A UK regulator now treats chatbots as search
The CMA wants Google to put AI assistants on the choice screens Android and Chrome users see when they set up a phone or open a browser. In the same week both of the big labs cut their prices on the same day.
- 01
The CMA wants AI assistants on Google's choice screens
On 23 September the CMA published strengthened proposals under its Google search case, and the new part is that AI assistants meeting its technical and security criteria would be allowed to appear on the choice screens users see when they first set up an Android phone or open Chrome, with an annual prompt to pick a default. Chief executive Sarah Cardell said the regulator is "responding quickly to the rise of AI assistants, which people are increasingly using as their search service". Comments close on 9 October and a final decision is expected by the end of the year. Read plainly, a British regulator has now put chatbots and search engines in the same box.
Competition and Markets Authority - 02
The ICO becomes the Information Commission on 30 September
A written statement to Parliament on 14 September confirmed the date the data protection regulator changes shape, moving from a corporation sole, where authority sat with one named Commissioner, to a body corporate where executive and non-executive members decide together. Seven non-executive members take up their posts on 30 September, with an interim chief executive in place while recruitment for a chair continues. Nothing changes in what your business has to do, and the regulatory functions carry over untouched. It matters because this is the body now drafting the UK's statutory code on AI and automated decisions, and it is rebuilding its own decision-making at the same time.
Information Commissioner's Office - 03
Anthropic released Opus 5.5 at 20 per cent off the sticker price
Claude Opus 5.5 arrived on 22 September at $4 per million input tokens and $20 per million output, down from $5 and $25, with cached reads cut from $0.50 to $0.20. Anthropic says that at default settings it works out 40 per cent cheaper than Opus 5 on typical work and produces output more than 30 per cent faster, and it is available on AWS, Google Cloud and Azure from the start. The interesting number is not the benchmark, it is the cache price falling by 60 per cent, because repeated work on the same documents is what most business use actually looks like.
Anthropic - 04
OpenAI halved its API prices the same day
Also on 22 September, OpenAI put out GPT-6 Sol at $2 per million input tokens and $10 output, and GPT-6 Luna at $0.10 and $0.50, roughly half what the models they replace were charging, and it says these are permanent rates rather than a promotion. Sol is aimed at coding and agent work, Luna at high volume clerical jobs like summarising documents and pulling fields out of them. Both are in the API and in ChatGPT Work and Codex for paid plans, with Luna reaching Free and Go users through the desktop app. Two flagship launches in one day, both leading on price, which tells you where the competition has moved.
OpenAI - 05
Anthropic is paying Accenture to audit it from the inside
Announced on 18 September, Anthropic named Accenture its first embedded evaluator, with each side committing at least $1bn over five years to the arrangement. Accenture staff get employee-like access while models are being trained, running adversarial testing and checking safeguards before deployment decisions are final, and the deal is not exclusive. Whatever you make of a consultancy auditing its own commercial partner, the shape of it is worth noting: paid external assurance on AI is becoming a product, and products end up in procurement questionnaires.
Anthropic - 06
California ordered work on a kill switch and onsite auditors
Governor Newsom signed an executive order on 18 September giving the state's Government Operations Agency until 16 November to recommend whether California law should require a shutdown mechanism for frontier models, place independent verification organisations onsite inside the largest developers, and widen the list of safety incidents those firms must report to include loss of control events. It is a study order rather than a rule, so nothing binds anybody yet. Set it next to the Accenture deal and the same idea is arriving from two directions at once, one bought voluntarily and one written into state law.
Office of the Governor of California - 07
The UN's science panel invoked the precautionary principle on AI agents
The Independent International Scientific Panel on AI, forty experts co-chaired by Yoshua Bengio and Maria Ressa, published its first thematic brief on 21 September as leaders gathered in New York, in an advance unedited version. It examines this year's OpenAI and Hugging Face incident as an early example of capable agents persistently pursuing goals that conflict with what their operators intended, and argues that loss of control is exactly the sort of problem the precautionary principle exists for, where harm could be severe and irreversible while the odds stay uncertain. No obligations follow from a brief. It is still the first time the UN's own scientists have said install the safeguards before you understand the risk properly.
United Nations
Of everything this week, the CMA proposal is the one that touches an ordinary British firm, and it has nothing to do with models or safety. It is about where a customer starts. The regulator wants Google to show a choice of search providers when somebody sets up an Android phone or opens Chrome, to ask again once a year, and to let AI assistants onto that screen if they clear its technical and security bar. Cardell's reason for the change was that people are already using assistants as their search service. That sentence is doing more work than the proposal is.
Think about what happens to a firm that gets most of its enquiries through Google. For twenty years the job has been to rank, and the tactics were dull but knowable: a decent site, a Business Profile, reviews, the right words on the right page. If a share of customers pick an assistant at setup and never open a search engine again, the question stops being where you rank and becomes whether the assistant knows you exist and describes you correctly. Those are not the same skill. From direct experience in construction, the part of the trade that gets overlooked is that reputation travels through specific channels, and when the channel changes the reputation does not automatically come with it. Nothing is decided yet, comments close on 9 October and the CMA will not rule until the end of the year, so there is no job to do this week. The thing worth doing is checking what a couple of assistants say about your company today, because you can still influence that answer while nobody is competing to.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.