The Bank of England and FCA Say 86% of Finance Firms Use AI, and Only About Half Check It Once It Is Live. OpenAI's New Mid-Price Model Fills Forms From a Screen at Half the Cost.

On Monday 26 October the Bank of England and the Financial Conduct Authority published the results of their fourth survey of AI and machine learning in UK financial services.

196 firms responded, 101 larger and 95 smaller. The headline figures are based on the 105 that completed the full survey.

86% of respondents reported using AI, up from 75% in 2024. Among small firms the figure is 73%, and 58% of those said they predominantly use large language models, the chat-style tools such as ChatGPT, Claude and Gemini. The regulators caution that the 2024 and 2026 samples may not be the same firms.

Operations was the most common area for current or planned use, named by 64%. Generative AI, software that writes text, images or code, made up 56% of all reported AI systems, against 17% for foundation models in 2024.

80% reported having AI agents deployed, software that carries out a task step by step on its own, and a further 14% plan to within 12 months. Almost all are decision support, where a person still makes the call. Only 2% reported agents with broad autonomy. 16% use agents for automated customer communications now, rising to 34% including planned use. No more than 9% expect agents to approve or decline customer outcomes, execute trades or initiate payments.

Third-party systems, tools bought in rather than built, made up 39% of all AI systems and 53% of the high-materiality ones, the systems a firm says matter most. For third-party systems, 81% of firms rely on vendor assurance and contracts, and only 24% rely on deep knowledge of how the system works.

On governance, 94% have firmwide AI policies and more than 90% reported human approval, pre-deployment review and vendor due diligence. But ongoing monitoring, leakage testing and periodic review were each reported by only about half. In the report's words: "Pre-deployment data risk controls were reported more widely than post-deployment monitoring and review controls."

The top risks firms named were data protection and security, third-party and supply chain risk, and generative AI-specific risks such as hallucinations (confident wrong answers) and prompt injection (instructions hidden in a document or email that the AI then follows). 30% of small firms said data protection and IT systems were their main constraints, and 42% of all firms rated UK data protection law a high compliance burden.

On 26 October OpenAI released GPT-6.1 Terra, the middle tier of its range (Luna, Terra, Sol and Astra, from cheapest to most capable). It replaces GPT-5.6 Terra.

OpenAI describes it as "a fast, efficient model that combines detailed visual reasoning and precise computer use". Computer use means the model looks at a screen and acts on it: clicking, typing and moving between apps. OpenAI's own examples are drafting proposals, interpreting technical drawings, filling out forms and editing videos.

Through the API, the route developers use to build it into tools, it costs $1 per million input tokens and $5 per million output tokens, half the price of GPT-5.6 Terra. A token is roughly three-quarters of a word. OpenAI says it completes tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol, and performs on par with Astra at one-tenth the price per token.

On OpenAI's own internal test of error-inducing prompts, the error rate fell from 53.6% with GPT-5.6 Terra to 16.7% at low reasoning effort, and to 4.3% at maximum effort.

The independent measurement is more modest. Artificial Analysis found a 6-point gain over GPT-5.6 Terra on its Intelligence Index at maximum effort, 4 points below GPT-6.1 Sol, at $0.40 per task against $1.40. On its hallucination test the rate of made-up answers fell from 88% to 38%. Better, not solved.

It is available now to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-terra. OpenAI still recommends Sol or Astra for the most complex coding.

On 26 October Google changed a default in the Gemini API, the route developers use to build Gemini into apps and chatbots. Paid projects created on or after 26 October now store every request and response by default.

Google's changelog wording: new projects "on the paid tier now store requests by default (store=true), matching the Interactions API. Existing projects keep their current setting."

A stored log holds the full prompt, the response and the previous turns of the conversation. Google's policy page says logs are kept for a default maximum of 55 days, configurable to 7, 14, 28 or 55, are private to the developer's Google Cloud project, and "are not used for product improvement or development" unless the developer chooses to share a dataset with Google, at which point human reviewers may read it.

To switch it off, the developer sets store=false on each request, or turns off Generate Content API logging in the Settings panel of the Logs and Datasets page in Google AI Studio. Logging only applies to billing-enabled projects.

On 26 October Microsoft said its Frontier programme, the opt-in scheme that gives staff preview features before general release, will from November also cover eligible public previews for new Outlook for Windows, OneDrive, SharePoint and Microsoft Teams. Microsoft's pitch: one programme, one admin setting, one place to try what is coming.

Microsoft's admin guide, updated the same day, sets the default to "No access". An admin can choose "All users" or "Specific users and groups" in the Microsoft 365 admin centre under Copilot, then Settings, then search for Frontier. A Copilot licence is no longer required to enrol.

The same guide says Frontier experiences "may not be covered by standard support commitments or service level agreements", that HIPAA Business Associate Agreement coverage is not included, and that "Frontier features do not comply with ISO 27001, ISO 27002, or ISO 27018 requirements."

The timing matters because Microsoft is also retiring Targeted release, its older early-access option. From November 2026 admins can no longer add or remove users from it, and in January 2027 it ends. Users on it default back to Standard or Deferred release.