Claude Fable 5.1
Anthropic's Mythos-class top tier for the longest, hardest problems where cost is secondary.
Overall Intelligence
#4
of 56
0.5 pts behind #1
12
races led
240
weeks at #1
88.2
score
- Released
- Aug 20, 2026
- 2 mo ago
- Context
- 1M
- 128K max output
- Price
- $10 / $50
- per 1M in/out · $20.00 blended
- Speed
- 54 tok/s
- 8.8s to first token
- Modalities
- text · image
- outputs text
- Weights
- Closed
- API and apps
The 30-second version
Anthropic's top tier, built for the hardest and longest problems. Fable 5.1 is the model you reach for when a task would take a skilled human days, and getting it right matters more than what it costs.
- 1Leads our long-horizon autonomy race: it can work productively on a single task for many hours.
- 2Top-tier reasoning and research quality; excellent at synthesising large bodies of evidence.
- 3Expensive ($10 / $50 per million tokens) and slower than Opus, so use it selectively.
- 4Pairs well with Opus or Sonnet as the escalation path when cheaper models get stuck.
What you need to know
Purpose
A 'Mythos-class' research and engineering partner for problems at the edge of what models can do: novel algorithms, deep investigations and long autonomous projects.
What Anthropic optimised for
Maximum capability and depth of reasoning, with heavy investment in safety evaluations for autonomous use. Speed and cost were explicitly secondary.
Unique strengths
- Sustains a line of investigation for hours, keeping notes and revisiting assumptions.
- Excellent at research synthesis: reconciling conflicting sources and stating confidence.
- Strong at novel problems where there is no pattern to copy.
Where it ranks
Current snapshot · illustrative scores · click any race to watch it
Fable 5.1Best on each axis
Overall
Coding
Hard Reasoning
Science
Writing & Language
Multimodal
Tool Use & Agents
Speed & Cost
Trust & Factuality
Overall Intelligence over time
Best use cases
01
Multi-day engineering projects
New subsystems, complex migrations or performance work that needs sustained focus.
02
Deep research
Literature reviews, due diligence and technical investigations with many sources.
03
Escalation tier
The model your agent calls when cheaper models have failed twice.
Strengths
Long-horizon autonomy
Leads the field on how long it can work unaided before needing help.
Research quality
Strongest at weighing evidence and producing well-argued syntheses.
Hard reasoning
Consistently in the top three on our frontier reasoning races.
Weaknesses
Cost
One of the most expensive models in the race; easy to overspend on routine work.
Speed
Around 50-55 tokens per second; not suited to interactive chat or quick edits.
Overkill for simple tasks
Can over-think straightforward requests that Sonnet handles in seconds.
Prompts that play to its strengths
Open-ended investigation
Investigate why our recommendation model's click-through rate dropped 8% after the March retrain. You have read access to the training configs, data pipeline code and the evaluation notebooks. Keep a running log of hypotheses, what you checked and what you found. Do not stop at the first plausible cause; rule out the alternatives. End with your conclusion, your confidence, and the experiment that would confirm it.
Why it works: Long, evidence-driven investigations with a log are exactly what it is built for.
Design review
Here is the design doc for our new event-sourcing system. Act as the most demanding principal engineer in the company. Identify the three decisions most likely to hurt us in two years, explain the failure scenario for each, and propose a concrete alternative with its trade-offs.
Why it works: Rewards depth of reasoning and foresight rather than speed.
When to choose it, and when not to
The problem is genuinely at the frontier and Opus has already struggled, or the task will run for many hours.
You need a sustained research or engineering investigation rather than computer use or contest math.
Pick something else when…
- Routine coding, chat or summarisation: Opus 5.5 or Sonnet 5.5 deliver nearly the same quality for far less.
- Latency-sensitive products: it is one of the slowest models in the race.
Update history
The distilled version first; open any release for the deeper breakdown.
The Claude Fable family