The Month AI Got a New Name, a New Interface, and Little New Evidence

Three things about AI changed in the past month, and none of them was an independent test.

The first was its name. The second was its interface. The third was how quickly new models arrive with impressive numbers attached. I run BestSuperintelligent.com, a research site that publishes no score without evidence, so I watched all three from an awkward seat: the one where you must decide, every day, what you are actually allowed to say.

We began publishing on October 10. This is not a retrospective of a long track record. It is what a month of AI news looks like from a desk with a strict rule, and what that rule turned up.

What changed in AI over the past month?

Here are five developments, each with who is making the claim and what kind of evidence stands behind it.

Four of five rows are the vendor describing its own product. For none of them did we find an independent measurement.

Does Executive Order 14434 make AI superintelligent?

No, and the text of the order says why. Section 2(a) directs executive agencies to use the terms Super Intelligence and SI instead of Artificial Intelligence and AI in non-statutory documents, to the extent the law permits. Section 3(a) then defines the new terms by pointing back to the existing statutory definition of AI in 15 U.S.C. § 9401(3). Section 3(b) asks for proposed legislative language on a federal definition within 60 days, which lands around November 28. Section 4(c) says the order creates no enforceable rights.

Our reading: it renames a term for federal use and changes nothing a product can do. We set that out section by section in Executive Order 14434: what it changes for AI buyers, and what it doesn't, and Tal Slavin, our writer and editor, took the political angle in Trump Wants to Rename AI “Super Intelligence.” What Actually Changes?. The order is recorded with its primary source in our policy tracker.

The practical effect for buyers arrives through marketing, not law. Expect vendors to adopt the new word. A label is not evidence of capability, which is why we wrote Superintelligence and today’s tools: keeping the terms apart. Our own name is a brand, not a claim, and the site says so on its homepage.

What does Intelligent UI change for people choosing AI tools?

A chat reply that arrives as a calculator, a form or an editable chart is software generated on the spot. That moves the question a buyer has to ask. It is no longer only whether the words are right. It is whether the arithmetic inside the generated widget is right, where the numbers you typed into it go, and whether a retention policy covers them.

We have not tested Intelligent UI, and we have found no independent measurement of it, so we will not pretend to a verdict. What we can offer is the checklist. Five questions about AI data retention and training applies to any tool that takes your inputs. How to trial an AI research or writing tool fairly gives a repeatable method for testing on your own work. If the tool touches company data, how to read a vendor security page separates what a vendor states from what is evidenced.

Why are this month’s benchmark numbers hard to compare?

Two headline figures arrived within days of each other: 72.6% for one model and 72.4% for another. They look like a tie. They are not a comparison. One is reported on an offline subset of OSWorld 2.0, the other on OSWorld 2.1, and in both cases the vendor ran the test.

That single pair breaks three of the eight questions on our benchmark-claims checklist: who ran it, which version was used, and whether the comparison is like for like. The others on that list are worth asking of any number:

• How was success judged, and could a different judge reach a different result?
• How many items and how many runs? Small samples produce unstable numbers.
• Could the test material have leaked into training data? Benchmark contamination explained covers how that happens.
• Is the gap between two numbers large compared with run-to-run variation?

None of this says the vendor figures are wrong. It says they are claims, and a claim is a different thing from a result. We label the difference on every page: vendor-reported, independently observed, independently measured, third-party research, editorial interpretation, or unverified. Vendor claims vs. verified findings explains each label.

What did building BestSuperintelligent.com teach us?

Five things, in the order they hurt.

“Not tested” is a sentence you have to be willing to print. Our evaluation queue lists nine tools, among them Claude Code, Cursor, GitHub Copilot, OpenAI Codex and Gemini Code Assist, and every one is marked Planned, with zero independent tests. That looks unimpressive next to a launch announcement. It is accurate. See the full evaluation queue.

An editorial ranking and a scored ranking are different products. Our ranking of the 15 best AI apps of 2026 weighs adoption, reach, breadth and practical usefulness, and says in its first lines that it is not tested results. A scored ranking has a higher bar under our methodology: at least two fully evidenced, published evaluations in the same category. No category has met it.

Policy moves faster than product reviews, and commentary goes stale. Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved the AI Act’s high-risk rules to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems, and added new prohibitions from 2 December 2026. Plenty of articles still quote August 2026. We wrote the correction up in EU AI Act high-risk dates: what the Digital Omnibus changed and track the act in the policy tracker.

Security claims need scope, not slogans. A vendor being secure does not mean each of its products holds a given authorization. Our security evidence page records claims by service, deployment, scope, evidence and verification date, and the procurement guidance lists the questions to settle before you buy.

Corrections belong in public. Every change to the site goes on a dated updates and corrections page. It is the least glamorous page we have and the one I trust most.

Where the site stands today

Those zeros are the point. They are what the site’s own data reports, and they will change only when a test produces something worth publishing.

What should you do with this month’s AI news?

1. Treat launch numbers as claims until a second party reproduces them. Ask who ran the test and on which version.
2. Check what a new word is defined as. If a document says Super Intelligence, find the definition it relies on before drawing conclusions.
3. Trial tools on your own work. For code, choosing a coding assistant: what to test in your own code is a practical way to start. If you are weighing agents, what is an AI coding agent and what is an agentic workflow define the terms and the controls that matter.
4. Read the data terms before you paste anything sensitive into a tool that now builds interfaces around your inputs.
5. Check dates against the primary source. If you sell into the EU or to US agencies, the dates and definitions above are the ones to verify with counsel.

For more on governance and liability when agents act on your behalf, see our sister publication’s pieces on how enterprises govern AI agents in practice and who is liable when an AI agent fails. For background terms, start with the concept pages on artificial intelligence and large language models.

What happens next?

Three dates and one commitment. The 60-day proposal under Executive Order 14434 is due around November 28, and we will update the policy entry if its status changes. Mistral says Large 4’s open weights arrive at the end of October, which would let outside parties test it themselves for the first time. And the coding category sits at the front of our queue, with GitHub Copilot vs. Cursor and Claude Code vs. OpenAI Codex framed as questions, not verdicts.

The commitment: no winner gets declared until published test conditions and linked evidence support one.

Evidence in this article Editorial interpretation Our reading of what the evidence means. Judgment, clearly separated from the evidence itself.

Frequently asked questions

Sources

1. Executive Order 14434, Inaugurating the Era of Super Intelligence Federal Register, 91 FR 63129, October 2, 2026.
2. Regulation (EU) 2026/1744, Digital Omnibus on AI EUR-Lex.
3. OpenAI opens limited access to GPT-6 Astra Forklog, secondary report of OpenAI’s announcement.
4. OpenAI launches GPT-6 Astra DataNorth, secondary report.
5. OpenAI brings GPT-6 and Intelligent UI to ChatGPT Let’s Data Science, secondary report.
6. AI News Today, October 8 AI Weekly, secondary report on Claude Haiku 5.5 and Mistral Large 4.
7. AI Model Releases Timeline PromptZone, API pricing and dates.

Related

• Reading benchmark claims critically
• Superintelligence and today's tools: keeping the terms apart