Opus 5 moves the price of performance, not the frontier of capability

Anthropic released Opus 5, and Ars Technica frames it mainly as a token efficiency move rather than a major capability jump. The model is priced at 5 dollars per million input tokens and 25 dollars per million output tokens, in line with the previous Opus, while aiming to offer performance close to more expensive top tier models.

Opus 5 is selling a better price to performance ratio

According to Ars Technica, Anthropic’s benchmarks show iterative improvement over Opus 4.8 and competitors such as GPT-5.6-Sol, but not a radical breakthrough on the scale of earlier agentic jumps. The main argument is different: something close to the highest tier at roughly half the cost of Fable.

That matters more than it sounds. In coding and agentic work, teams increasingly do not ask whether a model can write a piece of code. They ask how much it costs to let the model think, repair itself and call tools repeatedly.

Ars also notes a cybersecurity tradeoff. Anthropic reportedly avoided the sharpest training on exploitation tasks, leaving Opus 5 behind Mythos and Fable in that area.

Model routing is becoming a requirement, not an optimization

For developer organizations, the practical impact is clear. Not every prompt needs the strongest model. If Opus 5 offers enough capability at a lower effective cost, it belongs in the layer that decides when to use an expensive model and when to use a cheaper alternative.

Ars mentions that Cursor and Meta are building model routers: systems that choose a model based on the task. That is less flashy than a new leaderboard, but more important for enterprise costs.

When an agent runs for a long time, the difference between a good model and the best model compounds across tokens, retries and parallel attempts. Inference economics becomes a product feature.

A cheaper Opus does not remove open-weight pressure

Anthropic’s weak spot is competition from below. Ars cites Kimi K3 at 15 dollars per million output tokens with similar performance. If cheaper open-weight models keep improving, customers will move routine tasks away from the most expensive frontier APIs.

That does not end Opus 5. It ends the comfortable assumption that the top model should receive every request just because it is best. Performance now has to justify the bill.

Adoption inside routed workflows will decide

The next signal is not just another benchmark. What matters is whether Opus 5 lands in model routers as the default for harder coding tasks, or merely as another line in the price sheet.

If teams use it where Fable feels expensive and smaller models fail, Anthropic has found the real problem. If not, this is another release that looks good on a chart and less convincing on an invoice.

Lilith's verdict

Opus 5 is aimed at the person who no longer admires the demo, but holds the invoice for millions of tokens. At that moment, the smart model is the one that fits on the receipt.