The blackout ends - with a patch, a launch, and a precedent
The most consequential AI story of the summer just resolved. Anthropic has restored access to Fable and Mythos, its highest-capability models, closing an eighteen-day pause that began when a US export-control directive forced them offline on June 12. The trigger was a method documented by Amazon researchers that bypassed Fable 5's safety controls, getting it to identify software vulnerabilities and supply exploitation code - though security evaluations during the shutdown found the behaviour was not unique: Claude Opus 4.8, GPT-5.5, and Kimi K2.7 all duplicated the results. The fix is an automated safety classifier trained against the specific bypass, which internal validation says blocks the technique in more than 99% of trials. It runs with a deliberately wide safety margin, which cuts both ways: ambiguous prompts get flagged more often during routine development, and when that happens the platform automatically reroutes the workload to Opus 4.8 to keep applications running.
Sonnet 5: the commercial centre of gravity
Alongside the restoration, Anthropic shipped Claude Sonnet 5, aimed squarely at production agentic workloads. It scores 63.2% on SWE-bench Pro and 80.4% on Terminal-Bench 2.1 - clear gains over Sonnet 4.6's 58.1% and 67.0%, approaching Opus 4.8's 69.2% and 82.7% at a fraction of the price: $3 input / $15 output per million tokens, with introductory rates of $2 / $10 through August 31, 2026. The deployment stories are unusually concrete. Rakuten ran it against dozens of its hardest production pull requests, with the model executing tests and verifying results before humans signed off. Zapier had it complete multi-stage admin sequences - updating Salesforce account tiers, then generating and sending launch announcements - end-to-end where prior models stalled midway. Zed watched it independently reproduce, fix, and verify an active bug in a single pass, and Factory reports it finishing long codebase tasks that previously timed out. The system card adds a notable safety datapoint: Anthropic omitted specialised offensive-cybersecurity data from training, and in public Mozilla assessments the model failed to build a single working exploit against Firefox 147 - a zero percent success rate.
Governance is becoming infrastructure
The episode's lasting legacy may be institutional. Anthropic, Amazon, Microsoft, and Google are jointly building an objective framework for scoring model security breaches across four criteria - capability gain, breadth of that gain, ease of weaponisation, and discoverability - with automated mitigations mandated for high-severity exploits threatening things like financial systems or power grids. Add a new HackerOne vulnerability programme, a 24-hour threat-monitoring team, and formalised agreements giving federal researchers early access to frontier models before public release, and the relationship between labs and the state has visibly changed shape.
Your playbook for the new model landscape
- Treat regulatory shutdown as a standing vendor risk, not a black swan - it has now actually happened. Note that Anthropic's own remedy is multi-model fallback routing; architect the same into your stack, with a tested secondary model and provider.
- The Sonnet 5 intro pricing window (through August 31) is a genuine migration opportunity for agentic workloads: near-Opus capability at Sonnet prices is the kind of cost-performance shift worth a re-benchmark of your current deployments.
- Budget engineering time for classifier false positives: the widened safety margin will occasionally flag benign development prompts, so build retry-and-reroute handling rather than treating blocks as hard failures.
- If you operate in a regulated sector, watch the four-criteria breach framework - it is the likely seed of tomorrow's compliance vocabulary, and aligning your internal AI incident taxonomy with it early is cheap insurance.