The Week AI Showed Its Work

Astra's machine-checked proofs, DeepMind's whole-body robots, a rebuilt agent protocol, and the EU's transparency deadline — July 27 to August 2, 2026.

Some weeks in AI are all launch trailer. This one was mostly receipts. Between July 27 and August 2, a model produced mathematics you can machine-verify, a robotics stack learned to use its legs, the protocol underneath most AI agents got rebuilt from the floor up, and the EU's transparency rules stopped being hypothetical. Here is what actually happened, and why it matters.

OpenAI's Astra published proofs, not a demo

On August 1, OpenAI previewed its next model family, Astra, in about the least product-launch way imaginable: no keynote, no pricing page — just ten solutions to open problems in mathematics and theoretical computer science, each unsolved for at least a decade. The results span group theory, von Neumann algebras, high-dimensional geometry, quantum complexity, lattice cryptography, and extremal combinatorics. The headline is the first explicit construction of a non-sofic group, a question left open since Mikhail Gromov introduced the notion of soficity in 1999. Astra also disproved Connes' Rigidity Conjecture.

What separates this from the usual benchmark victory lap is the verification. OpenAI published a 249-page manuscript alongside Lean 4 certificates for every result, so the proofs are machine-checkable rather than merely persuasive. The company says the tokens behind all ten solutions would have cost roughly $2,000 at its Sol API rates. Thomas Bloom, who runs erdosproblems.org, called the results big news. The fair caveat: this arrived by announcement rather than peer review, and humans still did the work of turning model output into papers.

DeepMind gives robots the rest of their body

Two days earlier, on July 30, Google DeepMind shipped Gemini Robotics 2 — three models rather than one: a vision-language-action model, an embodied reasoning model called Gemini Robotics ER 2, and an on-device version. The pitch is whole-body control. The previous generation moved a robot's arms; this one plans from feet to fingertips, so a humanoid can twist, lean, and reach as one motion instead of a sequence of poses.

The published numbers are refreshingly honest about how hard this still is. On an Apollo 2 humanoid with dexterous hands, unscrewing a light bulb succeeds 92% of the time — but picking an object up off the floor lands at 45.7%, and several fine-manipulation tasks sit in the thirties. ER 2 is available now through Google AI Studio; the action models are early-access only. Real progress. Not a solved problem.

The agent stack got a new foundation, and a lower price

On July 28, the Model Context Protocol landed its 2026-07-28 specification, the largest revision since launch. MCP is now a stateless request/response protocol: requests describe themselves, so any server instance behind a load balancer can handle one without shared state. Tasks, Apps, and enterprise authorization move into a formal extensions framework, while Roots, Sampling, and Logging are deprecated with a twelve-month offramp — as is the legacy HTTP+SSE transport. If you run MCP servers in production, this is a week to read the changelog closely. The payoff is that agent infrastructure finally scales like ordinary web infrastructure.

Costs moved too. On July 30, OpenAI cut GPT-5.6 Luna by 80%, to $0.20 per million input tokens and $1.20 per million output, and GPT-5.6 Terra by 20%, to $2 and $12. Flagship Sol pricing held steady. Read that as a plain statement about where competition actually bites right now: not at the frontier, but at the volume tier, where the margin lives.

The bill, and the fine print

Nvidia spent the week reminding everyone who is funding all of this. Reports on July 27 said the company is in talks to guarantee as much as $250 billion in lease payments so OpenAI can occupy a 10-gigawatt data center SoftBank is building in Pike County, Ohio — a project that could exceed $500 billion all in — alongside a roughly $500 billion AI initiative with SK Group in South Korea. These are talks, not signatures. And the structure is precisely what critics mean by circular financing: the chipmaker underwrites the customer who buys the chips.

Meanwhile the rules arrived. The EU's Digital Omnibus on AI entered into force on July 27, and as of August 2, Article 50 transparency obligations apply — tell people when they are talking to an AI, machine-mark synthetic content. The trade was time: high-risk obligations for standalone Annex III systems slipped from this month to December 2, 2027. And Apple, per the Financial Times, capped bug-report submissions and added a 30-day cool-off after a deluge of AI-generated security reports, becoming the first major vendor to formally rate-limit AI-assisted disclosure. Researchers can request higher quotas.

---

The takeaway

The interesting work this week was not in the demos. It was in the certificates, the success rates, the deprecation notices, and the compliance dates — the unglamorous scaffolding that decides whether any of this holds weight. When a lab ships Lean proofs instead of a highlight reel, and a robotics team publishes its 45.7% next to its 92%, that is a field learning to show its work.