Anthropic · Intelligence Briefing · September 2026
Claude Fable 5.1:
one frontier model,
two carefully different fables
Anthropic's newest flagship pairs record reasoning benchmarks with cache reads 75% cheaper than its predecessor — while a tightly restricted twin, Claude Mythos 5.1, quietly carries the capabilities Fable was built to hold back.
66
Intelligence Index, Max effort
↓75%
Cache-read cost vs. Fable 5
1M
Token context window
5
Configurable reasoning levels
I.
The split foundation
Same underlying weights, two names. Fable 5.1 ships broadly, wrapped in extra safeguards; Mythos 5.1 stays behind an approval wall reserved for vetted cybersecurity and life-sciences organizations who need the version without the training wheels.
How a request gets routed
1
Request submitted to Claude Fable 5.1
Developer or consumer sends a prompt at a chosen reasoning effort level.
2
Automated safety classifiers evaluate it
Rebuilt safeguards cut false-positive interventions by 85% on benign biology requests, while still supporting source-code security analysis without enabling exploit generation.
3
A trigger reroutes the request automatically
Configured through the Fallback API; routing can stay "sticky" for the rest of a multi-turn conversation.
Cybersecurity trigger → Claude Opus 4.8
Biology trigger → Claude Opus 5
4
Response returns with full transparency
Consumer apps show a model-switch notice; the API returns a fallback content block naming which model actually answered, its reasoning effort, and tokens used.
Enterprise data governance
Standard deployment
30-day data retention for safety monitoring.
Enterprise Frontier Safeguards
Customer-controlled infrastructure for sensitive data.
Zero-data-retention
Available to qualifying organizations on request.
II.
The benchmarks
At maximum reasoning effort, Fable 5.1 posts the highest Artificial Analysis Intelligence Index score Anthropic has published for the family — and roughly doubles its predecessor on the hardest agentic and scientific tests.
Artificial Analysis Intelligence Index (0–100, Max effort)
ARC-AGI, at Max effort
97.5%
ARC-AGI-1
Near-total clearance of the original benchmark.
90.0%
ARC-AGI-2
The harder successor benchmark, still cleared at 90%.
Fable 5 → Fable 5.1, hardest agentic & scientific tests
Claude Fable 5
Claude Fable 5.1
Knowledge work (AA-Briefcase, 1,694 Elo — tied with Opus 5's 1,685)
Fable 5.1 attempts 93.4% of knowledge questions, versus Opus 5's 87.8% — and gets 67.2% of what it attempts right.
Analytical quality
2,025 Elo
vs. Opus 5 at 1,980 — Fable 5.1 ahead
Presentation quality
1,495 Elo
vs. Opus 5 at 1,572 — Opus 5 ahead
III.
The ledger
Reasoning effort is now a dial, not a fixed setting — and Anthropic cut the cost of remembering by three-quarters.
Five reasoning effort levels
Fable 5.1 pricing, per million tokens
| Service | Rate |
|---|---|
| Standard uncached input | $10.00 |
| Generated output | $50.00 |
| 5-minute cache write | $12.50 |
| 1-hour cache write | $20.00 |
| Cache read −75% | $0.25 |
| Batch input (−50%) | $5.00 |
| Batch output (−50%) | $25.00 |
Cache-read cost across the Claude line-up
Fable 5.1
$0.25
Fable 5
$1.00
Opus 5
$0.50
Sonnet 5
$0.20
What that means in practice
~25%
Estimated cost reduction on typical cached workloads
~45%
Savings ceiling on highly agentic, cache-heavy workloads
Where it runs
Claude API
Amazon Bedrock
Google Cloud
Microsoft Foundry
Claude Platform on AWS
OpenRouter
IV.
The discoveries
Outside the benchmark suite, early-access partners pointed Fable 5.1 — and its restricted sibling Mythos 5.1 — at three genuinely hard scientific problems.
Planetary science
Venus, remapped from 30-year-old radar
Fable 5.1 reprocessed NASA's Magellan radar imagery — gathered more than three decades ago — into a new high-resolution elevation map covering roughly a third of the planet, released under a Creative Commons license.
2–3 km
New observable detail scale, vs. 10–20 km before
+25%
Improvement in height accuracy
CC
Released under Creative Commons license
Molecular design · Claude Mythos 5.1
Protein binders at 3–5× the usual hit rate
Tested across 12 protein targets and validated through external laboratory testing, with binding affinity up to ~10× stronger than baseline on three of them.
Computational biology
GPU optimization in days, not weeks
Seven open-source deep-learning models re-optimized for genome-scale analysis, at a fraction of the usual manual engineering time.
Up to 2.5×
GPU performance improvement
30–60%
Estimated GPU cost savings
V.
The field
Ten early-access organizations, one pattern: Fable 5.1 stays on task longer, and keeps finding the thing nobody was looking for.
Millennium
Investment managementSurfaced a vendor library bug that had gone unresolved for years.
MongoDB
Database softwareBuilt a functional prototype in roughly 3 days.
Ramp
Financial technologyCompleted a 38-hour unattended ML investigation.
Jane Street
Quantitative financeSharper coding and trading-intuition performance.
IMC
Trading researchSurfaced novel analytical approaches.
Rakuten
Clinical researchIdentified a research gap earlier review had missed.
Red Hat
Enterprise softwareDiagnosed root causes across multiple builds.
Datadog
Cloud monitoringDiagnosed complex production incidents.
Plaid
Financial technologyTraced workflows across multiple codebases.
Shopify
E-commerceMaintained task continuity over an extended session.
Field dispatch
Inside Ramp's 38-hour unattended run
1
Opens with the assigned hypothesis
2
Reassesses it against the data — and doubts it
3
Finds a labeling artifact skewing the results
4
Corrects the analysis, launches 6 parallel experiments overnight
5
Returns with findings and a recommended next step
Unattended, start to finish.
VI.
The verdict
Faster benchmarks and cheaper caching come with fine print worth reading before you scale a workload on it.
1
Subscription math differs from API math
API developers capture most of the cache-read savings. Max and premium seats can run Fable 5.1 for up to 50% of their weekly limit; Pro and standard seats fall back to pay-as-you-go usage credits — and adaptive thinking runs at High effort by default, with thinking tokens counted against that usage.
2
Fallback routing is designed to be visible
A model switch shows as a notice in consumer apps; the API returns a dedicated fallback content block; usage metadata logs every model actually queried, and routing can stay sticky across a conversation.
The asterisk on every score above
Published benchmarks ran with production safeguards switched on. Some OSWorld tasks scored zero when a safeguard intervened mid-task, and cybersecurity or biology triggers routed those specific tasks to Opus 4.8 or Opus 5 instead. That means Fable 5.1's real ceiling is likely higher than the figures reported in Chapter II.
"
Fable and Mythos share one foundation model split across two safeguard profiles — not two different levels of intelligence. It's how Anthropic ships frontier capability broadly without lowering the floor for everyday use, while keeping the sharpest edge of it behind a door only a few vetted teams can open.