Opus 5 costs a third of the price — and that’s actually the problem
The New Stack Amanda Caswell ● Covered by 17 sources
Anthropic dropped Opus 5, a cheaper Claude that codes for hours on its own. Cheap and persistent is a great combo — until the AI doesn't know when to stop spending your money.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic's Opus 5 landed on Friday, just two months after Opus 4.8 and hot on the heels of June's Mythos 5, Fable 5, and Sonnet 5. It's cheaper than Fable 5 — $5 per million input tokens, $25 per million output — and it beats the flagship on agentic benchmarks like OSWorld 2.0 while costing roughly a third as much to run. On ARC-AGI 3, a benchmark built around novel problem-solving, Opus 5 scores triple the next best model. During testing, when Anthropic deliberately blocked the model from seeing a diagram it needed, it didn't stall out — it wrote its own computer vision pipeline to reconstruct the missing image data.
That persistence is the whole pitch, and also the whole problem. Cheaper tokens mean teams can let Opus 5 grind through much bigger coding tasks without a human checking in every few minutes. Fine, until you remember that an AI working unsupervised for hours doesn't have an internal sense of "I should probably stop now." Standard API logs won't catch that kind of runaway behavior in time, which means platform teams need telemetry that actually understands what the workflow is doing — something closer to a semantic circuit breaker than a simple usage alert.
Anthropic is clearly aware of the operational mess this creates. The company is rolling out Automatic Fallbacks in beta: if a prompt trips a safety classifier, the API quietly reroutes the job to Opus 4.8 instead of just killing the pipeline outright. Opus 5 also keeps the zero-retention data policy from its predecessor, skipping the 30-day logging requirement that applies to Fable and Mythos. And there's a growing shift toward short-lived, revocable credentials and microVMs — isolated environments that can be torn down the moment a task wraps — because letting an agent roam across systems for hours needs access controls that expire fast, not standing permissions.
On the science side, Anthropic says Opus 5 is a real step up for biology work: 10.2 percentage points better on organic chemistry benchmarks involving spectroscopy data, and 7.7 points better at predicting how protein variations affect function. But the company also found real limits on long-running autonomous research, which it treats as the highest-risk category, so Opus 5 keeps most of the same biological safeguards as Opus 4.8. Those guardrails will trigger about 85% less often than they did on Fable 5, according to Anthropic, though Mythos 5 remains the model of choice for open-ended work like drug design campaigns.
The pattern across this release cycle is unmistakable: model capability is outrunning the infrastructure built to contain it. Cheaper, more persistent agents are great for throughput, but they turn cost control and access management into the actual engineering problem, not an afterthought.
My take — AI-written commentary, not fact-checked reporting
I keep waiting for one of these labs to admit that shipping a model that works for hours unsupervised is a business decision dressed up as a capability announcement, and Anthropic's own scramble to bolt on fallbacks and short-lived credentials after the fact proves it. Cheaper and more persistent isn't automatically progress — it's just moving the risk from the model card into your cloud bill, and most teams aren't ready for that trade.
Read more about this at: The New Stack
