Technology RadarTechnology Radar

Claude Fable 5

frontierllmcodingagenticreasoningmultimodal
Adopt

Claude Fable 5, released June 9, 2026, is the first publicly available Mythos-class model from Anthropic — a step-function above Opus 4.8 on every major coding and reasoning benchmark. It is the most capable generally available model Anthropic has released.

Why It's in Adopt

Fable 5 is Anthropic's public answer to the question "what happens when Mythos-class capability ships with safety scaffolding?" The result is the highest-scoring coding model generally available:

  • SWE-bench Verified 95% — surpassing Claude Mythos Preview's 93.9%, which was previously restricted to ~40 Project Glasswing partners. The largest SWE-bench jump to reach general availability.
  • SWE-bench Pro 80.3% — the headline agentic-coding number, 11 points clear of Opus 4.8.
  • "Days of work" agentic endurance: Fable 5 can run agentic coding tasks for extended durations with minimal human intervention — a qualitative shift from the Opus-class multi-step sessions.
  • Safety scaffolding: Ships with conservative safety filters that, on average, trigger in fewer than 5% of sessions. When triggered, requests redirect to Opus 4.8. For the vast majority of developer workflows, the scaffolding is transparent.

Fable 5 is broadly available — Claude.ai, the Anthropic API, Amazon Bedrock, Vertex AI, and GitHub Copilot all launched same-day. The 2× price premium over Opus 4.8 ($10/$50 vs $5/$25) is the only meaningful trade-off.

Performance

Benchmark Fable 5 Opus 4.8 Mythos Preview
SWE-bench Verified 95.0% 88.6% 93.9%
SWE-bench Pro 80.3% 69.2% 77.8%
FrontierCode Diamond 29.3% 13.4% —
GPQA Diamond — 93.6% 94.5%

FrontierCode Diamond is the hardest 50-task subset of the FrontierCode benchmark — a set of "expert-level" multi-step coding challenges designed to resist saturation. Fable 5's 29.3% represents a ~5× improvement over Opus 4.8 (13.4%) and a ~5× improvement over GPT-5.5 (5.7%). Multiple independent trackers (Weights & Biases, Vellum AI, TrueFoundry, aivy.com.au) agree on the 29.3% figure.

Anthropic retired GPQA Diamond from official benchmark reporting after the Opus/Mythos family crossed 94%, declaring the benchmark saturated and no longer a useful differentiator. Third-party aggregators report Fable 5 in the 92–93% range — consistent with the Opus 4.8 / Mythos Preview band.

Fable 5 outperforms Claude Mythos Preview (previously only available through Project Glasswing) on the two coding benchmarks, suggesting the safety scaffolding has not degraded coding capability below the restricted predecessor.

Pricing

Property Value
Input $10 per million tokens
Output $50 per million tokens
Context window 1,000,000 tokens
Prompt caching Up to 90% savings on repeated context
Batch processing 50% savings for async tasks

Pricing is 2× Opus 4.8 and matches Claude Mythos 5 (glasswing-restricted).

When to Use Fable 5 vs Opus 4.8

Use Opus 4.8 Use Fable 5
Cost-sensitive production usage Highest-quality code generation matters
Batch processing at scale Long-running autonomous coding sessions
Tasks with sensitive content (lower scaffold trigger rate) Complex multi-file refactors
Most routine agentic workflows SWE-bench-class engineering challenges

July 2026 Status Update

Export controls (June 12–July 1, 2026): On June 12, the U.S. Commerce Department imposed emergency export controls on Fable 5 and Mythos 5 after Amazon researchers discovered a jailbreak enabling the models to write proof-of-concept exploit code. Access was restricted globally. On June 30, the controls were lifted after Anthropic deployed a safety classifier blocking the technique in more than 99% of attempts. Fable 5 was fully restored worldwide on July 1, 2026. Mythos 5 remains restricted to ~100 U.S. critical infrastructure organizations. See security headline #45.

Subscription access change (took effect July 19, 2026): The July 19 cutoff took effect. Fable 5 access is now split permanently by plan tier: Max and Team Premium subscribers retain Fable 5 at 50% of weekly limits with no additional cost; Pro and Team Standard subscribers received a one-time $100 usage credit and now access Fable 5 at API billing rates ($10/$50 per 1M tokens). The split was reached after two extensions — original July 7 cutoff pushed to July 12, then to July 19 — following user backlash over abrupt access removal. The 50% Claude Code weekly usage limit increase became permanent for all plan tiers. The model remains available via the Anthropic API and consumption-based Enterprise contracts without restriction.

Performance after restoration: Early user reports describe a noticeable capability regression relative to the pre-export-control Fable 5 release. Systematic benchmarks are not yet available. The regression is likely attributable to the safety classifier update deployed to address the jailbreak trigger. Watch for updated benchmark data from Anthropic.

How to Access

  • Claude.ai — Max and Team Premium: Fable 5 retained at 50% of weekly limits (permanent); Pro and Team Standard: one-time $100 usage credit applied, then API billing ($10/$50 per 1M tokens); enable via Settings → Usage
  • Anthropic API — claude-fable-5 model ID
  • Amazon Bedrock — available at GA
  • Google Cloud Vertex AI — available at GA
  • GitHub Copilot — available at GA (June 9, 2026)

Key Characteristics

Property Value
License Proprietary
Context window 1,000,000 tokens (1M)
Pricing $10/M input, $50/M output
Provider Anthropic
API model ID claude-fable-5
Release date June 9, 2026

Further Reading