In the span of three weeks, Anthropic's model lineup went from its biggest launch to a forced global suspension to a full restoration — with a new model launched in between. Here is a complete timeline of what happened and what it means for developers and subscribers.
Timeline: Three Weeks of Rapid Changes
June 9, 2026 — Claude Fable 5 launches. Anthropic's first Mythos-class model goes live. Fable 5 offers a 1M token context window, extreme reasoning benchmarks (78.5% on SWE-bench Pro), and dynamic subagent workflows. Available on Pro and Max plans and via API at $10/M input, $50/M output.
June 12, 2026 — Fable 5 suspended globally. Three days after launch, the US Department of Commerce issues an export control directive targeting Fable 5 and Mythos 5. Anthropic, unable to verify user nationality in real time, disables both models for all users worldwide — subscribers and API users alike. Claude Opus 4.8 becomes the effective top model on all plans.
June 30, 2026 — Claude Sonnet 5 launches. While Fable 5 remains suspended, Anthropic ships Claude Sonnet 5 — the most capable mid-tier Claude model to date. Sonnet 5 features a 1M context window, adaptive thinking with five selectable effort levels, and benchmarks that approach Opus 4.8 at a significantly lower API price.
July 1, 2026 — Fable 5 restored worldwide. The US Department of Commerce rescinds the export controls. Anthropic confirms that the suspension was triggered by a jailbreak technique reported by Amazon that circumvented Fable 5's safety classifiers. Anthropic has shipped a new classifier that blocks this specific technique in over 99% of cases. Access is restored globally.
July 7, 2026 — Pricing structure changes. From this date, Fable 5 access on Pro and Max plans shifts from being included in the flat subscription to a usage-credits model. Subscribers can still access Fable 5, but consumption draws from a separate credit balance rather than the monthly plan allowance.
What Was the Jailbreak?
Amazon's security team identified a technique that could cause Fable 5 to follow instructions it was designed to refuse — a category of vulnerability known as a jailbreak. The specific technique allowed prompts to bypass Fable 5's safety classifiers under certain structured input conditions.
Anthropic confirms the new classifier targets and blocks this exact technique. The company has not published specific technical details of the jailbreak to prevent replication, but has made the classifier's effectiveness figures public: over 99% block rate in internal testing.
The restored model is described by Anthropic as a "battened-down version" of Fable 5 — functionally equivalent for standard use cases, with stronger classifier coverage on the specific vulnerability class.
Claude Sonnet 5: What Developers Need to Know
Sonnet 5 is not a minor update — it represents a substantial capability jump over Sonnet 4.6:
- Context window: 1M tokens (up from 200k on Sonnet 4.6), with context compaction and 128k maximum output tokens.
- Adaptive thinking: Five selectable reasoning effort levels — low, medium, high, max, and x-high — allowing developers to trade response time for reasoning depth per request.
- Benchmark performance: SWE-bench Verified 85.2%, Humanity's Last Exam (with tools) 57.4%, OSWorld-Verified 81.2%. Approaches Opus 4.8 on several benchmarks.
- API pricing: Introductory rate of $2/M input, $10/M output through August 31, 2026. Standard rate of $3/M input, $15/M output from September 1, 2026.
For developers who were routing standard tasks to Sonnet 4.6, Sonnet 5 is a direct upgrade at the same or lower cost (introductory pricing is below Sonnet 4.6's standard rate). For developers routing frontier tasks to Opus 4.8, Sonnet 5 handles a meaningful portion of those cases at roughly 40% of the cost.
What This Means for Your Plans
If you use Claude.ai Pro ($20/month): Fable 5 is accessible again. Until July 7, usage is included in your plan at up to 50% of your weekly limit. From July 7, Fable 5 requires usage credits. Opus 4.8 and Sonnet 5 remain fully included in the flat plan with no additional cost.
If you use Claude Max 5x or Max 20x: Same as Pro — Fable 5 is available now, shifts to credits from July 7. The primary benefit of Max plans (5× or 20× higher usage limits) continues to apply to Opus 4.8 and Sonnet 5 on the flat subscription.
If you use the Claude API: Fable 5 is available at $10/M input, $50/M output with 90% prompt caching discount on cached input. Sonnet 5 is available at $2/M input, $10/M output through August 31. Consider whether Sonnet 5 can handle your workload — at $2/M vs $10/M, routing to Sonnet 5 where quality is sufficient saves 80% on input cost.
Should You Route to Fable 5 or Sonnet 5?
The practical routing decision after these changes:
- Use Sonnet 5 for most professional and production tasks: coding, document analysis, summarization, agent workflows, RAG applications. Sonnet 5's benchmarks are close enough to Opus 4.8 on standard tasks that it handles the majority of real-world use cases effectively. The introductory pricing makes it the clear economic choice.
- Use Opus 4.8 when you need Sonnet 5's quality plus more reliable handling of very long context or maximum precision on complex professional tasks.
- Use Fable 5 specifically for Mythos-class reasoning: extreme mathematics, the hardest coding architecture problems, or tasks where the quality ceiling of Opus 4.8 is genuinely not sufficient. At $10/M input, reserve it for cases where it clearly earns its cost.