The open-weight frontier stopped being a compromise
Kimi K3 landed, GLM shipped twice in a week, and self-hosting is now a serious answer for more workloads than it was in spring.
The open-weight tier is no longer where you go to save money at the cost of quality. It is where you go when you need control, and the quality penalty has become small enough that the argument now turns on data residency and cost rather than capability. I have moved two client workloads onto open weights this month and neither one lost anything measurable on the customer's own eval set. Neither decision was about ideology. Both were about not wanting a vendor to change my model underneath me without asking.
GLM-5.3 on the 14th, then GLM-5.2 Turbo on the 17th. Shipping a headline model and a cheaper fast variant in the same week is a deliberate pricing move, and it is the right one. Most production traffic does not need the frontier tier. It needs something good enough that runs at a price you can put in a spreadsheet.
Worth being clear about what did and did not happen on the second of August, because a lot of the commentary got it wrong. The transparency duties and the AI Office's enforcement powers over general-purpose providers arrived on schedule. The high-risk obligations for stand-alone systems did not. Those moved to December 2027, and high-risk AI embedded in already-regulated products moved to August 2028. If you have been treating those as imminent, you have more time than you think, and if you have been treating the GPAI duties as distant, you have less.
Hardened the cost controls in Hiveclaw again, this time around the case where an agent retries into a rate limit. The general principle I keep relearning: every retry path needs a budget, not just the happy path. Failures cost tokens too.
One email a week. No noise, easy unsubscribe.