Anthropic Opus 5 update focuses on token costs for coding tasks
Anthropic released Opus 5 with modest benchmark gains and Fable-like coding performance at lower token costs, according to Ars Technica.
By Hana Yoshida · Markets Reporter
3 min read
Anthropic Opus 5 has been released as the latest version of a model widely used for coding and software development work, according to Anthropic and Ars Technica. The update matters less as a major capability jump than as a bid to offer near-frontier performance at a lower token cost.
Ars Technica reported that Opus 5 appears to be a meaningful update, but not the kind of agentic coding advance associated with the earlier Opus 4.5 jump. Anthropic’s own benchmark chart shows Opus 5 roughly matching or slightly exceeding the coding-task performance of its Fable model on tests including Frontier-Bench and DeepSWE.
The same Anthropic chart indicates that Opus 5 beats Opus 4.8 and OpenAI’s GPT-5.6-Sol across most categories. Ars Technica described the gains as incremental rather than dramatic, with the main selling point being performance close to Fable for about half the price.
How much does Anthropic Opus 5 cost?
Anthropic prices Opus 5 at $5 per million input tokens and $25 per million output tokens, Ars Technica reported. That matches the prior Opus pricing while undercutting Fable.
A token is a unit of text processed by an AI model, so token pricing affects how expensive it is to run prompts and receive answers at scale. For software teams that use coding agents heavily, output-token costs can become a direct budget concern.
Ars Technica noted that cheaper competition is increasing pressure on Anthropic. The recently announced Chinese open-weight model Kimi K3 is priced at $15 per million output tokens while offering similar performance, according to the report.
What are Opus 5’s cybersecurity limits?
Anthropic did not give Opus 5 the same cutting-edge cybersecurity training as some of its other models, according to Ars Technica. As a result, the model trails Fable and Mythos in that area.
Anthropic says Opus 5 is relatively capable at finding cybersecurity vulnerabilities, Ars Technica reported. The company also says training choices leave it “substantially behind Mythos 5” on exploiting those vulnerabilities.
Because Opus 5 lacks the same cybersecurity profile as Fable, it also does not carry all of Fable’s more disputed safeguards, according to Ars Technica. Those included a policy of retaining data for 30 days for review if an incident occurred.
Why token efficiency is becoming central
Ars Technica reported that software developers and engineering managers are paying close attention to AI coding costs, while open-weight models, local models and other alternatives continue to improve. That puts pressure on frontier-model providers to lower token expenses or deliver more performance for the same price.
Some companies are trying to cut costs with model routers, according to Ars Technica. Cursor and Meta have been building systems that choose among models of different sizes and capabilities based on the prompt, reducing the need to send every request to a top-tier model such as Fable.
Ars Technica reported that Anthropic and similar companies will need to keep lowering token costs, or provide better results without raising prices, to sustain growth. If smaller or open models become good enough for routine development tasks, users may shift more work away from the most expensive frontier systems.
This story draws on original reporting from Ars Technica.