What to watch next in AI tools, and how to judge predictions

admin
What to watch next in AI tools, and how to judge predictions cover graphic

This article was planned as a look back at our own AI tool predictions and how they held up. We have to be straightforward: we have not published any predictions, so there is nothing to score. Writing a fake scorecard would be dishonest, so we have written something more useful and more modest instead.

Below are open questions to watch in AI tools over the coming months, each tied to something a dated source reported as of 11 November 2026, followed by a method for judging any prediction you read, from us or anyone else. These are questions to watch and not forecasts. We are not saying how any of them will turn out.

Key takeaways

  • We have no published predictions, so this article does not grade any. It lists questions to watch instead.
  • Each question below is anchored in a dated report, such as an introductory price that ends on 30 November or a court filing due after an August ruling.
  • A good prediction names a measurable outcome, a deadline and the evidence that would prove it wrong.
  • Confident forecasts about fast-moving technology deserve more scepticism than specific, checkable claims about the near term.
  • Keep your own short log of what you expect and what happens. It is the most honest way to learn how reliable your judgement is.

Summary

Predictions about AI tools are everywhere, and they are rarely checked afterwards. A confident claim that a certain tool will win, a price will fall or a rule will change is easy to forget once the moment passes. We think readers are better served by two things: a short list of concrete things that can be checked soon, and a way of reading forecasts critically.

The questions below come from our own recent news roundups. Each links to the full article, where the dated sources are listed. Where we mention a date, it is the date a source gave. We have not tried to say what will happen.

Why there is no scorecard

A “how our predictions held up” article needs three things: predictions published in advance, a clear outcome for each and an honest tally. We have the first missing. Any list we wrote now and then marked as correct would be hindsight. So we have not done that, and we want to say so openly. If we publish forecasts in future, we will date them, state how they could be proved wrong and revisit them.

Questions to watch: overview

Area Open question Anchored in Earliest checkpoint
Pricing What happens to Gemini 4 Argon API pricing when the introductory rate ends? Report of 13 October 2026 30 November 2026
Pricing Do credit and compute-based allowances spread across more consumer plans? GitHub, Google and Anthropic changes in 2026 Vendor pricing pages, quarterly
Models Did Moonshot publish the Kimi K3.1 Pro weights it promised? Tracker updated 28 October 2026 4 November 2026, now passed and unconfirmed by us
Security Does the next OWASP quarterly roundup show more observed exploitation? Q2 roundup of 8 July 2026 The next roundup
Courts What do the amended complaints in the OpenAI litigation look like? Tracker post of 7 August 2026 Fifteen days after the August rulings, check court records
Work Does Gallup’s organisational adoption figure keep rising? 47 percent in May 2026 The next Gallup quarterly release

Pricing and plans

Introductory prices ending

A report of 13 October said Gemini 4 Argon had introductory API pricing through 30 November, with standard pricing to be higher and no figures given. The question is simply what the new price is and how it compares with rivals. You can check Google’s pricing page on 1 December. For background on why vendors’ prices move, see our roundup of AI pricing changes in 2026.

Credits and compute allowances

Google moved to a compute-based model in May, GitHub moved Copilot to credits on 1 June and Anthropic separated third-party tools from its flat plans in April. Whether other vendors follow is an open question you can check by reading plan pages. A fair test is whether a plan states its limits in a form you can plan around.

Models and releases

The late-2026 release run included three Claude 5.5 models, Gemini 4 Argon, Meta’s Muse Spark Open and GPT-6.1 Terra. Questions worth watching include whether independent evaluations of these models appear and whether they agree with the vendors’ own claims, and whether preview-only products such as Genie 4 reach wider availability. Our model release roundup has the dated details and notes where trackers disagreed. One concrete check: a tracker said Moonshot promised full Kimi K3.1 Pro weights by 4 November, and we did not confirm whether that happened.

Security

The OWASP roundup for the second quarter of 2026 listed six cases, only two with observed malicious activity. A fair question for the next roundup is whether the balance shifts. Another is whether the gap between a vendor learning of a flaw and fixing it narrows, since one incident monitor entry cited a report of more than eight months in a Copilot case, which we could not verify. See AI security incidents in 2026 for the cases.

Several matters are open: appeals in Thomson Reuters v. Ross, review sought in the Kadrey case, a UK appeal in Getty v. Stability AI and the next steps in the consolidated OpenAI litigation after the August rulings. The most useful way to follow them is to read court records and reputable trackers, not summaries of summaries. Our article on where AI copyright cases stand lists what we found and what we could not confirm. This is general information, not legal advice.

Work and support

Gallup’s May figures, 52 percent individual use and 47 percent organisational integration, will be updated in future quarters. Gartner’s 2026 findings on customer support, including the 24 percent of leaders who reported positive financial returns, invite a simple follow-up: does that number change in later surveys? Whether companies that planned AI-driven layoffs went ahead is something to look for in later reports. These are questions, not predictions about what the next survey will say.

How to judge a prediction

You can apply the same checks to any forecast about AI, whether it comes from a vendor, an analyst, a newsletter or us.

Check What a good prediction looks like Warning sign
Specific Names a measurable outcome, such as a price, a date or a count Vague words such as “transform” or “disrupt”
Time-bound Gives a deadline “Eventually” or “soon”
Falsifiable States what would show it wrong No outcome could contradict it
Sourced Rests on dated evidence you can open Relies on unnamed insiders or “people say”
Track record The forecaster has published earlier calls and revisited them Only successes are mentioned
Incentives You know who benefits if you believe it The forecaster sells the product being predicted
Honest uncertainty Admits what is not known Total confidence on a fast-moving topic

Vendor claims are not predictions, but treat them alike

Launch announcements often include promises: “coming in the coming weeks” or “available later this year.” In the reports we read, vendors also made performance claims that we could not verify, such as benchmark scores. Treat both as claims to check, using a trial run for products and careful review for outputs.

Beware hindsight

After events, almost everyone remembers having expected them. The only protection is writing the expectation down beforehand, with a date.

Keep your own prediction log

A simple log is the best way to calibrate your judgement and to improve how you buy tools.

Field Example entry
Date written 11 November
Claim “The vendor’s introductory API price will rise after 30 November.”
How I will check Read the vendor’s pricing page on 1 December
Confidence Low, medium or high, written before checking
Outcome Filled in on the checkpoint date
What I learned One sentence, including whether the confidence was justified

What it means for tool buyers

  • Do not buy on a forecast. Decide from what a tool does for your task today, using a small trial.
  • Prefer reversible choices. Monthly plans and exportable data cost less if a prediction fails.
  • Re-check on a schedule. A quarterly look at your tools’ pricing, limits and terms catches most changes.
  • Be wary of urgency. Predictions that you must act now or fall behind deserve extra checking.

Our article on AI tools that shut down or merged in 2026 reinforces this: the safest plan assumes change.

Frequently asked questions

Why does this article not score your predictions?

Because we have not published any. Scoring predictions made after the fact would be misleading.

Are the open questions your forecasts?

No. They are things a dated source suggests are worth checking. We have not said how they will turn out.

How far ahead can anyone predict AI tools?

We do not know, and we would be sceptical of anyone who claims to. Near-term, checkable claims such as scheduled price changes are more reliable than broad claims about what the technology will do.

Where can I verify the facts you cite?

Each linked article lists its sources with dates, and the Sources section below lists the pages we opened for this one. Confirm anything important with the vendor or the original document.

Will you do a real review of predictions later?

If we publish dated forecasts, we intend to revisit them. We have not committed to any yet.

Conclusion

We cannot tell you how our predictions held up because we made none. What we can offer is a short list of checkable questions, a method for judging forecasts and a habit worth adopting: write down what you expect, check it and learn from the result. If you are choosing tools in the meantime, start with how to choose an AI tool, and revisit the questions above on the dates given.

Sources

Keep reading