GPT-5.6 lands, ChatGPT Work debuts, and agent evals get harder

Tiered effort levels turn model selection into a per-call cost/quality knob — plan routing logic, not just a default model. The cache-write pricing changes prompt architecture math; long system prompts get materially cheaper if you actually pin them.

Grok's serious enterprise problem remains the brand and data-governance posture, not the tokens. But if Cursor is training against it, expect its coding-agent numbers to be real — worth including in bake-offs for teams already comfortable with xAI.

Voice is the surface where model quality changes are most obvious to end users. If you have a voice deployment on prior Realtime API models, budget an eval cycle — barge-in behavior and prosody changes tend to break scripted flows.

The open frontier keeps compressing. For workloads where you already run vLLM and need sovereign inference, Hy3 plus GLM-5.2 give you two credible options to benchmark against closed models — with the caveat that Chinese-origin weights raise procurement questions in some verticals.

This is OpenAI's direct move on the 'AI coworker' territory Anthropic staked with Claude Cowork. For enterprise buyers already on ChatGPT Enterprise, expect procurement pressure to consolidate agent spend here rather than stand up a separate agent platform.

For teams working on physical AI, the value here is the eval methodology, not any single policy. Cutting hardware iteration cycles is where robotics economics have been stuck.

Any vendor pitch citing SWE-Bench Pro numbers this quarter needs a follow-up question. For internal model selection, weight your own private repo-based evals higher than public leaderboards — this is the third major coding benchmark to get called into question in six months.

If your architecture still treats the LLM as the scarce resource, your cost model is a year out of date. The design shift worth prioritizing: durable agent memory, structured tool catalogs, and observability primitives that assume 10x more model calls per user action.

The scope is the story. Most enterprise AI still lives as bolt-ons; DT is publicly committing to workflow-level redesign. Worth pulling apart with your own transformation roadmap to see which pillars they attacked first.

If your Copilot pilots stalled on quality, rerun the eval — the underlying model is materially different this week. Also worth pressure-testing whether the new model changes your data-loss-prevention and prompt-injection threat model in Office.

For regulated financial services, MUFG going public matters more than the specifics — it lowers the internal risk-committee bar at other banks. Expect procurement conversations that previously ended at 'no LLM in production' to reopen.