Cognition's Devin Restores Claude Fable 5 After a Messy Government Ban

Devin, Cognition's AI software engineering platform, has re-enabled Claude Fable 5 across all its surfaces: Devin Cloud, Devin Desktop, and Devin CLI. The timing is significant. As of June 30, the export controls on Fable 5 and Mythos 5 were lifted, with Fable 5 becoming available again starting July 1 to users globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Devin is one of the first coding agent platforms to restore the integration.

Claude Fable 5 was originally made available in Devin across Cloud, Desktop, and CLI on June 9, but as of June 12, access was removed following Anthropic's announcement and the US government directive. That directive came fast and hit hard: on June 12, the US government applied export controls to Fable 5 and Mythos 5, requiring Anthropic to restrict access to foreign nationals. Because the order took effect immediately and there was no reliable way to verify nationality in real-time, Anthropic suspended access to both models for all users.

What triggered the ban , and why it was lifted

Government officials enacted the restriction after researchers at Amazon documented a method to bypass the safety controls of Fable 5, causing the model to identify software vulnerabilities and supply exploitation code. The irony? Anthropic's testing confirmed that many less capable models , including Claude Opus 4.8, GPT-5.5, and Kimi K2.7 , could identify the same vulnerabilities as Fable 5 did in the report.

To get Fable 5 restored as quickly as possible, Anthropic trained an improved safety classifier that targets and blocks the jailbreaking behavior that led the government to apply export control to the models. The AI can now block the technique described by Amazon in more than 99% of cases, though the trade-off is a more sensitive AI with more false positives on legitimate development or debugging tasks.

Why Fable 5 matters inside Devin

Devin's pitch for Fable 5 is not just about raw model capability , it's about what that capability unlocks inside an agent harness. Claude Fable 5 earned the top score on FrontierCode, Cognition's proprietary benchmark that grades models on real-world tasks based on quality and mergeability. It excels at long-horizon reasoning and debugging, and it generalizes to unfamiliar tools out of the box, from MCP integrations to computer use.

FrontierCode is worth understanding: unlike benchmarks that test isolated coding problems, it grades whether a model's output is actually mergeable into a real codebase , closer to what a code reviewer would care about than a test suite pass rate. FrontierCode's emphasis on mergeability and overall quality is still early, so it is unclear how widely teams will adopt it versus existing coding benchmarks. That said, the signal it captures is genuinely useful for evaluating agents in production contexts.

Claude Fable 5 is built for long-horizon, multi-step tasks that require sustained reasoning, tool use, and error recovery , not just one-shot responses. Where most models degrade as context grows, Fable 5 maintains more consistent decision quality across long sessions , which matters for coding tasks where the model needs to remember what it decided three hours ago and why, and act accordingly.

The Ultra agent: Fable 5's home in Devin

Fable 5 is specifically wired into Devin Cloud's Ultra agent , the highest-capability tier in Devin's agent stack. Devin Ultra is Cognition's smartest and most capable agent, excelling at long-horizon tasks and debugging. Cognition tuned the harness so Ultra costs only ~40% more than the default Devin agent. That's a meaningful constraint: Fable 5 is expensive at the API level ($10/M input tokens, $50/M output tokens), so the efficiency of the surrounding agent harness matters a lot.

Fable 5 is also available in Devin Desktop (formerly Windsurf) and Devin CLI, giving teams flexibility in how they integrate it into their workflows. The practical use cases where it shines:

• Large-scale code migrations , tasks that span thousands of files and require tracking decisions made hours earlier
• Long-horizon debugging , multi-step root cause analysis across interconnected systems
• MCP and computer use integrations , generalizing to unfamiliar tools without custom prompting
• PR review and visual QA , catching bugs and organizing diffs for human review
• Incident triage , routing Slack bug reports and investigating Datadog alerts autonomously

The safeguard trade-off you should know about

The returning Fable 5 is not identical to the June 9 version. The returning version is a battened-down version of Fable 5, out of "an abundance of caution," redesigned to more effectively handle and abort cybersecurity tasks. Anthropic has arranged for Fable to fall back to Opus 4.8 on queries related to sensitive topics, with safeguards tuned conservatively , they'll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions.

For most software engineering work, this is invisible. Anthropic tuned Fable 5's safety classifiers specifically to protect cybersecurity and biology domains; for general software engineering, these guardrails remain largely invisible. But teams doing legitimate security tooling or penetration testing infrastructure should expect occasional fallback to Opus 4.8 on specific requests , not on the whole project.

Access and pricing

On the Devin side, Fable 5 is available now across all paid plans that include Devin Cloud. Devin's pricing tiers start at $20/month (Pro) and $200/month (Max), with Teams at $80/month plus per-seat costs. The Ultra agent mode that powers Fable 5 runs at roughly 40% above standard Devin agent pricing.

On the Anthropic API side, Claude Sonnet 5 is available at an introductory price of $2 per million input tokens and $10 per million output tokens. Fable 5 itself carries a steeper price tag: input tokens cost $10.00 per million and output tokens cost $50.00 per million, with a 90% discount on input costs available via prompt caching for repetitive contexts. For Pro, Max, Team, and select Enterprise plans, Fable 5 is included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits.

The broader context here matters too. On the same day Anthropic made Fable 5 and Mythos models available to customers again, it released Claude Sonnet 5 as the most agentic Sonnet model yet. Sonnet 5 narrows the gap: its performance is close to that of Opus 4.8, but at lower prices. For teams where Fable 5's cost is prohibitive, Sonnet 5 is now a much more capable fallback than it was a month ago , and it's also available inside Devin as a model option.

The Fable 5 saga has also accelerated something bigger: Anthropic is using this opportunity, together with Amazon, Microsoft, Google and other Glasswing program partners, to propose a common framework for assessing the severity of jailbreaks. How AI labs, governments, and cloud providers negotiate model access going forward will likely look a lot like what just played out with Fable 5 , fast, messy, and consequential.