Claude Fable 5 Long Context and Memory: A Deep Dive
How Claude Fable 5 maintains focus across millions of tokens, and what its 3x Slay the Spire result over Opus 4.8 reveals about file-based memory.
TL;DR
Claude Fable 5 maintains focus across millions of tokens, making it the strongest long-context model Anthropic has shipped. In a test playing Slay the Spire with file-based memory, it performed 3x better than Claude Opus 4.8. Combined with its 80.3 percent SWE-Bench Pro score, this is what unlocks codebase-scale tasks like Stripe's one-day, 50-million-line migration.
Focus, Not Just Capacity
Long-context claims usually describe capacity - how many tokens fit in the window. Anthropic's claim for Fable 5 is about something harder: maintaining focus across millions of tokens. Models with large windows often degrade in practice, forgetting early instructions, losing track of intermediate state, or drifting off-task deep into a long run. Fable 5's signature improvement is that quality holds up across the whole span.
This is why Anthropic highlights long and complex tasks as the area where Fable 5 most outperforms benchmarks' expectations, and why Cursor CEO Michael Truell says it has "opened up a class of long-horizon problems that were out of reach."
The Slay the Spire Result
The most concrete memory data point is a game: playing Slay the Spire using file-based memory, Fable 5 performed 3x better than Opus 4.8. The setup matters. Slay the Spire is a roguelike deck-builder where good play requires remembering deck composition, relic synergies, and lessons from earlier encounters - far more state than fits comfortably in working context. The model must decide what to write to its memory files, keep those notes organized, and actually consult them later.
A 3x gap on that loop means Fable 5 is much better at the meta-skill: knowing what is worth remembering, recording it usefully, and retrieving it at the right moment. That skill transfers directly to multi-session coding agents, long-running research tasks, and any workflow where an agent must pick up where it left off.
What This Unlocks in Practice
- Codebase-scale work: Stripe's 50-million-line Ruby migration completed in one day instead of an estimated two-plus months, which is only possible if the model stays coherent across an enormous task
- Overnight and multi-day agent runs: state-of-the-art results on long and complex tasks mean fewer babysitting checkpoints
- Persistent assistants: file-based memory plus long-context focus lets an agent maintain a project across many sessions without re-explaining everything
Tips for Long-Context Work with Fable 5
- Front-load the full task specification; the model can hold it across the entire run
- Give agents a writable memory directory for anything multi-session - the Slay the Spire result shows Fable 5 exploits it far better than prior models
- Use the June 9-22 free window on Pro, Max, Team, and Enterprise plans to stress-test your longest real workloads before committing budget at the 10 and 50 dollar per-million-token rates