autonomously for days without losing the thread. Anthropic tested this with a complex strategy game — Fable 5’s performance improved three times more from memory access than Opus 4.8 did under identical conditions. For founders building agents: this is the sleeper feature with the highest enterprise value.
3. The Benchmarks That Actually Matter
Benchmark tables can mislead — vendors test their own models, harnesses vary. Read these directionally, not as gospel.
On SWE-Bench Pro (real software engineering tasks): Fable 5 scores 80.3%. GPT-5.5 scores 58.6%. Gemini 3.1 Pro scores 54.2%. That is a 22-point gap between Fable 5 and its closest competitor on the benchmark that most closely mirrors actual production code work.
On FrontierCode Diamond (production-quality code, not just passing tests): Fable 5 scores 29.3%. GPT-5.5 scores 5.7%. That is a 5× gap. If you’re shipping real software, that number matters more than any other on the benchmark sheet.
On Humanity’s Last Exam (graduate-level multi-domain reasoning): Fable 5 scores 59% without tools, 64.5% with. GPT-5.5 scores 41.4% without tools. The gap is wide and consistent across domains.
One important caveat most articles are missing: some of Anthropic’s benchmark table entries show Mythos 5 scores, not Fable 5. On cybersecurity and biology rows marked with an asterisk (*), the figure belongs to the restricted model you cannot access. On those rows, Fable 5 performs closer to Opus 4.8 because the safety classifiers redirect those queries. When you see someone claim Fable 5 dominates security benchmarks, they are benchmarking a model that isn’t available to them.
4. The Developer Setup — What You Need to Know
Copy this. Screenshot it. Send it to your tech lead.
- API Model ID:
claude-fable-5 - AWS Bedrock ID:
anthropic.claude-fable-5 - Context window: 1M tokens (default)
- Max output: 128k tokens per request
- Pricing: $10/M input · $50/M output (90% discount on cached tokens)
- Data retention: 30-day (not available under zero data retention)
- Available on: Claude API, Amazon Bedrock, Vertex AI, Microsoft Foundry, GitHub Copilot
Two breaking changes if you’re migrating from Opus 4.8:
First, Adaptive Thinking is always on and cannot be disabled. If your existing integration passes thinking: {"type": "disabled"}, that parameter is not supported on Fable 5 — you’ll need to remove it. Use the effortparameter to control thinking depth instead.
Second, refusals come back as HTTP 200, not errors. When the safety classifier triggers, the API returns stop_reason: "refusal" as a successful response. Your current error-handling logic will miss it. Add explicit refusal checking, and configure the fallbacks parameter to automatically retry on Opus 4.8 — Anthropic has a beta server-side fallback that handles this for you.
Also worth noting: the 30-day data retention is a hard constraint. If you operate in healthcare, legal, or defense with zero-retention contracts, Fable 5 is not available to you under those terms. Plan before you build.
5. Pricing & The Free Window Closing Fast
There is a deadline and most people don’t know it.
Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026. After June 23, usage requires credits. You have 12 days of free access to run real benchmarks on your actual workflows — before you commit to the token spend.
At $10/M input, Fable 5 costs twice what GPT-5.5 does on input tokens. But the 90% prompt caching discount changes the math entirely for applications with stable context — repeated codebases, contract libraries, product docs. Cached input tokens cost $1/M, not $10. Build your architecture around caching from day one and the pricing becomes genuinely competitive.
The strategic calculation: if Fable 5 completes a task in one pass that previously took three passes on a cheaper model, it’s not more expensive — it’s cheaper in total. The right approach is routing: Fable 5 for long-horizon, complex work; Opus 4.8 or Sonnet for everything else. Flat-routing everything through Fable 5 is how you overspend.
6. What Fable 5 Means for Your Startup — by Category
DevTools & engineering startups: Agentic coding pipelines that previously needed human checkpoints at every multi-file step now have a model that can run unsupervised across longer horizons. The Stripe migration example isn’t a demo — it’s a new reference point for what’s possible.
LegalTech: On the Legal Agent Benchmark, Fable 5 scores 13.3%. GPT-5.5 scores 2.1%. Gemini 3.1 Pro scores 0.0%. If you’re building in contracts, compliance, due diligence, or regulatory intelligence, there’s no close second right now.
FinTech: The IMC Trading and Hebbia results are commercial signals, not academic ones. Multi-document financial reasoning, chart extraction, root-cause analysis on complex data — Fable 5 performs at a level previous models couldn’t sustain without hallucinating critical numbers.
Sales & GTM automation: Persistent memory plus million-token context plus multi-day autonomous sessions equals a qualitatively different class of sales agent. One that remembers every account interaction, tracks context across weeks, and reasons about pipeline in ways that weren’t possible before last week.
BioTech / HealthTech: The headline drug-discovery results belong to Mythos 5, which is still locked behind Project Glasswing. But Fable 5’s vision improvements — extracting precise numbers from complex scientific figures — are immediately usable for clinical data parsing, lab report analysis, and research document workflows.
The Bottom Line for Founders
Claude Fable 5 is the first AI model that crossed from “impressive” to “structurally changes what a small team can build.” The Stripe example alone — two months of engineering, compressed to one day — should prompt every founder to rethink their hiring timeline, their product roadmap, and their moat.
The benchmarks are real. The free access window closes June 22. The breaking API changes will catch teams off guard if they don’t prepare. And the classifiers firing on 5% of sessions means anyone building at scale needs fallback routing configured from day one.
Run Fable 5 on your hardest workflow this week. The data you get in the next 12 days will tell you more than any benchmark table.
Related Articles on IMFounder
- Google’s $15 Billion AI Data Centre Raises Water Concerns in a Country Where over 600 Millions Already Face Water Stress
- The Shocking Cost of ChatGPT and Claude Inside Modern Companies
- 7 Explosive AI Updates in May 2026 That Will Shock Every Founder
- Founder Review: Get Expert Feedback Before You Build
- Kimi AI vs Claude AI
- Anthropic Mythos Preview: Revolutionary New AI Model from Anthropic
- Apple’s “Unbreakable” M5 Chip Was Cracked by an AI in Just 7 Days
Sources: Anthropic official announcement · Claude API Docs · Project Glasswing
