Yesterday, June 9, 2026, Anthropic let its most powerful model loose on the streets. It's called Claude Fable 5, and it's not just a juiced-up Opus — it's the first Mythos-class model available to the public, a category they unveiled in April under lock and key that until yesterday only existed behind guarded doors.
The launch comes with fine print. A lot of it.
Fable 5 ships on a leash. Built-in safety classifiers that block queries about offensive cybersecurity, dual-use biology, and model distillation, automatically rerouting them to Opus 4.8. It's like selling someone a Ferrari with an electronic speed limiter that kicks in whenever the GPS sniffs a school zone.
The move is ambitious and a little schizophrenic, landing right when Anthropic is in the eye of the storm: they just confidentially filed their IPO prospectus with the SEC, are valued at $965 billion with a $47 billion revenue run rate, and five days ago published a report basically saying AI is on the verge of learning to improve itself without asking us first.
Welcome to 2026, the year everything happens at once.
Mythos: The Monster They Kept in the Vault
In April 2026, Anthropic unveiled Mythos, a model so good at finding and exploiting software vulnerabilities that they decided not to release it publicly. They gave it only to a select group of critical infrastructure companies under a program called Project Glasswing. The reasoning: its offensive hacking capabilities were way too dangerous for anyone not playing defense.
Wall Street went nuts. CrowdStrike and Palo Alto Networks started integrating Mythos into their workflows. Mozilla used it to find 271 bugs in Firefox. The message was crystal clear: forget the chatbot that explains recipes, this is a different beast entirely.
Two months later, they decided to open the door. With conditions.
Fable 5 and Mythos 5: Twins with Different Lives
They dropped two models the same day:
| Feature | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Who gets it? | You, me, anyone with an API key | The VIP club (Project Glasswing) |
| Safeguards | Active (cyber, bio, chemistry, distillation) | Lifted in several areas |
| Pricing | $10 / $50 per million tokens (input/output) | Same |
| API model ID | claude-fable-5 |
Not for sale |
| Context window | ~1M tokens | Same |
| Max output | ~128K tokens per response | Same |
Same underlying architecture. Different clearance levels. Fable 5 is the one they sell you with insurance included. Mythos 5 is for those Anthropic already vetted and said "ok, you're good."
The Safeguards: Shield or Straitjacket?
Fable 5 ships with safety classifiers baked into the model itself — not an external wrapper. When it detects something off, it flat-out refuses and hands the call over to Opus 4.8.
What you won't be able to ask it:
- Offensive hacking (exploiting vulnerabilities, writing malware)
- Dangerous biology and chemistry (synthesizing toxins, pathogens)
- Model distillation (using Fable's responses to train another model)
Anthropic says this affects less than 5% of sessions and that they ran over 1,000 hours of red-teaming without anyone finding a universal jailbreak. They also admit the classifier errs on the side of caution — some perfectly legitimate queries will bounce to Opus 4.8 because the system plays it safe.
And there's a controversial twist: mandatory 30-day traffic retention, even if you previously had a zero-retention agreement. Anthropic swears the data won't be used for training, only to "defend against complex attacks and reduce false positives." But the precedent is uncomfortable: more power = less privacy.
The Numbers (Unofficial, but Juicy)
Anthropic hasn't released the system card yet. What we have are third-party benchmarks. Take these as directional, not gospel.
| Benchmark | Claude Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Pro (agentic coding) | 80.3% | 69.2% | 58.6% | 54.2% |
| Terminal-Bench 2.1 | 88.0% | — | 83.4% | — |
| Humanity's Last Exam (no tools) | 59.0% | — | 52.2% | — |
| FrontierCode Diamond | 29.3% | ~14% | — | — |
| Spatial reasoning | 38.6% | 14.5% | — | — |
SWE-Bench Pro at 80.3%. That's over 20 points above GPT-5.5. Not an incremental bump — a different league. In agentic coding, Fable 5 is playing alone.
What early testers are saying:
- Hex (analytics): first model to break 90% on their core analytics benchmark. "A 10-point jump over Opus."
- Cognition (creators of Devin): highest score ever on FrontierBench.
- CursorBench: state of the art. "Opened up long-horizon problems that were impossible with previous models."
- Base44 (vibe-coding): "One-shots complete apps. Tool calling is excellent."
- Hebbia (finance): "The strongest finance-first model we've tested."
- Rakuten: "At maximum effort, Fable reflects on and validates its own work. The extra thinking pays for itself."
Weren't We Supposed to Hit the Brakes?
On June 4 — five days before dropping Fable 5 — Anthropic published a report warning that AI systems are showing signs of being able to design, build, and train their own successors without human intervention. They call it Recursive Self-Improvement (RSI).
The same report calls on all major labs to establish a "coordinated and verifiable brake" on frontier model development. Dario Amodei has been on this beat for months: the window to act is closing.
The contradiction is glaring and Anthropic has been getting hammered with questions: how do you call for a global pause while simultaneously launching the most powerful model on the market? Diane Penn, head of product, answers with "race to the top." Releasing powerful technology with robust safeguards is better than leaving a void for someone without safety standards to fill.
Buy it or not — that's up to you.
Money: Twice the Price of Opus
| Input (1M tokens) | Output (1M tokens) | |
|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
Want US-only inference? 10% surcharge.
Until June 22, Pro, Max, Team, and Enterprise subscribers get Fable 5 at no extra cost. Starting June 23, it's on you.
Penn insists the final cost per solved task is lower because Fable needs fewer attempts. "You simply get higher ROI with more intelligent models." Sure, Diane. Let's see those June bills.
What's It Actually Good For?
Agents. Running in Claude Code or Managed Agents, Fable 5 can work for days without supervision: planning across stages, delegating to sub-agents, checking its own output. It's not a copilot — it's an employee you hand a project to and come back when it's done.
Coding. Large-scale migrations, complex implementations, refactors that used to need a team. Writes its own tests, implements designs with high fidelity, and uses vision to compare its output against what you asked for.
Enterprise. Deep research, document analysis, deliverables ready for review. You feed it a mountain of PDFs and it hands you back a report. You don't supervise — you review.
Vision. Reads diagrams, tables, and charts nested in PDFs. Also uses vision to check its own code output. Target market: finance, legal, analytics, architecture.
Should You Use It?
For day-to-day work, stick with Opus 4.8 or a fast model like Gemini 3.5 Flash. Cheaper, faster, gets the job done.
Fable 5 is for when the problem is genuinely hard and long-horizon: multi-hour autonomous refactors, complex agent pipelines, tasks where Opus 4.8 puts its hand up and gives up. At $10/$50 per million tokens, it's not your default — it's your secret weapon for when output quality justifies every cent.
And if your project even grazes offensive cybersecurity or synthetic biology, assume Opus 4.8 will answer half your queries.
The Full Board
Fable 5 didn't land on a quiet day:
- OpenAI confidentially filed its own IPO, valued at $852 billion.
- SpaceX/xAI debuts on the stock market this Friday.
- Google positioned Gemini 3.5 Flash as the fast, cheap workhorse.
- GPT-5.5 Pro competes on math and general knowledge, but hasn't published agentic coding numbers.
Fable 5 is the king of autonomous coding today. But this throne shifts every month.
The Bottom Line
Claude Fable 5 is ambitious, expensive, and ships with more restrictions than any Anthropic model before it. It's also, by the numbers we have, the most capable agentic coding model out right now.
Anthropic is juggling "this is dangerous, we need regulation" with "here, take it, buy it, use it now." If it works, Fable 5 will be remembered as the moment Mythos-class models went from corporate secret to work tool. If it doesn't, it'll be the perfect case study of why safety safeguards arrived too early — or too late.
The only certainty: the bar for autonomous coding just went up. Again.
Enjoyed this? Subscribe to the newsletter at guayoyo.tech or follow us on X.
Top comments (0)