Cheaper models, and new limits on agents

Three labs made strong models cheaper this week, each in a different way. Agents got a Senate bill, a sandbox warning and a call for hard spending limits.

1. OpenAI's GPT-6.1 Sol comes close to GPT-6 Astra at a fifth of the price

   TechCrunch

   For most features the cheaper model is the right place to start. I'd move a task up to the flagship only when a test on real inputs shows the gap.

2. Claude Sonnet 5.5 keeps its token price but needs fewer tokens per task

   Anthropic

   This is why price per token is the wrong number to compare. What matters is the cost of one finished task, and that can drop by up to 30% with the price list unchanged.

3. Gemini 4 Argon launches with an introductory price and limited access

   Google

   The $2 and $10 per million tokens is an introductory price, and it goes to $4 and $20 after. I'd budget on the second number, and access starts with cyber defenders anyway.

4. Cloudflare open sources Clef, models that pick an action instead of writing text

   Cloudflare

   A lot of what an agent does is choose from a short list, like reply, ask a question or hand off to a person. A model that can only return one of the allowed options is easier to test and can't make up a new one.

5. Agents in separate sandboxes passed instructions through a shared package cache

   Matthew Green

   Isolating where an agent runs isn't enough. Anything two agents can both read and write, like a cache, a file or a shared memory, works as a channel between them.

6. A Senate bill would make agent developers and operators liable for hacking damage

   The Daily Caller

   If it passes, whoever runs an agent answers for what it does. A log of every action and a person approving the risky ones would stop being optional.

7. Simon Willison makes the case for default hard budget caps on cloud services

   Simon Willison

   An agent stuck in a loop can spend a month's budget overnight. I'd rather a service stop at a limit I set than send a bill nobody approved.

Written by Ibrahim El Khansa, Technical Product Manager in Beirut, Lebanon.