Introduction

OpenAI made computer use cheaper, Google released its strongest model to a subset of vetted cyber defenders and nobody else, and Mistral put a trillion parameters on Hugging Face. Three labs, three ideas of who should hold the keys: everyone, almost no one, and anyone with eight H200s.

The question did not stop at model access. The UK put money behind both AI-powered cyber defence and the monitoring of rogue agents, and GitHub gave automated services a way to use Copilot without borrowing somebody’s credentials. This week’s common thread is control: who gets access, what an agent may do, and who can switch it off when things go wrong.

I will cover them in the order they landed.

In this issue:

1. GPT-6.1 Terra: Computer Use at $1 per Million Input Tokens
2. Gemini 4 Carbon: A Frontier Model, First for Cyber Defenders
3. Mistral Large 4: A 1T-Parameter Model with Open Weights
4. UK Budget: £600m+ AI Package, Including AI Security and Agentic Cyber Defence
5. GitHub Copilot SDK Gets Server-to-Server Authentication

Diagram 1. Three ways to reach a powerful model: rent it, qualify for it, or host it. Each route trades availability against control and the operational work you take on.

---

Frontier Models

1. GPT-6.1 Terra: Computer Use at $1 per Million Input Tokens

Introducing GPT-6.1 Terra — OpenAI, 26 October 2026

GPT-6.1 Terra: delivering fast and capable computer use at a lower price — OpenAI Developer Community

GPT-6.1 Terra model page — OpenAI developers

OpenAI releases GPT-6.1 Terra, cutting price 50% and benchmark error rates — Let's Data Science, 26 October 2026

OpenAI bets on AI economics with lower-cost GPT-6.1 Terra — InfoWorld

GPT-6.1 Terra: OpenAI's cheaper mid-tier model and all six of its benchmark charts — On The Wire, 27 October 2026

OpenAI released GPT-6.1 Terra on 26 October. It is a computer-use model. According to OpenAI, GPT-6.1 Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol in its computer-use tests on Agents’ Last Exam, a benchmark of professional tasks. At maximum effort, OpenAI also says Terra is over three times as token-efficient as Claude Opus 5.5 at one-fourth the price per token, which it puts at roughly 9x lower cost per task. That is a vendor comparison with a competitor’s model, and I found no independent check of it. OpenAI’s announcement, which is also posted in the OpenAI Developer Community, gives the availability: Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and the API as gpt-6.1-terra. According to OpenAI’s model page, the API price is $1 per million input tokens and $5 per million output tokens, and prompts above 272K input tokens cost more. OpenAI’s announcement states that the “50% lower” price is against GPT-5.6 Terra. InfoWorld quotes an analyst, Greyhound Research, who notes the rates are also half of GPT-6.1 Sol’s, which is why some headlines use Sol as the baseline.

The accuracy figures come from an OpenAI internal benchmark built from error-inducing prompts, as reported by Let’s Data Science and InfoWorld. Both say the error rate at low reasoning effort falls from 53.6% to 16.7%, and InfoWorld names GPT-5.6 Terra as the comparison. Those two reports do not give the Astra and Sol numbers, but On The Wire’s summary of OpenAI’s benchmark charts does. On its chart of difficult factual questions, GPT-6.1 Terra has a 4.3% error rate at maximum reasoning effort, almost identical to GPT-6 Astra (4.2%) and GPT-6.1 Sol (4.4%). The difference is just 0.1 percentage points in either direction. These are vendor-reported results from a deliberately challenging test set, not independent measurements of everyday hallucination rates. The context window is 1,050,000 tokens, with up to 128,000 output tokens, according to OpenAI’s model page.

These figures measure different things. A lower error rate on a prompt-based evaluation does not mean a computer-use agent will complete the same share of real tasks. A workflow of screenshots, clicks, forms and confirmation steps can fail even when the model answers correctly, for example by clicking the wrong button or losing its place on the screen. On OSWorld 2.1, a desktop-task benchmark, On The Wire’s reading of OpenAI’s charts puts Terra at 74.2% at maximum effort, against 80.9% for Astra and 78.0% for Sol, so near-identical error rates did not mean equal desktop success. Token price is also only part of the bill, because one task may need several screenshots, model calls and verification steps. Wall-clock time matters too: a screen-driving loop spends much of its time capturing the screen, acting and waiting for the interface to settle, so faster inference alone does not make a whole task proportionally faster. The useful metric is cost per verified successful task, including retries and human corrections.

Why this matters

A computer-use agent is a temp who can operate the till, the stockroom terminal and the email client, which is useful until one of them goes wrong and you are the one signing off the accounts. Cheap screen-driving agents move the economics of tedious back-office automation, such as copying values between a legacy portal and a spreadsheet, from “not worth the engineering” to “worth a pilot”. My advice: pilot it on a task where a wrong click is cheap and reversible, log every action, and measure error rate on your own screens. An error-inducing prompt set is not your invoicing system.

---

Security

2. Gemini 4 Carbon: A Frontier Model, First for Cyber Defenders

Gemini 4 Carbon: our next frontier model — Google, 28 October 2026

Gemini 4 Carbon puts Google back on top, but only cyber defenders can use it for now — WinBuzzer, 28 October 2026

Gemini 4 Carbon vs Argon, GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 — Artificial Analysis

Google announced Gemini 4 Carbon on 28 October, available first to a subset of cyber defenders through its Fairwind Program, with access widening gradually to additional trusted partners. WinBuzzer reports more than 650 partners, who must pass background checks, use phishing-resistant multi-factor authentication, and may not share access. Google gives no date for general availability, and WinBuzzer says Google’s reason is that the ability to find and fix vulnerabilities “could also help attackers”. Its previous model, Argon, was announced for Fairwind partners only on 30 September and opened broader access on 13 October, per WinBuzzer, but Google has not committed to the same timetable for Carbon. Introductory pricing is $2 per million input tokens and $10 per million output tokens, doubling later.

Google-reported cyber results (Google says its external benchmarks were run without guardrails) are 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71% on CWE-bench (patching). Independent numbers, as reported by WinBuzzer, are more mixed. Artificial Analysis puts Carbon at 59 on its Intelligence Index (v5.0), ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57), but its Coding Agent Index has Carbon at 59 behind Claude Code with Opus 5.5 at 62, and WinBuzzer also reports an 11% hallucination rate for Carbon against Argon’s 4%. I could not confirm those two figures on Artificial Analysis’s own pages, and its September launch article gave Argon 15% on an earlier index, so treat them with care. Google says its own internal testing found Carbon cuts unsupported claims by approximately 2x over Argon, so the two measures differ. On the four-attempt CWE-bench leaderboard, Grok 4.7 (81%) and Opus 5.5 (79%) lead, with no Carbon result listed.

The rankings need a little care. Google’s page gives no evaluation setup or attempt count for its 71% CWE-bench result. WinBuzzer describes it as Google’s self-reported, programmatic, single-attempt score on 120 held-out tasks, while that leaderboard allows four attempts per task. Those are different testing conditions, so the scores should not be compared as if they came from one experiment. The more useful question is how many attempts, tools and resources a model needs to produce a working fix.

Google also reports a 0.0% attack success rate for Carbon on Gray Swan’s indirect prompt-injection benchmark, against 0.7% for Argon. That is a vendor-reported result on one test, not a claim of immunity, but it fits this week’s theme of controlling what agents will do.

Why this matters

Think of a locksmith course open only to people with references: sensible, until you remember that a lock-picking guide is already on the shelf next door. Staged access for a model that scores 54.8% at exploitation is a defensible call, and I would rather see it than a surprise public launch. The pattern is the news, though: the same capability ships to defenders first, and the rest of us wait, with no date from Google. If you build security tooling, ask now whether your product can use a gated model, and plan a fallback for the weeks when you cannot. Treat “state of the art” with care, since the lead here depends on how success is measured.

---

Open Models

3. Mistral Large 4: A 1T-Parameter Model with Open Weights

Mistral Large 4 open weights — Mistral AI, 29 October 2026

Mistral Large 4 on Hugging Face — Hugging Face, 29 October 2026

Mistral released the weights of Mistral Large 4 on 29 October, after an API preview earlier in the month. It is a sparse mixture-of-experts model with about 1T total parameters and a 1M-token context window. The Hugging Face article lists 49B active parameters, while the repository name says A52B, and I found no explanation for the gap, so treat both as approximate. In a mixture-of-experts model, only a fraction of the parameters is used for each token, which reduces compute per token, but all expert weights still have to be stored and accessible. That is why memory, not arithmetic, is the constraint here. The licence is the Mistral Ecosystem Open-Weight License, a modified MIT: use, modification, fine-tuning, and commercialisation are allowed, but businesses running a model-as-a-service offering need a separate commercial licence.

Diagram 2. A mixture-of-experts model holds many specialist networks but activates only a few per token. Fewer active parameters reduce computation, but storing the whole model still takes substantial memory.

Official FP8 and NVFP4 checkpoints exist, with GGUF quantisations. A 2-bit quant runs on a single Apple Silicon Ultra with 512 GB of unified memory (Hugging Face measured about 15 generated tokens per second on an M3 Ultra with Q2_K), the vLLM NVFP4 checkpoint needs NVIDIA Blackwell GPUs, and Hugging Face’s Inference Endpoints recipe uses 8x H200. Mistral reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average, and 41% on Terminal-Bench 4.0, up 13 points from the preview. The API costs $0.70 per million input and $2.10 per million output tokens.

Also worth watching: Moonshot’s roughly 3T-parameter Kimi K3.1 Pro arrived on 23–24 October (Unite.AI’s article date and text disagree), with a company-reported 65.7% on Terminal-Bench 4.0 (Unite.AI). Full weights are promised by 4 November, which is a commitment rather than a release. Terminal-Bench figures from different vendors come from different harnesses, so do not read 41% against 65.7% as a straight contest.

Why this matters

“Open weights” at this size is a bit like being given the keys to a cargo ship: yours to use, though you will need a harbour. For a well-funded workstation, quantised inference is now possible even at the trillion-parameter scale, though at around 15 tokens per second it is slow next to a hosted API. Fine-tuning is another matter: Hugging Face’s published training example uses a node of eight B300 GPUs, and running a 2-bit quant locally does not mean the same machine can train the model. Its serving example also sets --max-model-len 131072, which is a configuration choice rather than a stated limit, so check what fits before counting on the full 1M. The licence clause is the catch for anyone planning to resell inference. If you sell hosted access, read it before you build a business on it, and budget for the conversation with Mistral.

---

Governance

4. UK Budget: £600m+ AI Package, Including AI Security and Agentic Cyber Defence

Budget 2026 document — GOV.UK, 28 October 2026

UK government announcement on AI funding — GOV.UK, 24 October 2026

UK government sets out AI security funding — The Register, 26 October 2026

Chancellor announces over £600m package to back British AI, security and skills — London Tech News, 28 October 2026

Budget 2026 funds £40m unit to monitor rogue AI agents — Resultsense, 29 October 2026

The government set out the package in an announcement on 24 October, ahead of the Autumn Budget on 28 October. The Register, writing on 26 October, lists £120m for the AI Security Institute, a £60m National Cyber Shield using agentic AI (including a £25m challenge fund), a £70m National Exposure and Real-Time Vulnerability Centre, and £165m for biosecurity, including a £100m AI Biosecurity Lab. London Tech News and the 24 October announcement both give £60m for the Sovereign AI Fund, while the Budget text mentions an additional £140m for a Sovereign AI Unit. I could not reconcile the two, so I would not rely on either figure. Outlets quote totals between £600m and £700m, so treat the headline figure as approximate.

The Budget also confirms £40m for a new National Agentic Security Unit, which will investigate and respond to rogue agentic incidents. Resultsense covered it separately, and the other two articles do not mention it. This is a different job from the Cyber Shield: one programme puts agents to work defending infrastructure, the other investigates what happens when agents themselves become the problem. The Budget gives the funding and purpose, but I found no start date, operator or autonomy limits, and The Register notes it is unclear how much freedom the planned security agents will have to act without human intervention, or when they will be deployed.

Why this matters

Money for agent monitoring signals that agents acting outside their remit have moved from conference-talk worry to a Treasury line item, even if the line is thin on detail. Developers should care about the cyber side. The government wants agents hunting for flaws in critical infrastructure, which is the kind of capability a gated model such as Gemini 4 Carbon is built to offer, although none of the sources says which model the programme will use. I would watch for what the unit measures, since “agent acted outside controls” needs a definition that your logging could also meet.

---

Developer Tooling

5. GitHub Copilot SDK Gets Server-to-Server Authentication

GitHub Copilot supports server-to-server authentication — GitHub Changelog, 27 October 2026

On 27 October, GitHub added server-to-server authentication for services built with the Copilot SDK, in public preview. A service can now make requests for an organisation without a signed-in user’s credentials, through a GitHub App installed on the organisation. The app needs the read-only “Copilot Requests” organisation permission, and a short-lived installation access token is passed to the SDK. GitHub’s documentation says installation tokens expire after 1 hour, so a long-running worker needs to request a fresh one rather than cache it. Usage is billed to the organisation that owns the installation, and its model policies, budgets and spending controls still apply.

Owners can identify app-attributed usage by the app_client_id field in the AI usage CSV, and stop an app by suspending or uninstalling it. The SDK itself has been generally available since 2 June 2026, so this is an addition to an existing product. The changelog does not say whether personal-account apps are supported.

The permission that lets an app call Copilot is not the same as access to repository contents. An app may be allowed to use the SDK and still need separate repository permissions for its actual work, which is my reading rather than something the changelog spells out. For example, a scheduled CI job could use a GitHub App to authenticate with the Copilot SDK, request a code review, and report its findings without borrowing a developer’s credentials. Server-to-server authentication solves the identity problem. It does not replace least-privilege permissions, sandboxing or human approval for sensitive actions.

Diagram 3. A GitHub App gives a scheduled agent its own identity and a short-lived token, but repository access, sandboxing and approval still need their own controls.

Why this matters

This is unglamorous, and it is what stops an internal agent running on someone’s personal token until that person leaves the company. Bot identities with narrow permissions, short-lived tokens, billing you can attribute, and a kill switch are the same pattern the other four stories point towards. If you run Copilot-based automation in CI or an internal tool, move it off a human’s credentials as soon as the preview allows. The app_client_id field also gives finance a way to see which agent spent what, which is the part nobody enjoys finding out at month-end.

---

Closing Thoughts

The three model announcements differ most in who can get at them:

Model	Access	Main trade-off
GPT-6.1 Terra	Hosted API, ChatGPT Work and Codex	Cheap to run, but provider-managed, with no independent testing reported yet
Gemini 4 Carbon	Initially a subset of cyber defenders through Fairwind	Frontier cyber capability, but restricted access and no public date
Mistral Large 4	Downloadable weights under a modified MIT licence	Control over your infrastructure, but heavy hardware and a licence clause for hosted services

Terra makes screen-driving agents cheap, Carbon shows a strong model can be released to defenders before everyone else, Mistral Large 4 puts a trillion parameters in public hands, the UK budget funds agentic defence and agent monitoring, and GitHub’s token change gives an agent its own identity. This week, AI got cheaper, larger and more capable. The most important question may no longer be what an agent can do. It is who gives it the keys, how closely they watch it, and whether they can take those keys back. Let me know what you think.

---

References

1. Introducing GPT-6.1 Terra — OpenAI
2. GPT-6.1 Terra: delivering fast and capable computer use at a lower price — OpenAI Developer Community
3. GPT-6.1 Terra model page — OpenAI developers
4. OpenAI releases GPT-6.1 Terra, cutting price 50% and benchmark error rates — Let’s Data Science
5. OpenAI bets on AI economics with lower-cost GPT-6.1 Terra — InfoWorld
6. GPT-6.1 Terra: OpenAI’s cheaper mid-tier model and all six of its benchmark charts — On The Wire
7. Gemini 4 Carbon: our next frontier model — Google
8. Gemini 4 Carbon puts Google back on top, but only cyber defenders can use it for now — WinBuzzer
9. Gemini 4 Carbon vs Argon, GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 — Artificial Analysis
10. Mistral Large 4 open weights — Mistral AI
11. Mistral Large 4 on Hugging Face — Hugging Face
12. Moonshot AI releases Kimi K3.1 Pro — Unite.AI
13. UK government sets out AI security funding — The Register
14. Chancellor announces over £600m package to back British AI, security and skills — London Tech News
15. Budget 2026 funds £40m unit to monitor rogue AI agents — Resultsense
16. GitHub Copilot supports server-to-server authentication — GitHub Changelog
17. Copilot SDK is now generally available — GitHub Changelog
18. Budget 2026 document — GOV.UK
19. UK government announcement on AI funding — GOV.UK