- Agent bills need hard caps
+ Kolibri: open weights from Germany
- COSMIC: no LLM content in PRsA cloud spending cap can save your budget while putting your project on a recovery clock. AWS's new project limits stop resources at the ceiling and initially preserve the data. Its documentation adds a condition: after 90 days paused without action, AWS permanently deletes that project's data.
Today's news peg is Simon Willison's October 3 call for default hard budget caps. AWS announced its simplified builder experience on September 16, and Google announced a related feature in July. The discussion is fresh; the launches have their own dates.
The bill needs an actual stop
Willison describes the familiar midnight budget warning: your email arrives, the service keeps spending, and you wake up as the involuntary sponsor of an infinite loop. Coding agents make it easier to provision useful software, including software that calls paid APIs or spins up cloud resources. The account owner still gets the invoice.
AWS's project limit provides actual enforcement. The new experience is rolling out to a limited number of customers, and creating a limit requires a paid plan. Check availability for your project before treating an existing account as covered.
The ceiling applies to pre-tax charges and excludes credits. Its minimum is the greater of $20 or a conservative estimate of usage. That estimate considers activity and running resources; a sandbox already full of machines may need a cleanup before you can set a smaller limit.
Optional early controls can block new launches about a week before a forecast breach. Existing resources keep running, although autoscaling may lose the ability to expand. Other options pause idle resources and top cost drivers. A forgotten SageMaker endpoint with no invocations for 14 days qualifies as idle. The demo finally gets an exit interview.
Google's cap has a different boundary
Google Cloud's July public-preview announcement describes a cap on one service within one project. Supported services include Gemini API and Cloud Run. Google says AI-service caps trigger within minutes; that is its claim about enforcement timing.
Outside-scope services stay unaffected. Fixed contractual commitments continue billing. The cap can halt new on-demand usage while the contracts you've already signed keep their place on the invoice.
For experimental agents, I'd use a separate project, check the covered charges and assign someone to recover it when a limit fires. Production services need their own decision about whether an interruption is acceptable.
Kolibri's small bird needs a big cage
Aleph Alpha released Kolibri on October 3: an open-weight German-English model under Apache 2.0. About 3.46 billion parameters are active per token, out of 78.1 billion total. The published FP8 weights occupy about 78 GB, and the full model must be held in memory.
Tejas Kumar independently measured its tokenizer on Germany's Basic Law and the English translation. Kolibri used about 15% fewer tokens than OpenAI's tokenizer on the German text; on the English translation they were effectively tied. He says he hasn't run inference with the model. That's useful evidence for German text efficiency, with a clear limit on what it proves.
COSMIC puts a boundary on contributions
The COSMIC compositor's actual pull-request template asks contributors to certify that they included no LLM-generated content, explicitly covering code, comments and descriptions. Contributors also promise to understand and test their work.
The cosmic-flatpak template contains a specific exception for a contributor's own app source code and manifest. Read the relevant repository's template. The machine can produce another patch in seconds; somebody still has to review it.
The countdown is in the documentation
AWS says increasing the spending limit can reactivate a paused project, with some resources needing manual restarts. The 90-day no-action condition follows that recovery guidance. A stopped bill and a retained backup are separate jobs; keep important backups outside the paused project's failure boundary.
Verdict: NEEDS REVIEW. I welcome enforced spending limits. I'd verify availability, covered charges and recovery before handing an agent the account. A 90-day deletion clock deserves a very visible warning.
Sources
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.