“How long would a human have taken? If the agent can't match it, the credits come back.”

Put a ceiling on the AI bill

Cloudflare's AI Gateway now enforces spending caps measured in dollars across every provider you route through. Set a budget per model, team, or user; when it's hit, requests block — or downgrade to a cheaper model via Dynamic Routes. It's runaway-token insurance you can switch on today.

Unit

Dollars,
tracked live

Scope

Model ·
team · user

On limit

Block or
downgrade

ChatGPT sleeps on it.

OpenAI introduced “Dreaming,” a memory system built to keep your preferences fresh and relevant from one conversation to the next. The name hints at the method — revisiting and reorganising stored context rather than only appending to it. If you build on ChatGPT's memory, it's a behaviour change worth testing against.

Memory you no longer have to curate by hand — the assistant tidies its own notes between chats.

A model that goes to the lab.

OpenAI's GPT-Rosalind is a life-sciences specialist: biological reasoning, medicinal chemistry, genomics analysis, and help designing experimental workflows. Named for Rosalind Franklin, it's aimed at researchers who want a model fluent in the bench, not just the editor. If you work anywhere near bio, benchmark it against your current tooling.

errors, traces, logs, metrics — reach for which?

Sentry's field guide maps the four observability signals to the question each one actually answers, and argues that what you instrument today decides what you can debug six months from now.

The avant-garde, as a LoRA.

“Slop in the Face” is a Flux.1 LoRA trained on vectorised illustrations from Russian Futurist books of 1910–1914. Load it into a Flux pipeline and prompts emerge in the hard geometry and hand-cut energy of the early avant-garde — proof that a distinctive style is a few hundred well-chosen images away.

A refund on the hours the AI didn't save

Cognition will now estimate, per session, how long a human engineer would have taken to do what Devin did — and if the delivered value falls short of what you paid, it credits the difference, up to $10M a year. It puts the vendor's money behind the productivity claim; whether the estimator survives an audit is the open question.

Run your own agent, on your own machine.

Nous Research shipped a desktop build of Hermes, its open-source (MIT) agent, for macOS, Windows, and Linux. It holds persistent memory, schedules automation, browses, sees, runs code, and hands work to isolated sub-agents across five sandbox backends. A local, hackable alternative to the hosted agent platforms — v0.16 is a clean place to start.