Anthropic's Claude Fable 5 Beats the Hardest Coding Benchmark in Under 2 Hours
Mechanize's GBA Eval benchmark shows Claude Fable 5 hitting 74.5% — beating Opus 4.8's 24-hour score in under 2 hours, with a notable audio regression.
Takeaways
- New record: Claude Fable 5 scores 74.5% on GBA Eval, the highest score to date, beating Opus 4.8's 24-hour score in under 2 hours.
- Gameplay dominance: Fable 5 scores 91.5% on gameplay accuracy (vs. Opus 4.8's 77.4%) and is the first model to achieve perfect play on Varoom 3D.
- Audio regression: Fable 5 scores only 44.5% on audio — worse than Opus 4.8 (69.1%) and GPT-5.5 (58.9%), a clear step backward.
- What GBA Eval is: Models get 24 hours in a Docker container to write a GBA emulator in Rust/WASM from scratch, graded frame-by-frame against the Mesen2 reference emulator.
- Pricing and access: Fable 5 costs $10/M input tokens and $50/M output tokens, available via
claude-fable-5on the Claude API, AWS, GCP, and Azure — but requires 30-day data retention. - Persistent failure: No model tested so far, including Fable 5, can build an emulator that correctly runs Spout, a homebrew cave-flying game, scoring just 7.6%.
Anthropic's Claude Fable 5 just set a new record on one of the most demanding coding benchmarks in existence: writing a Game Boy Advance emulator from scratch. Mechanize, the AI evaluation company, ran Fable 5 through its GBA Eval and the results are striking , both for what the model got right, and for one surprising regression.
The hardest coding test you've never heard of
GBA Eval is not a typical benchmark. Models are tasked with writing, from scratch, a Game Boy Advance emulator in Rust that compiles to WebAssembly. That alone is a multi-week project for a skilled human engineer. The grading is equally rigorous: accuracy is measured using a combination of existing open-source test suites and test cases of gameplay patterns in real ROMs, evaluated using a custom harness.
The key insight that makes this benchmark fair is the GBA's hardware determinism. This kind of grading is tractable because the GBA console itself has no entropy source , there is no RTC, wall clock, or analog input on the console. That means the emulator's output on any given input sequence is perfectly reproducible, so Mechanize can compare the model's emulator frame-by-frame against Mesen2, a reference-quality open-source GBA emulator.
The setup each model gets is specific: each model gets a Docker container with the Rust and wasm32 toolchain, the ABI specification, a BIOS stub, dev ROMs, and an oracle CLI , a black-box wrapper around Mesen2 that the model can run on any ROM with any input sequence to see Mesen2's behavior. The model cannot read Mesen2's source and does not have access to the internet. Each model runs for 24 hours, with checkpoints taken every 15 minutes.
74.5% and a two-hour knockout
Claude Fable 5 scored 74.5% on GBA Eval, the highest score any model has achieved. But the more telling number is the speed: it beat Opus 4.8's 24-hour score in under 2 hours. For context, Claude Opus 4.8 scored 70.9% on GBA Eval , given 24 hours, it writes an emulator that plays most games with working audio on all of them, and it beat the previous best (GPT-5.5 at 53.2%) in under an hour. Fable 5 has now leapfrogged that bar entirely.
The breakdown by category tells a more nuanced story:
- Overall score: 74.5% (previous best: Opus 4.8 at 70.9%)
- Gameplay accuracy: 91.5% (Opus 4.8: 77.4%)
- Audio accuracy: 44.5% (Opus 4.8: 69.1%, GPT-5.5: 58.9%)
- Varoom 3D (3D racing game): 100% , first model to achieve perfect gameplay on this title (Opus 4.8: 25%)
- Spout (homebrew cave-flyer): 7.6% , a persistent failure across all models tested
The audio regression nobody expected
The most surprising finding is the audio score. Fable 5 scores 44.5% on audio , worse than both Opus 4.8 (69.1%) and GPT-5.5 (58.9%). This is a clear regression, not just a plateau. It suggests that whatever training changes drove Fable 5's extraordinary gameplay accuracy may have come at the cost of audio emulation fidelity , a classic capability trade-off that shows up when models are optimized hard for one axis of a multi-objective task.
The Spout failure is a separate, older problem. Every model tested so far diverges shortly after the loading screen on this homebrew cave-flying game. It's the benchmark's current "unsolved" case, and Fable 5 doesn't crack it either.
What Fable 5 actually is
Anthropic launched Claude Fable 5 as a Mythos-class model made safe for general use. Its capabilities exceed those of any model Anthropic has ever made generally available, with state-of-the-art performance on nearly all tested benchmarks, including software engineering, knowledge work, vision, and scientific research.
The longer and more complex the task, the larger Fable 5's lead over other models. That framing maps directly onto what GBA Eval is measuring: not a quick code snippet, but a sustained, multi-hour engineering effort with real correctness constraints.
The model comes with a notable architectural wrinkle. The general-purpose Fable 5 wraps the same underlying Mythos-class capability in new safeguards. Requests involving certain high-risk areas , including cybersecurity, biology, chemistry, and model distillation , are automatically routed to Claude Opus 4.8 instead, with users notified when that happens. For most coding and knowledge work, you're getting the full Mythos-class model.
Run Claude Fable 5 in an agent harness like Claude Code or Claude Managed Agents, and it can work for days at a time: planning across stages, delegating to sub-agents, and checking its own work. The GBA Eval result is essentially a live demonstration of exactly that capability , a 24-hour autonomous coding session on a genuinely hard systems programming problem.
Real-world signal beyond the benchmark
The GBA Eval result isn't an isolated data point. During early testing, Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand.
In GitHub's internal benchmarks on autonomous coding workflows, Fable 5 completed equivalent work with fewer tool calls and lower token consumption than previous Opus-tier models. Token efficiency matters a lot at Fable 5's price point.
Availability and cost
Claude Fable 5 is available on the Claude Platform natively, through available marketplaces, and in Amazon Web Services, Google Cloud, and Microsoft Foundry. It is priced at $10 per million input tokens and $50 per million output tokens, with the existing 90% input token discount for prompt caching.
To call it via the API, use claude-fable-5.
Claude Fable 5 and Claude Mythos 5 share the same specs: a 1M token context window by default, and up to 128k output tokens per request. One important operational note: both models carry 30-day data retention and are not available under zero data retention. That's a meaningful change for enterprises that previously ran Claude under zero-retention agreements.
What the benchmark is really measuring
The deeper point Mechanize is making with GBA Eval goes beyond any single model score. They've argued for some time that current evals, benchmarks, and RL environments for LLM capabilities have significant quality problems, and that these problems pose a major barrier to future improvements in model capabilities. GBA Eval is their answer: a single, carefully constructed task where the grading is provably correct, the problem is genuinely hard, and there's no way to reward-hack your way to a high score.
Fable 5's 74.5% , with near-perfect gameplay but a surprising audio regression , is exactly the kind of signal a well-designed benchmark should produce: not a clean sweep, but a detailed capability profile that tells you something real about where the model is strong and where it still has room to grow.