Anthropic ships Opus 5: a new release for long-running agents

The Opus tier was refreshed within two months

Anthropic announced Claude Opus 5 on its own newsroom and made the model available across its platforms. The framing is explicit: this release is a step change for the Opus tier that runs long-running agents. The highlighted areas are coding and knowledge work; the company also says the model performs strongly in scientific research, including organic chemistry and protein analysis, and can produce interactive visualizations and 3D models. The release arrived roughly two months after Opus 4.8 shipped at the end of May.

A model that verifies before it moves on

Anthropic's evaluation claims run as follows: the model leads on coding and knowledge-work measures including Frontier-Bench and GDPval-AA; on ARC-AGI 3, an evaluation of novel problem solving, it scores three times as high as the next best model; on Zapier AutomationBench it reaches a 1.5 times pass rate against the next best; and on OSWorld 2.0, a computer-use benchmark, it surpasses Fable 5's result at roughly a third of the cost. The behaviour the company underlines is the one that matters most for agents: the model verifies its own output and iterates carefully until it succeeds.

Pricing is set at 5 dollars per million input tokens and 25 dollars per million output tokens, the same level as Opus 4.8, while a fast mode that runs at two and a half times the speed costs double the base rate. Two further points stand out in the trade coverage: safety classifiers are expected to engage significantly less often than with Fable 5, and a beta feature called Automatic Fallbacks routes restricted requests to smaller models instead of returning an error.

Agent economics are measured in finished work, not unit price

With long-running agents the real expense is not a single response but the total tokens spent until a job is done. That makes headline unit pricing misleading on its own: a model that finishes the same work with fewer attempts and fewer round trips can cost less overall even when it looks more expensive per token. The verify-then-iterate behaviour highlighted in this release targets exactly that calculation.

The second point is access friction. In enterprise use, agents often stall not because of missing capability but because of refusals that arrive mid-flow. Routing a restricted request to a smaller model instead of returning an error is meaningful for keeping the flow intact. The third is that the trade-off between speed and cost is becoming a setting: the same model can be run with different economics depending on how urgent the job is.

What it means for businesses in Türkiye

Model versions change fast; the system you build around them lasts. The practical conclusion for businesses in Türkiye is to stop designing an agent project around one model version and instead assume the model will change. A structure that makes the provider and the version selectable in configuration avoids rewriting the work with every release.

The second practical conclusion is measurement. Because the cost of running agents now depends largely on the number of attempts, you cannot compare options without seeing how many steps and how many tokens each task consumed. Running two models under identical conditions on three to five representative tasks from your own operation tells you more than any launch post. The third is boundaries: however capable the model is, hard-to-reverse steps such as payments, cancellations and price commitments should stay under human approval. Turkish output quality and data processing that complies with Türkiye's data protection law also need retesting at every version change.

Sources

Anthropic, "Claude Opus 5" announcement · TechCrunch, "Anthropic launches Opus 5"