September 2026 AI updates: superintelligence, DevDay, and the real cost curve.

A freight broker in Ohio, a medical-billing clerk in Phoenix, a parts distributor in Rotterdam: the last week of September 2026 gave all three the same two headaches. The vocabulary moved again, and the invoice moved with it. This note separates what is on the record from what we are seeing in client bills.

We verified every reported claim below against primary pages fetched on 30 September 2026. Where we could not verify, we say so, and we frame our own view as analysis — never as reporting.

Superintelligence becomes policy

On 22 September 2026, at the UN General Assembly, President Trump said the United States would call artificial intelligence Super Intelligence. The verbatim lines from that address — and the ones that did not check out — are in our companion note, What is superintelligence? The UN AI rename, explained.

Reported, and per White House social posts in the closing days of September, the President has kept using superintelligence as the operative word for AI at public events since the UN address. The White House video library lists a 43-minute announcement address dated 28 September 2026. We could not obtain a verified transcript of those remarks at writing time, so we do not quote them here — the pattern, not the wording, is the news. One speech is a headline; a repeated word is a vocabulary.

Why it matters for a business is simple: when the US government adopts a term in its documents and events, vendors rewrite their marketing to match, then the UK and EU follow with their own framing. Your next compliance form or enterprise procurement template will likely ask about "superintelligence" posture. The definition we stand by — it is still artificial — is the one to keep in your own documents.

OpenAI Dev Day and 5.5

OpenAI's DevDay 2026, on 29 September, was — per the company's own recap — its biggest yet, with "more than 20 major announcements across ChatGPT, Codex, our models, and entirely new forms of working with AI." Two structural changes stand out for small businesses: agents that take on ongoing responsibilities, and ChatGPT opening as a shared surface where developers can launch experiences to what OpenAI calls its "collective 1.2B weekly users." If your customers live inside ChatGPT, your storefront may soon need to as well.

The model news was led by GPT-6.1 Sol, introduced the same day: near-frontier intelligence "for a fifth of the price," with cached input at $0.10 per million tokens. On Terminal-Bench Science, OpenAI reports it averaging $5.47 per completed task, against $23.21 for Opus 5.5 — a real, benchmark-level drop.

The 5.5-generation model that set this year's bar is GPT-5.5. OpenAI's own numbers: 82.7% on Terminal-Bench 2.0 (versus 75.1% for GPT-5.4), matching GPT-5.4's per-token latency while using significantly fewer tokens to finish the same coding tasks — and, per the release, state-of-the-art intelligence at roughly half the cost of rival frontier coding models on Artificial Analysis's index. That is genuine performance, and we do not discount it.

The cost curve: dollar-per-task

The rest of this section is NexFlow's assessment, not reported fact.

What we're seeing in our own workflows and client bills: the effective dollar-per-task cost of AI work is rising for OpenAI, Anthropic's Claude, and Grok — even as every vendor's per-token price falls. The models are cheaper per token and more capable per call. But agentic work is not a single call. A modern agent reads more context, calls more tools, retries failed steps, and re-plans when a step stalls. Token prices drop; tokens consumed per completed task climb faster.

There is even a documented example of the hidden-cost pattern, from OpenAI's own GPT-6.1 Sol page: the published benchmark cost for Claude Fable 5.1 "understates its actual cost, as it omits the cost of fallbacks, which occurred on ~40% of tasks." In our assessment, that is the whole September story in one sentence — the sticker metric and the real bill have quietly parted ways.

In our assessment, the new subscription tiers read as intentionally vague. What a Pro plan includes is generally being reduced — caps, priorities, and quotas reshuffled in language vague enough that most buyers cannot comparison-shop across vendors. Vagueness is not an accident; it is a pricing strategy.

The fix is measurement, not cynicism. Track one number monthly: dollars per completed task — per triaged inbox, per drafted quote, per reconciled invoice. Our Field Notes exist because that number is knowable, and because it behaves differently from the one on the pricing page.

What this means for small business

A two-person brokerage in Columbus, Ohio does not need a position on superintelligence. It needs to know that the follow-up emails still go out, the load-matching still runs, and the monthly AI bill stays predictable. That is a workflow question, and workflows are what you own — see how we build AI assistants you actually own.

Three practical moves for the next quarter. First, keep the rename in procurement language only; your internal docs should still say what the system is. Second, when a vendor restructures its tiers, re-measure dollar-per-task before you re-up — the plan you are renewing is usually not the plan you bought. Third, keep a human on the approval step for anything that leaves the building. Super or artificial, the audit trail does not care.