Anthropic Opus 5 is now available, giving the company’s coding-focused Claude line a new model that appears aimed at better economics more than a dramatic jump in raw capability. Anthropic’s own benchmark chart shows Opus 5 landing around the level of its higher-profile Fable model on coding tasks, while costing less to run.
According to Anthropic, Opus 5 performs slightly ahead of Fable on some benchmark views and beats Opus 4.8 and OpenAI’s GPT-5.6-Sol across a range of tasks. Ars Technica’s Samuel Axon described the gains as incremental rather than a repeat of the larger agentic-coding step seen with Opus 4.5.
What is Anthropic Opus 5?
Opus 5 is Anthropic’s latest update to its Opus model, which has become a common choice for programming and other software development work. In practical terms, the pitch is that users can get performance near Fable without paying Fable-level prices.
Tokens are the small text units that language-model APIs count for billing and compute. Anthropic lists Opus 5 at $5 per million input tokens and $25 per million output tokens, the same price as its predecessor, according to Ars Technica. Fable is described as more expensive, while the Chinese open-weight model Kimi K3 was recently announced at $15 per million output tokens with similar performance claims.
That makes Opus 5 a cost-efficiency release in a market where software teams are watching AI bills as closely as model scores. A model that is modestly better for the same price can matter if it is being called thousands or millions of times inside coding tools, test agents, and developer workflows.
Why does Opus 5 trail Fable and Mythos on cybersecurity?
Anthropic did not give Opus 5 its most advanced cybersecurity training, according to Ars Technica’s account of the release. Anthropic says the model can still find security vulnerabilities relatively well, but it is “substantially behind Mythos 5” at exploiting those vulnerabilities.
In this context, exploitation means moving from identifying a flaw to using it in a working attack. That distinction matters because a model that can only point out a bug creates a different risk profile than one that can help weaponize it.
Because Opus 5 does not carry the same cyber capability profile as Fable, it also does not come with all of Fable’s more controversial protections. Ars Technica cited Fable’s policy of retaining data for 30 days for review in case of an incident as one measure that does not apply in the same way to Opus 5.
Why token costs are driving model choices
The release lands as developers and engineering managers are trying to keep model spending under control. Ars Technica noted rising interest in open-weight models, local models, and other alternatives to frontier systems once those alternatives can handle routine development tasks.
One response is the model router. Cursor and Meta have been building systems that pick from multiple models depending on the prompt, according to The New Stack reporting cited by Ars Technica. The mechanism is blunt but useful: send easy jobs to cheaper or smaller models, and reserve expensive frontier models for tasks that need them.
That puts pressure on Anthropic to lower token costs or deliver more capability at the same price. Opus 5 appears to take the second route: a better model on paper, a familiar bill, and no claim of a coding breakthrough large enough to end the cost argument.
This story draws on original reporting from Ars Technica.
