Anthropic released Claude Opus 5 on Friday, its fourth Claude 5-series model in under two months and the clearest sign yet that the frontier lab has decided the interesting margin isn’t raw capability but cost per task. Opus 5 holds Opus 4.8’s $5/$25 pricing while more than doubling its predecessor’s Frontier-Bench v0.1 score, and it closes most of the gap to Fable 5 at roughly half the price.
The pitch, per Bloomberg, is aimed squarely at “cost conscious” enterprise buyers as Chinese competition intensifies. Anthropic expects Opus 5 to become the default for day-to-day office work, and it’s already the new default on Claude Max and the strongest option on Claude Pro.
The benchmark posture is aggressive. Opus 5 claims state-of-the-art on Frontier-Bench v0.1 and GDPval-AA, roughly three times the next-best model on ARC-AGI 3, and about 1.5x the next-best on Zapier’s AutomationBench at the same cost per task. It still trails the larger Mythos 5 on cybersecurity work. Anthropic illustrated the reliability story with a Frontier-Bench task in which the model, denied direct image access, wrote its own computer vision pipeline to reconstruct a machine part as a 3D FreeCAD model. No competing model solved it in five attempts under the same setup. The company describes the new Opus as “much stronger at verifying its work and iterating carefully until it succeeds.”
The more consequential shipment may be the effort dial. Per Fortune, API and product customers can now pick low, medium, or high at runtime, choosing where to sit on the intelligence-versus-token-spend curve. The context here’s the industry’s quiet pivot: business customers had criticized Fable 5’s token burn rate, and OpenAI made economical token usage a marketing point when it shipped GPT-5.6 on July 9.
The release cadence tells its own story. Opus 4.8 launched May 28. Mythos 5, Fable 5, and Sonnet 5 all shipped in June. Only Haiku is still waiting for a 5-series upgrade. Fable 5 itself had a rough debut, pulled from market on June 12 after Amazon researchers demonstrated safeguard bypasses that drew temporary U.S. export controls, and re-released June 30 with hardened security. Anthropic still recommends Fable 5 for the most demanding autonomous work, the kind that runs for days on end.
Also new: a beta feature called Automatic Fallbacks, which routes API requests to a less powerful model when a prompt trips the safety classifier, rather than returning an error. It’s a small operational tweak that reads as an admission of where enterprise friction actually lives. Not in benchmark ceilings, but in the calls that don’t come back.