AI this week: Claude Opus 5.5 leads, one-job models, two-model coding agents

Anthropic’s Opus 5.5 tops overall intelligence, TypeSafe’s Jev classifies anything cheaply, and Cognition pairs two models in one agent.

September 25, 2026 · By Bohol AI

A new leader in overall intelligence, a bet on small models that do one thing, and evidence that combining two models can match the best at lower cost.

Claude Opus 5.5

Anthropic released Claude Opus 5.5, a lower-cost successor to Opus 5 that outscores Claude Fable 5.1 and every other current model on overall intelligence. Unlike Fable, it doesn’t keep users’ data for 30 days. It arrived about a week and a half after Anthropic CEO Dario Amodei proposed slowing down AI development.

Jev: a model that only classifies

TypeSafe, founded by OpenAI alumnus Diogo Almeida, released Jev, a model that does one job: it reads text and answers questions about it from a fixed set of options. It is meant to be faster and cheaper than a large language model for that job.

Two models, one coding agent

Cognition introduced SWE-2, a model built for software engineering, and opened up Devin Fusion, a setup that runs a powerful model and a cheaper one together in one session. An independent evaluator found the pairing matched top models at lower cost.

Research: agents that pass notes

Carnegie Mellon researchers let parallel AI workers share partial results with each other directly, removing a bottleneck where every result had to pass through one central point.

What it means for Bohol

Many small-business jobs are sorting jobs: is this message a booking, a complaint or a question? A focused classifier can do that cheaply.
You don’t always need the most expensive model. Combining a strong model for planning with a cheap one for routine steps can cut costs.

Sources

All articles