Custom silicon, Chinese open weights, and a double-blind eval push reset the frontier

Take vendor-published silicon benchmarks with salt, but the direction matters: if OpenAI can serve its own workloads on Jalapeño at claimed efficiency, per-token pricing pressure on frontier APIs accelerates through 2027. Model your inference budgets assuming another 30–50% price cut cycle, not stability.

The intelligence-per-dollar gap between US closed models and Chinese open weights is now small enough that self-hosted GLM-5.3 is a defensible option for coding-heavy internal tools where data residency matters. Worth an actual bake-off, not a dismissal.

If your rollout has been blocked on 'we can't audit who's spending what,' this removes the excuse. Get baseline usage telemetry live in week one — you cannot govern what you don't measure, and the shadow-IT ChatGPT problem becomes a shadow-agent problem fast.

This matches what delivery teams see firsthand: prompting, retries, and tool wiring move task success more than most model-version bumps. Version your harness like you version your model, and log both in evals — 'we upgraded to Claude 4.5' is not a sufficient changelog entry.

A sub-$400 bipedal RL platform meaningfully changes who can run embodied AI experiments — universities, hobbyists, and small robotics teams can now iterate on locomotion policies without a Boston Dynamics budget. Watch for the training data flywheel this creates.

If you're using public benchmarks to pick models for enterprise deployments, stop. Build a private eval set from your actual tasks — the double-blind push is an admission that even the labs no longer trust the numbers they publish.

A frontier lab voluntarily pausing training runs for infrastructure security is a governance datapoint worth citing in your own AI risk committees. The 20% monitoring overhead is also a useful anchor when internal stakeholders push back on observability costs.

Add 'change of control' scenarios to your model-vendor risk register. If your product depends on a single frontier API, understand what happens if your company — or your vendor's competitor — gets acquired by someone on the wrong side of a lab's strategic map.

The interesting pattern here is not the 68% — it's using AI to route around resource constraints on fixed-deadline work. That's a cleaner ROI story than 'productivity' gains that get absorbed into scope creep. Ask your teams where the deadline-plus-no-headcount pockets are.