By A 50-million-line Ruby codebase. Two months of team effort. One day with Claude Fable 5. That is the number Stripe put out after Anthropic’s June 9 release, and it is the kind of claim that either lands or crumbles under scrutiny. In this case, the benchmarks back it up. Fable 5 sits at 80.3% on SWE-Bench Pro — 11 points ahead of the nearest competitor and 22 points ahead of GPT-5.5. That is not a marginal lead. That is a tier shift.
What Fable 5 Actually Is
Fable 5 is Anthropic’s first Mythos-class model to reach general availability. It does not slot into the Opus family — it sits above it, built on the same base model as the restricted Claude Mythos 5. The distinction matters: Mythos 5 operates without a safety classifier ceiling and is gated to vetted Project Glasswing partners — authorized penetration testers and biomedical researchers. Fable 5 is that same model with three active safety classifiers covering cybersecurity exploitation, biology/chemistry dual-use, and distillation attacks. Fewer than 5% of sessions trigger a fallback. When one fires, you get Claude Opus 4.8 on that request — not an outright refusal — and you are billed at Opus rates.
Under the hood: a 1M token context window, chain-of-thought reasoning enabled, and the same API integration path as existing Claude models. Swap in claude-fable-5 and you are there — no SDK changes required. The model is available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot, where it landed the same day as the release.
The Benchmark Story
The coding numbers are the headline, so here they are:
| Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 |
|---|---|---|---|
| SWE-Bench Pro | 80.3% | 69.2% | 58.6% |
| FrontierCode Diamond | 29.3% | 13.4% | 6.3% |
| Terminal-Bench 2.1 | 88.0% | — | 82.7% |
The FrontierCode Diamond gap is the one worth flagging: 4.6x better than GPT-5.5 on the hardest real-world coding tasks. SWE-Bench measures whether a model can resolve GitHub issues autonomously. FrontierCode Diamond pushes that into production-grade complexity — multi-file refactors, architecture decisions, dependency-aware changes. That is the gap developers on Hacker News noticed first: a 46x reduction in memory allocations during a database migration, bugs caught that Opus 4.8 and GPT-5.5 both missed, a CRDT fuzzer written unprompted to verify correctness. One developer put it plainly: “the first model that feels like it’s coming for my job.”
One important caveat Anthropic buries in the footnotes: their benchmark table shows “the higher of Fable 5 and Mythos 5 scores” for security and biology categories. Fable 5’s actual cybersecurity performance drops near Opus 4.8 territory once the classifiers engage. The security headlines belong to Mythos 5 — which most developers will not have access to.
The Pricing Reality Check
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Claude Fable 5 | $10 | $50 |
| Claude Opus 4.8 | $5 | $25 |
| GPT-5.5 | $5 | $30 |
Fable 5 is exactly double Opus 4.8 on both token types. A 90% prompt caching discount applies to input tokens, which helps in high-repetition agentic workflows, but the output premium stays. The routing strategy that makes sense: deploy Fable 5 where task complexity means Opus 4.8 would need two or more retry passes. Single-pass Fable 5 often beats two-pass Opus 4.8 on total cost once you account for the token volume from repeated attempts. For PR reviews, quick explanations, routine classification — anything that does not need sustained autonomy over a complex task — stick with Opus 4.8 or GPT-5.5. Fable 5 is not a drop-in replacement for your entire API budget.
Where It Earns the Premium
Community testing makes the use-case boundary clearer than Anthropic’s marketing does. Fable 5 demonstrably pulls ahead on:
- Large codebase migrations requiring multi-step reasoning and cross-file awareness
- Long-horizon agentic tasks that previously needed human checkpoints
- Vision-based frontend work — it can rebuild a web application from a screenshot alone
- Document-heavy analysis: finance, legal reasoning, and medical records where a single pass needs to be right
The pattern that keeps surfacing in community testing: the longer and more open-ended the task, the wider the lead. Evaluate it on quick Q&A and you will conclude it is an expensive Opus. That is a fair conclusion for quick Q&A — it is just the wrong evaluation for the model’s actual design target.
The Safety Fallback You Will Hit
The classifier system is worth understanding before you commit a use case to Fable 5. The three trigger domains — cybersecurity, biology/chemistry, distillation — sound narrow but have real surface area. Developers reported false positives on health data analysis, MRI image segmentation, lab automation protocols, and biology coursework. These are legitimate professional tasks. The fallback to Opus 4.8 is better than a hard refusal, but you lose the capability jump you paid for, and there is no explicit warning in the response when a fallback occurs. Plan for this in any health or research-adjacent product.
How to Access It
The integration path is frictionless for existing Claude users. The model identifier is claude-fable-5 — drop it into your API call and it works. Full details on API parameters and advanced configuration are in TrueFoundry’s technical guide. GitHub Copilot users have access today — Fable 5 landed in Copilot on June 9 and handles the complex tasks that Copilot routes to frontier models. For teams managing cost governance across model tiers, an AI gateway layer makes routing by task complexity much cleaner than manual prompt engineering. A detailed head-to-head comparison with GPT-5.5 is worth reading before you commit your routing strategy.
Data policy note: Anthropic uses a 30-day retention window on Mythos-class traffic, does not train on this data, and deletes it after the retention period in almost all cases.
The Bottom Line
Claude Fable 5 is a genuine capability step for complex, long-horizon coding work. The SWE-Bench lead is real. The Stripe result is real. The community results are real. What is also real: the safety classifier friction on health and security tasks, the $50/M output price, and the fact that it performs similarly to Opus on tasks that do not challenge it. Selective deployment is the skill here. Route hard, complex, multi-step tasks to Fable 5. Route everything else to models that cost half the price. That is the workflow shift that actually matters — not the model, the routing strategy. For a broader look at where this sits in the current frontier model landscape, the picture is one of genuine differentiation at the top of the coding tier for the first time in a while.
