By Claude Fable 5 launched today. It is Anthropic’s first publicly available Mythos-class model — the same architecture that found 271 Firefox vulnerabilities in a single month — now available on the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. It scores 80.3% on SWE-Bench Pro. That is 11 points higher than Claude Opus 4.8 and more than 20 points above GPT-5.5. If you build with Claude, today changes your benchmark baseline.
Fable 5 Is Not Mythos 5 — Know the Difference Before You Build
Anthropic released two models today. Claude Fable 5 is publicly available with safety classifiers that block high-risk queries. Claude Mythos 5 is the same underlying model with safeguards lifted, and it remains locked to Project Glasswing partners (AWS, Apple, Broadcom, Cisco, Google, JPMorgan, Microsoft, NVIDIA, and a handful of others) and select biology researchers.
The 271 Firefox zero-days headline you may have seen? That was Mythos 5, running with cybersecurity restrictions removed. Fable 5 cannot replicate that — offensive vulnerability research is explicitly in the blocked category. If you went into this announcement expecting a general-purpose offensive security tool, that is Mythos territory, and Mythos is not for public sale.
What Fable 5 can do is handle defensive code review, complex codebase refactors, financial document analysis, and long-horizon autonomous coding tasks better than any publicly available model in existence. That is a broad, practical mandate for the vast majority of developer workloads.
The Benchmarks Are Convincing
SWE-Bench Pro tests real GitHub Issues resolved autonomously — no hand-holding, just results. Here is where things stand today:
| Model | SWE-Bench Pro | Input (per 1M tokens) |
|---|---|---|
| Claude Fable 5 | 80.3% | 0 |
| Claude Opus 4.8 | 69.2% | |
| GPT-5.5 | 58.6% | |
| Gemini 3.1 Pro | 54.2% |
On the harder FrontierCode Diamond benchmark — designed to avoid training data leakage — Fable 5 scores 29.3% versus 13.4% for Opus 4.8 and 5.7% for GPT-5.5. The pattern is consistent: the harder the task, the larger the gap. This is the model for problems that previously required multiple passes and heavy human correction.
What the Safeguards Actually Block
Fable 5’s safety classifiers watch for queries in three domains: cybersecurity (offensive research, exploit development), biology (synthesis, enhancement), and chemistry (distillation pathways). When a classifier triggers, the model blocks its own response and falls back to Opus 4.8, returning a safety_refusal stop reason.
Anthropic says fewer than 5% of sessions hit this fallback. For standard business, coding, and analysis workflows, you will not encounter it. Security developers doing defensive code review — auditing dependencies, checking for SQL injection, reviewing auth logic — should be fine. Penetration testing tooling and active vulnerability research cross the line.
Handle refusals explicitly in your implementation:
if response.stop_reason == "safety_refusal":
# Log it, fall back to Opus 4.8, or surface a clear message to the user
passHow to Access Claude Fable 5 Now
The model ID is claude-fable-5. It is available today on the Claude API, Claude Code CLI, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Through June 22, it is included in all paid Claude plans — Pro, Max, Team, and seat-based Enterprise — at no extra cost. On June 23, it moves to usage-credit billing.
Migrating from Opus 4.8 is three changes:
response = client.messages.create(
model="claude-fable-5", # Step 1: update model ID
max_tokens=8192,
messages=[{"role": "user", "content": "..."}],
thinking={"type": "adaptive", "effort": "high"} # Step 2: replace budget_tokens with effort
)
if response.stop_reason == "safety_refusal": # Step 3: add refusal handling
# Fall back or inform user
passRemove any thinking: disabled logic — Fable 5 does not support disabling thinking. Start effort at high for most tasks that previously ran at xhigh on Opus 4.8, then tune down for cost optimization. The full migration guide is in the API docs.
The Pricing Math
At $10 per million input tokens and $50 per million output tokens, Fable 5 costs twice as much as Opus 4.8. With 90% prompt caching on input, cached tokens drop to $1/million. The honest question: does your workload benefit from the jump from 69% to 80% on SWE-Bench Pro? For long-horizon agentic tasks, autonomous code migrations, and complex document analysis, the answer is almost always yes. For quick queries or lightweight generation, stay on Opus 4.8 or Sonnet.
Early enterprise results back this up. Stripe ran a 50 million-line Ruby migration in a single day using Fable 5 on Claude Code — previously scoped as a two-month project. Hex hit its first score above 90% on its internal analytics benchmark. Hebbia took top marks on its finance reasoning suite. Cursor named it the top performer on CursorBench.
The full Anthropic announcement covers both models in detail. For AWS customers, the Bedrock launch post has region availability and IAM setup. The models overview page has the full context window and token limit specs.
