Anthropic's Opus 5 Offers a Third of Fable's Cost—But Raises New Safety Challenges

Anthropic's latest heavyweight model, Opus 5, flips the economics of agentic AI work. At just $5 per million input tokens and $25 per million output tokens—roughly a third of Fable 5's pricing—the model opens the door to longer, more autonomous coding sessions while raising fresh operational concerns around cost management and safety guardrails.

Performance gains at lower costs#

Opus 5 establishes new benchmarks in its performance class. On programming and reasoning evaluations like Frontier-Bench and GDPval-AA, it ranks among the field's top performers. More tellingly, it outperforms Fable 5 on the OSWorld 2.0 computer-use benchmark at one-third the cost, and achieves a score three times higher than the next-best model on ARC-AGI 3, where systems must tackle novel problems without training examples.

According to Anthropic, the model demonstrates notable self-correction abilities. In one test case, when prevented from viewing a detailed diagram, it independently wrote a computer vision pipeline to reconstruct the missing information rather than abandon the task.

New infrastructure demands for unsupervised work#

Cheaper execution costs mean teams can run agents through more complex, multi-step workflows—but that autonomy introduces control challenges. With AI systems working across multiple systems with minimal human oversight, the industry is shifting toward tighter access management: temporary credentials that expire immediately after use, and isolated execution environments like microVMs that can be destroyed as soon as a task completes.

Operators also need smarter observability. Standard API logs won't catch runaway token consumption in time. The field is gravitating toward semantic circuit breakers—systems that understand the context and intent of a workflow, not just raw API calls—to shut down agents before they exhaust budgets.

Managing costs and preventing silent failures#

Autonomous coding agents churning through background loops can accumulate massive token bills before surfacing results. To prevent outright pipeline failures, Anthropic is introducing Automatic Fallbacks in beta: if a prompt triggers a safety classifier, the API silently routes the request to Opus 4.8 instead of returning an error that breaks the entire job.

Opus 5 also preserves Anthropic's zero-retention model—data is not logged for 30 days as it is with Fable and Mythos, improving privacy for sensitive code tasks.