Claude Sonnet 5 Launches and Fable 5 Returns Tomorrow – What It Means for AI Models

In the early hours, Anthropic announced two major updates: the official release of Claude Sonnet 5 and the scheduled reinstatement of Claude Fable 5 after a nearly three‑week forced removal.

Sonnet 5 is positioned as the most agent‑friendly model in the Claude series, capable of autonomous planning, browser interaction, and terminal control without supervision—abilities that previously required an Opus‑level model.

Benchmark curves released by Anthropic show Sonnet 5 dramatically outperforming Sonnet 4.6 on BrowseComp (agent search) and OSWorld‑Verified (computer operation), achieving an effort level of xhigh and, on some tasks, matching Opus 4.8 performance.

Pricing is set at $2 per million input tokens and $10 per million output tokens until August 31, after which standard rates of $3 and $15 resume. Sonnet 5 is the default model for Free and Pro tiers, and is available in Claude’s web UI, desktop client, Claude Code, and via the API under the ID claude-sonnet-5.

The model also adopts a new tokenizer, similar to Opus 4.7, increasing token count by roughly 1.0–1.35× for the same text. Although per‑token prices are lower, the higher token consumption offsets the discount, and usage costs are expected to rise sharply once the August price increase takes effect.

On security, Anthropic notes that Sonnet 5 exhibits fewer mis‑alignments, hallucinations, and over‑eagerness than Sonnet 4.6, but its network‑security capabilities are deliberately throttled compared to Opus 4.8 and Mythos 5. The protection stack mirrors that of Opus 4.7/4.8 and is less restrictive than the safeguards originally placed on Fable 5.

Turning to Fable 5, the model was launched on June 9 alongside Mythos 5, with Anthropic describing Fable 5 as a safety‑hardened, publicly accessible Mythos‑level model. After three days of intensive testing (see the author’s earlier “Claude Fable 5实测” article), the U.S. Department of Commerce issued an export‑control order on June 12, citing national‑security concerns over a jailbreak, halting access for all foreign users, including Anthropic staff.

Negotiations ensued for nearly three weeks. Initially led by CEO Dario Amodei, the talks later shifted to co‑founder Tom Brown, reportedly due to Amodei’s high‑profile AI‑safety stance and political affiliations. On June 26, Commerce Secretary Katherine Lutnick granted a limited exemption for Mythos 5 to about 100 trusted partners, and on June 30 the export controls on both Fable 5 and Mythos 5 were fully lifted, a move confirmed by Anthropic on X.

The episode coincided with OpenAI’s limited rollout of GPT‑5.6, which also faced pre‑release restrictions. Both cases reflect the broader impact of the AI administrative order signed by former President Trump in early June, mandating voluntary model submissions for government review before public release, suggesting that future frontier models will undergo similar scrutiny.

For domestic Chinese models, the regulatory pause offers a brief window to catch up as U.S. models face repeated review cycles.

It remains uncertain whether Fable 5 will be re‑included in the standard Claude subscription; if access is limited to API‑based usage, its practical value may be limited.