Anthropic has reopened the top of its model stack while changing the lane most developers will actually use. Claude Fable 5 returned globally on July 1 after an unusual US export-control suspension, now with a classifier trained against the reported bypass. One day earlier, Claude Sonnet 5 became the default for Free and Pro users and the lower-cost API starting point for agentic work. The practical story is not that Fable is “back to normal.” It is that Sonnet now handles the daily lane, while Fable returns as a costly, guarded escalation and Mythos remains restricted.

Fable returns; Mythos stays gated

Anthropic launched Fable 5 and Mythos 5 on June 9. Fable exposed the same underlying model with stronger safeguards for general use; Mythos went only to a small group of Project Glasswing partners for defensive cybersecurity. On June 12, Anthropic suspended both after a US government export-control directive covering foreign nationals made immediate compliance impractical.

The trigger, according to Anthropic’s redeployment account, was an Amazon research report describing a way around Fable’s safeguards. The model identified several software vulnerabilities and produced exploit-demonstration code in one case. Anthropic says its review found no unique Mythos-level capability: other models found the same vulnerabilities and reproduced the demonstration. That is Anthropic’s technical interpretation, not an independently published Amazon finding. Axios reported that Amazon’s warning prompted the government response; AP independently confirmed the suspension and later restoration.

Anthropic trained a new classifier and claims it blocks the specific reported technique in more than 99% of cases. The company also acknowledges more false positives in ordinary coding and debugging, with blocked requests routed to Opus 4.8. Its broader statements about Fable’s safeguards and Mythos’s exceptional cyber capability remain vendor claims.

The controls were lifted June 30 and Fable returned July 1 on the Claude Platform, Claude.ai, Claude Code, and Cowork. Mythos did not become self-serve: access was restored only to approved US organizations. Axios reported the June 26 Commerce Secretary letter and the June 30 lifting; no corresponding public Commerce newsroom release was available when this article was reviewed. WIRED separately reported the unresolved government talks and an NSA review of the vulnerability claims. The June 2 executive order describes a voluntary frontier-model review framework and explicitly says it does not create mandatory preclearance, making this episode an unsettled precedent rather than a clean statutory process.

Sonnet 5 becomes the daily lane

Sonnet 5 is available across Claude plans, Claude Code, and the API as claude-sonnet-5. It costs $2 per million input tokens and $10 per million output tokens through August 31, then $3/$15. Anthropic made it the default for Free and Pro users; Max, Team, and Enterprise users can select it.

One launch comparison is useful if kept narrow. Anthropic’s current system card reports SWE-bench Pro scores of 63.2% for Sonnet 5, 69.2% for Opus 4.8, and 58.1% for Sonnet 4.6. Those are Anthropic-reported results on that named software-engineering benchmark—not a general agentic ranking. Selected partner testimonials in the launch post are also vendor-curated evidence, not independent validation.

The pricing needs a tokenizer footnote. Anthropic says equivalent text maps to roughly 1.0–1.35 times as many tokens as with Sonnet 4.6, depending on content. The introductory rate is intended to make migration roughly cost-neutral. List price alone therefore does not prove a lower cost per completed task.

What builders should change

Use a measured routing ladder:

LaneUse it forVerify before production
Sonnet 5Routine coding, tools, agents, and knowledge workToken counts, max_tokens, effort, and HTTP 200 refusals
Opus 4.8Hard review, architecture, debugging, and final arbitrationWhether the quality gain changes the outcome
Fable 5Rare cases where Opus demonstrably failsCredits, route/region, retention, classifier fallback, and cost per task
Mythos 5Approved-organization defensive cyber workExplicit organizational approval; it is not self-serve

Do not treat a successful HTTP response as proof that the requested Fable model completed the work. Log stop_reason, the model that actually answered, retries, and whether the result remained usable after an Opus fallback. Fable is included for up to 50% of weekly usage on eligible plans only through July 7; afterward it uses usage credits, and standard Enterprise seats require credits from the start. Anthropic says AWS, Google Cloud, and Microsoft Foundry re-enablement is still in progress, so cloud and region checks remain deployment gates.

Measure complete task cost—input, thinking, tools, retries, fallbacks, and output—rather than multiplying a prompt estimate by the headline token rate. The Sonnet 5 guide covers migration details; the Fable decision guide covers retention and buying constraints.

Open questions and developer takeaway

The unresolved issue is process. The June shutdown and June 30 reversal show that model availability can change through government review before a durable public standard exists. Anthropic’s over-99% claim addresses one reported bypass, not the classifier’s real-world false-positive rate across developer workloads.

For now, route daily work to Sonnet 5, escalate consequential failures to Opus 4.8, and test Fable only when the measured improvement can justify its price and guardrails. Mythos remains an approved-organization path; lifting the export controls did not make it generally available.

Sources