Yesterday's signals, distilled.

Yesterday's signals, distilled, A look back at April 23, 2026.

GPT-5.5 hit latency parity with its predecessor while stepping up intelligence. NVIDIA and OpenAI tightened their integration around Codex and agentic workflows. Anthropic publicly treated model regressions like an SRE incident, not a PR problem. At the same time, Meta and Microsoft re-cut their org charts around AI leverage, and capital flowed into embodied automation and knowledge-cloning agents.

The throughline: AI is no longer a “tool” bolted onto existing structures. It’s becoming the organizing principle for how compute is bought, how talent is managed, and how workflows are encoded.

Model performance is stabilizing at “fast enough.” The bottlenecks are shifting to UX, verification, and organizational willingness to actually remove humans from loops. Meanwhile, labor and knowledge are being repriced, either automated, cloned, or re-benchmarked against AI-native baselines.

If your 2026 plan assumes “we’ll experiment with AI and see where it fits,” you’re already behind. The game now is: what do you stop doing, who do you stop hiring, and which parts of your institutional memory do you turn into software before someone else does.

MODELS / STACK

GPT-5.5 makes speed a solved problem, your UX and control layer are now the constraint

OpenAI said “GPT-5.5 matches GPT-5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence,” per Techmeme.

In parallel, OpenAI detailed that GPT-5.5 is powering Codex on NVIDIA infrastructure, and that NVIDIA is already using it internally for agentic workflows and dev tooling, per NVIDIA.

The Bet: Model providers and infra vendors are assuming that “good enough” latency is table stakes and that differentiation moves to depth of integration and workflow ownership.

So What? Model speed is no longer a credible excuse for shallow automation. If you’re still gating deeper use cases on “LLMs are too slow,” you’re hiding a product and governance problem behind a technical one. The structural shift is that the default enterprise stack is converging on “frontier model + GPU giant + agentic orchestration,” and the value is accruing to whoever owns the workflow surface, not the raw model API.

The Risk: If you over-index on a single vendor stack without abstraction, you’re locking your roadmap to their pricing, safety policies, and outage profile. And if you rush into agentic automation without a verification layer, you’re just trading human error for opaque machine error at higher speed.

Action: • Audit every AI feature in your product: flag where “latency” is the stated blocker and force a redesign conversation around UX and ver