All insights

Weekly14 August 20261 min read

This week in AI: capable models get cheaper

New model tiers from Anthropic and OpenAI bring near-flagship ability at lower prices, and let you choose how much thinking a task gets. That changes the sums for public services.

What happened

Over the summer, the major labs have been filling in the middle of their ranges. Anthropic's Claude Opus 5 offers close to the capability of its flagship, Claude Fable 5, at about half the cost. OpenAI released GPT-5.6 Sol as its flagship, alongside lower-cost Terra and Luna versions. Last week it updated Sol in ChatGPT with a control that lets people choose how much thinking a response gets.

Our take

For most organisations, the price of AI is no longer set by one model. It is set by how well you match models to tasks. Drafting a reply to a resident, summarising a long consultation and checking a contract for missing clauses are different jobs. The cheapest model that does each job well is usually not the same model.

Two habits help. First, keep a small set of test cases taken from your own work, with the answers you would expect. When a new model or tier appears, run the tests and compare quality and cost. Second, design systems so the model can be swapped without rewriting everything around it.

The thinking control is a useful idea in its own right. A quick answer to a simple question should be cheap; a careful answer to a hard one is worth paying for. Services that can tell the difference will cost less to run and give better answers.

Keep reading

Related insights