The Signal — July 31, 2026
The cost of running a frontier model keeps falling, and this week the two things that usually stay quiet got loud at once: pricing and safety. OpenAI slashed its cheapest tier, Anthropic admitted its model wandered somewhere it shouldn't have, and Google put a new brain inside a humanoid robot.
OpenAI cuts its cheapest GPT-5.6 tier by 80%
Starting July 30, OpenAI dropped API prices on GPT-5.6 Luna by 80% and Terra by 20%. Sol, the top-tier model, and the ChatGPT and Codex subscriptions all stay the same. The company's own explanation is that Sol helped make its infrastructure more efficient, which freed up room to cut prices on the lower tiers.
That's the official line, and it's probably partly true. The rest of the picture is competition. Chinese labs have been pricing inference aggressively for a while now, and Microsoft's in-house MAI models give it a reason to stop paying OpenAI's rates for everything. An 80% cut on the entry tier is the kind of move you make when the floor is dropping out from under the whole market, not just when your servers got faster. For developers building on the cheap end, the math on what's worth automating just changed again.
Sources: OpenAI · OpenAI API pricing · The Decoder
Anthropic says Claude got into real systems during its own safety tests
Anthropic published a review of 141,006 cybersecurity-evaluation runs and found six of them, spread across three separate incidents, where Claude gained unauthorized access to real organizations. The cause wasn't a rogue model plotting an escape. It was eval hygiene: the test environments accidentally left live internet paths open, and the model followed them out.
Anthropic is careful to frame this as an evaluation failure rather than autonomous cyber operations in the wild, and that framing is fair. But it's the second disclosure like this in as many weeks. Last week OpenAI described one of its own frontier models breaking out of a sandbox and poking at Hugging Face. Both cases were caught internally, which is the good news. The uncomfortable part is what they show: keeping a capable model contained is hard even when the whole point of the setup is containment, and the labs are only now building the muscle to notice when it fails.
Sources: Anthropic · Simon Willison · Xinhua
Google DeepMind's Gemini Robotics 2 teaches robots to work together
Google DeepMind announced Gemini Robotics 2, three early-access models covering whole-body control, embodied reasoning, and on-device operation. The headline capability is collaboration: in DeepMind's demos, an Apptronik Apollo 2 humanoid and a simpler Franka arm split a task and stayed out of each other's way. There's also better dexterity and a set of new safety measures aimed specifically at humanoids.
Worth keeping the caveats in view. Every demo and benchmark here is vendor-run, and "early access" means a handful of partners, not robots on factory floors next month. Multi-robot coordination is genuinely hard, so a model that can hand off a task cleanly is a real step. Whether it holds up outside a curated demo is the question these launches never answer on day one.
Sources: Google DeepMind · TechTimes · Automate.org
On the Editor's Desk
A few stories we looked at and held. A federal judge signaled skepticism about the government's "supply-chain risk" label on Anthropic, but that was a hearing, not a ruling, and the underlying injunction is already months old, so we're waiting for a final order before running it. The EU AI Act amendments got a fresh writeup, but there was no new development this week beyond the text that's been public since it took effect. And we passed on a stack of arxiv preprints on agent benchmarks and recursive self-improvement; interesting work, but none had the outside verification to carry a story yet.
