The government opened the first four competitions under its £100m Sovereign AI R&D Procurement Scheme this week, and that, rather than OpenAI's new flagship, is the piece of UK AI news that changes what a British team does on Monday. The Register reported the launch on 2 September 2026 (opens in a new tab).
A day later, OpenAI shipped GPT-6 Astra at $10 per million input tokens and $50 per million output tokens, twice what GPT-5.6 Sol cost before the promotion it is currently running. Public money arriving, inference getting dearer, in the same seven days.
And the consultation on workplace monitoring technologies closes on 30 September 2026. It has had far less attention than either of the above, and it is the one with a date attached.
- Term: Sovereign AI R&D Procurement Scheme
- Definition: A £100m UK government programme that buys research and development from UK-based AI companies to solve named public sector problems, sitting alongside the separate £500m Sovereign AI venture fund.
What is the UK's £100m Sovereign AI R&D Procurement Scheme?
It's a £100m pot that pays UK-based AI firms to develop solutions to specific public sector challenges, with the first four competitions covering NHS productivity, defence, computing efficiency and the security of AI agents. The scheme was announced by the Chancellor, and the competitions opened in the week to 6 September 2026 — TechMarketView reported the opening on 1 September, The Register on 2 September.
| Challenge | What the competition document asks for | Challenge owner |
|---|---|---|
| AI for a more productive NHS | Systems that unlock NHS productivity at scale, whether automating workflows, coordinating care or supporting decision-making across services, aligned to the NHS 10 Year Health Plan | Department of Health and Social Care |
| Integrate AI at pace across Defence mission environments | Open, secure, modular-by-design AI that can operate on platform, in the field and by dismounted forces in close combat, against what the document calls critical gaps in edge AI deployment | Ministry of Defence |
| Increase compute efficiency | Novel hardware or software for compute efficiency and heterogeneous systems, supporting expansion of the AI Research Resource | Dept for Business, Innovation, Science and Trade, with ARIA's Scaling Inference Lab |
| Enabling safe AI agent adoption | An evidence-based, operational and ideally automated risk management approach for CISOs, covering resilience and cyber security risk for specific agents in specific contexts | National Cyber Security Centre |
The design point is the interesting one. The stated aim is to remove the barriers that keep small British firms out of public contracts, cash reserves and track record, which are exactly the two things a three-year-old company doesn't have and a systems integrator does.
That matters because of who has been winning the work so far. UK public sector bodies awarded 453 AI-related contracts worth £1.4bn in 2026 up to August, part of 2,129 contracts worth £5bn since January 2018, on Tussell's AI Procurement Tracker figures (opens in a new tab) as of August 2026. Against £1.4bn of live spend, £100m of R&D money is a rounding error with a press release attached.
Who should actually bid?
The scheme suits a UK consultancy or product company with something already running in a trust, a force or a council. A deck and a demo will not survive the evaluation: R&D procurement pays for delivery against a defined challenge, and the assessment asks what has shipped rather than what is planned.
The numbers are in the guidance rather than the announcement, which is part of why the coverage has been vague about them. Contracts run from £250,000 to £10 million, and the guidance expects most to land around £1 million to £3 million. The binding eligibility is narrower than the pitch: the lead applicant has to be a UK-registered company and the project has to take place in the UK. There is no minimum turnover, trading history, net assets or cash reserves, and no stated size test, despite the scheme being aimed at startups and smaller firms. Universities cannot lead a bid but can subcontract. Suppliers keep both background and foreground intellectual property, and government takes a licence to use the outputs for public sector purposes.
Why does this week's UK AI news matter more than a model launch?
Because a procurement route is durable and a model release is a Tuesday. My position: the Sovereign AI competitions are the story of the week for UK practitioners, and GPT-6 Astra is a budget event rather than a capability event for most teams. If you buy from or sell to the public sector, the procurement change alters your pipeline for the next two years.
There's a contradiction sitting inside the government's own pitch, and The Register put it in its subheading: Whitehall is funding startups to fix public services while its own internal AI adoption stays patchy. Buying novel systems from small suppliers is harder than buying seats from Microsoft, and the departments doing the buying are the same ones that have struggled to roll out the tools they already licence.
I'd still take the scheme seriously. The £500m venture fund gives equity to a handful of companies; the £100m procurement route gives revenue and a reference customer to more of them, and a reference customer is worth more than a cheque at Series A.
What does GPT-6 Astra cost?
OpenAI's published price for GPT-6 Astra is $10 per million input tokens and $50 per million output tokens, with cached input at $1 and cache writes at $12.50, as of September 2026. What it costs relative to GPT-5.6 Sol depends entirely on which Sol price you compare against, and that is not a pedantic distinction. Sol's $4 and $20 are promotional: OpenAI cut them from $5 and $30 on 21 August 2026, and its own documentation (opens in a new tab) says the promotional pricing runs at least until 21 November 2026. Against the pre-promotional list rate Astra is twice the price on input and about 1.67 times on output. Against the promotional rate it is 2.5 times both. OpenAI has published no rate for after the promotion ends, so the multiple is somewhere between those two and nobody outside OpenAI knows where. The model shipped on 3 September 2026.
- Term: GPT-6 Astra
- Definition: OpenAI's flagship model released on 3 September 2026, aimed at long-horizon agentic work including computer and browser use, with a context window of about 1,050,000 tokens and maximum output of 128,000 tokens.
| Model and tier | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra, up to 272K input tokens | 10.00 | 1.00 | 50.00 |
| GPT-6 Astra, above 272K input tokens | 20.00 | 2.00 | 75.00 |
| GPT-5.6 Sol, promotional to at least 21 Nov 2026 | 4.00 | 0.40 | 20.00 |
| GPT-5.6 Sol, pre-promotional list | 5.00 | not published | 30.00 |
Cache writes bill at 1.25 times the uncached input rate, so a caching strategy that thrashes will cost you more than no caching at all. That is the detail I'd put in front of whoever signs off your cloud bill this month.
On efficiency the answer depends on the task, and quoting one number for both is how this gets reported wrongly. Artificial Analysis (opens in a new tab) measured about 10% fewer output tokens on its general Intelligence Index, and roughly a third of Sol's tokens in the Codex coding harness at maximum effort. Those point in opposite directions. On the Intelligence Index it puts Astra about 75% more expensive per task than Sol, not the 150% a straight price comparison suggests. On the Coding Agent Index it found Astra costing about the same per task while scoring two points higher.
Should UK teams switch to GPT-6 Astra?
Not by default. Keep GPT-5.6 Sol as the workhorse and route to Astra only for long-horizon agentic tasks, browser and computer use, and the software engineering work where a failed run costs more than the tokens. Run a two-week cost comparison on real traffic before you change the default in production.
OpenAI put unusual weight on safety in the announcement, and Al Jazeera reported the release came amid rising scrutiny of frontier risk. That framing tells you nothing about whether the model is safe in your context, so test it against your own workload before you route production traffic to it.
Cheaper, less power-hungry ways to run models is one of the four things the UK is now buying. The week's flagship release went the other way on price.
Circle back to the procurement point for a second. Public sector buyers are about to write agent contracts against models whose flagship tier has just got materially dearer, and the compute efficiency competition exists because that maths does not work at scale. A bid priced on last quarter's rate card is a bid priced on a promotion that expires in November.
What else made UK AI news this week?
Two live deadlines. The UK government's consultation on workplace monitoring technologies closes on 30 September 2026, and the EU AI Act's transparency obligations have been enforceable since 2 August 2026, so any chatbot or synthetic media you ship into the EU is already in scope.
- Term: Workplace monitoring technologies (WMT)
- Definition: Digital tools employers use to collect, track, analyse or make decisions about workers and their activities, including algorithmic management and AI-assisted decision-making, as defined in the UK government's July 2026 consultation.
The consultation opened on 8 July 2026 under the Make Work Pay agenda and asks whether more regulation is needed on transparency, worker voice and accountability. If you've deployed productivity analytics, call scoring or an AI-assisted performance process, you're the subject of it. Responding is free and takes an afternoon, and the alternative is finding out what the rules are after they're written.
On the EU side, the European Commission's transparency rules (opens in a new tab) require interactive AI systems to tell people they aren't talking to a human and require AI-generated or manipulated content to be labelled. The high-risk obligations under Annex III are a different matter: Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved them from 2 August 2026 to 2 December 2027, with product-embedded systems under Annex I going to 2 August 2028.
Domestically, the ICO's published guidance pipeline (opens in a new tab) shows work in drafting with a public consultation due to launch in September 2026 and final guidance in spring 2027, alongside its programme on automated decision-making and profiling. If automated decisions touch your customers or your staff, that consultation is where you get a say.
AI and UK jobs: what the numbers actually say
The most quoted UK figure is still the CIPD's Labour Market Outlook from November 2025, where 17% of more than 2,000 UK employers said they expected AI to shrink their workforce over the following year. That's an expectation, not a headcount, and it's now ten months old.
The metric definition is doing a lot of work in any number published this autumn. A survey reporting employers who have already made AI-related cuts and one reporting employers who plan to make them are measuring two different things, and the plan-to figure is always the larger of the two.
What to do next week
- Read the four Sovereign AI competition documents and decide by Friday whether you have a credible delivery story, not a demo.
- Pull last month's API spend, model the same traffic at GPT-6 Astra rates, and set an explicit routing rule rather than switching the default.
- Respond to the workplace monitoring consultation before 30 September 2026 if you run any form of algorithmic management.
- Check whether every EU-facing chatbot and every piece of synthetic media you publish carries the disclosure the AI Act has required since 2 August 2026.
- Log the Annex III date change to 2 December 2027 in your compliance plan, and diarise the ICO consultation.
One last thing about the £100m. The government's stated aim is to remove the cash-reserve and track-record barriers that usually keep newer companies out of public contracts, and routes designed that way tend to narrow once the first cohort establishes itself. If you're going to bid, this round is the cheap one.