issue #012 · July 31, 2026
GPT-5.6 Luna drops 80%, and OpenAI wants your open weights regulated
OpenAI cuts Luna token costs 80% — then lobbies to fence off open-weight models
| AI Week Radar |
#12 Fri, Jul 31 |
OpenAI is playing both sides this week: an 80% price cut on GPT-5.6 Luna aimed squarely at Anthropic and Google, paired with a lobbying push against the open weights that keep it honest. Meanwhile Stripe eyes OpenRouter for ~$10B, and DARPA lets an AI fly an F-16.
— Jarek
Featured
Advancing the price-performance frontier with GPT-5.6 — cuts GPT-5.6 Terra prices 20% and Luna a full 80%, with OpenAI crediting inference wins it found using its own Sol model. Per the note, Sol optimized the forward pass — trimming memory movement and synchronization overhead. The takeaway for builders: model-driven kernel tuning is now a real weapon to undercut Anthropic and Google on token costs. — simonwillison.net
Articles & Videos
POLICY
OpenAI and Anthropic unite against open-weight AI risks to their bottom line — the two labs are reportedly lobbying for stricter rules on open-weight models under a national-security banner. Read cynically, it's regulatory capture: tighter limits on Llama-class weights would threaten local inference for privacy-sensitive enterprise apps — and, conveniently, the cheapest alternative to their APIs. — yogthos
RUMOUR
Stripe in talks to buy OpenRouter for ~$10B — would put Stripe at the orchestration layer of the AI economy at a reported ~$10B. For builders that likely means model billing and API management collapse into one bill — convenient, until it's the only game in town. — htrp
DARPA, U.S. Air Force fly AI-controlled F-16 — ran the first in-air combat maneuvers on an AI-controlled F-16. Beyond the headline, it's a proof point for reinforcement learning in high-stakes, real-time physical control — hardware-in-the-loop, not another sim demo. — r2sk5t
Tools & Papers
OneCLI – OSS credential gateway that keeps secrets out of AI agents — proxies API calls so agents get scoped tool access instead of raw keys — "you give them access to specific tools through OneCLI." A sane answer to the security gap that opens when agentic workflows leave the sandbox and you'd rather not hand plaintext secrets to a non-deterministic model. — Jonathanfishner
Go LLM SDK for streaming, tool-calling AI backends (plus frontend React lib) — brings structured streaming and tool-calling to Go backends, with a React lib for the frontend. Grafana's answer for teams that live in Go and want production plumbing without dragging LangChain into the stack. — matryer
Palmier Pro – Open-source macOS video editor built for AI — a native Swift macOS editor that treats LLM-driven timeline manipulation as first-class, not a bolt-on plugin. Aimed at automated content workflows where the model, not the mouse, moves the cuts. — harrisontin
In Brief
Bot-detection startup Spur nabs $200M from Insight — As LLM-driven scraping and automated account creation become harder to detect, Spur's massive funding signals a growing market for sophisticated anti-bot infrastructure. — techcrunch.com
See you next week — keep your API keys off the agents.
Read on web · Manage preferences · Unsubscribe in one click
© AI Week Radar · 8cells Jarosław Kaczmarski · ul. Ketlinga 33a, 96-313 Jaktorów, Poland