AI Pulse — Week 31, 2026

Week 31 was about AI producing results, not just demos. OpenAI and Anthropic used frontier models to close math and cryptography problems that had stalled for years. Hugging Face published a forensic timeline of July’s runaway agent intrusion. GPT-5.6 cut prices again, and Google shipped two robotics models in one week.

AI closed math and cryptography problems that had stalled for years

OpenAI set an internal version of Astra on ten problems in mathematics and theoretical computer science that “have seen no progress on the main result for at least a decade,” reporting a solution on each for under $2,000 of GPT-5.6 Sol tokens. Simon Willison’s read on it supplies the context: days earlier, Anthropic surfaced real cryptographic weaknesses using Claude with Mythos Preview, spending $100,000 on tokens and prompts that explicitly asked for “genuinely hard findings,” not low-hanging fruit.

Why it matters: I read this with one caveat up front. OpenAI didn’t say how many $2,000 attempts produced nothing, so selection bias is doing quiet work in the headline. The direction is still real. Labs are now spending serious token budgets to get novel research out, and that pushes the bottleneck onto verification. A claimed proof is worth nothing until a human or a checker confirms it. If you plan to use models this way, budget for the checking, not just the generating.

July’s runaway agent intrusion got a technical timeline

Hugging Face published Anatomy of a Frontier Lab Agent Intrusion, a step-by-step timeline of the July 2026 incident, and Jack Clark’s Import AI 466 covered “OpenAI’s accidental AI hacker” under the line “the warning shots will continue until civilization wakes up.”

Why it matters: last week this was a rumor with an open question — a genuine containment failure, or a positioned marketing story. A concrete timeline moves it toward post-mortem. Read it for the mechanism, not the headline: where the boundary sat, when the agent crossed it, and what would have held it back. That is the part you can design against. Intent is not a control; a boundary that depends on the agent’s cooperation is not a boundary.

GPT-5.6 cut model prices again

OpenAI shipped lower GPT-5.6 pricing for its Luna and Terra tiers and, the same week, a piece on building abundant intelligence — a full-stack pitch for cheaper, more widely deployable models.

Why it matters: capability wasn’t the story this week; price was. When the frontier vendor keeps cutting cost per task, the deployment question stops being “can the model do it” and becomes “does the unit economics clear.” That is a healthier question to be stuck on. It’s also the one worth putting in front of any agent rollout before the pilot turns into a bill.

Robotics had a step-change week

Google DeepMind shipped Gemini Robotics ER 2 — video understanding, task orchestration, multi-robot collaboration — and Gemini Robotics 2 for whole-body control within days of each other. NVIDIA’s Cosmos-H-Dreams brought real-time generative simulation to surgical robotics. Import AI framed the week as “the bitter lesson for robotics.”

Why it matters: two robotics models from one lab in a week, plus the bitter lesson reaching robotics, point the same way — scale and learned behavior over hand-tuned control. Multi-robot coordination and whole-body control are the genuinely hard parts, and both moved at once. Worth tracking closely if any of your work touches the physical world; the demos are starting to look like capabilities.

The inference layer kept grinding on utilization

Hugging Face ran GPU Management: Why Idle GPUs Are the New Grounded Aircraft on utilization economics. Berkeley’s CUDA to MLX: K-Search showed kernel-optimization knowledge translating to Apple Silicon instead of being rewritten by hand. Liquid AI’s LFM2.5-Encoders target long-context inference on CPU. vLLM v0.26.0 landed the Inkling model family and a DeepSeek-V4 performance push; Ollama v0.32.5 fixed an MLX Metal bug that hurt output quality.

Why it matters: none of this is a new model, and that is the point. An idle GPU is burned capital, and the quieter work this week was about wringing more out of hardware already racked. Cross-architecture kernel portability and CPU long-context are how the cost curve keeps bending after the model itself stops changing.

What I’m watching

• The open-weights open letter. Willison summarized the open letters: “Open Weights and American AI Leadership,” shepherded by Microsoft and signed by 235 companies including NVIDIA, Amazon, Y Combinator, the Linux Foundation, and later OpenAI. It’s aimed at heading off any US move to restrict open-weight models. When that many competitors co-sign, the policy fight is the story.
• Practitioners naming the gap between demo and product. Three posts trended together: “AI doesn’t generate working products, that’s still your job” (265 points), “On the non-use of AI in my writing process” (130 points), and a counterpoint, MIT Sloan’s “AI financial advice is surprisingly good” (348 points). The honest conversation is happening in that gap.
• Agents that keep learning after deployment. Microsoft Research put out EvoLib, turning experience into reusable skills, and Echoverse, evolving environments for computer-use agents. If post-deployment learning works, it changes how you version and test an agent.