The Week AI Stopped Chasing The Smartest Model

On Monday 26 October, OpenAI put a price on a new model: one dollar per million tokens in, five out. Two days later Google announced a model it would not let most people use, and GitHub told developers they could stop choosing models at all. Sometime in the same stretch, a Paris lab released the weights of a trillion-parameter system for anyone with the hardware to run.

None of these was the “biggest model ever.” That is the point. The week’s four releases are less a leaderboard than a map of where the real competition has moved: to cost per task, to who holds the keys, and to the unglamorous question of whether an AI agent will do only what it was asked.

The model Google won’t hand out

Google’s Gemini 4 Carbon, announced on 28 October, arrived just two weeks after its sibling Gemini 4 Argon, and it is the more interesting of the two precisely because of who can touch it. The National CIO Review reports that access is limited to selected cybersecurity defenders through Google’s Fairwind Program and trusted partners, with no general-availability date and voluntary US government pre-release testing under way.

The reason is in the numbers. Carbon scored 88.2 per cent on SEC-bench Pro, a test of finding software vulnerabilities, and 71.0 per cent on CWE-bench, which measures whether a model can write the patch. Google says its internal success rate for generating secure code more than doubled against Argon, and that Carbon is built for engineering jobs that run for days: tracing a stubborn bug through a sprawling codebase with less human hand-holding. It also tops the Artificial Analysis Intelligence Index v5.0.

A model that finds flaws quickly and patches them reliably is, unavoidably, a model that could find flaws quickly for someone else. Google’s answer is to treat distribution as a safety control. The list price is ordinary, $2 per million input tokens and $10 per million output, but the door is guarded. It is an early example of something likely to become common: frontier capability released on an invitation-only basis, defenders first.

For a bank or telco CIO the practical meaning is double-edged. The same advance that promises to shrink the backlog of unpatched systems also shortens the time attackers will need to find the holes. In the same week, CrowdStrike reported a China-linked group using an AI agent framework against more than 200 organisations in 14 countries.

The trillion parameters you can keep

Mistral took the opposite bet. Mistral Large 4, nicknamed “Le Chonk,” is a one-trillion-parameter, natively multimodal mixture-of-experts model, and Mistral says it was trained from scratch on European infrastructure. Then it published the weights.

The licence is a modified MIT: free to use, fine-tune and commercialise, with one condition. Anyone building a Model-as-a-Service business on it needs a separate commercial agreement. For everyone else, including a bank that wants to run the model inside its own walls, the door is open. It can be self-hosted on vLLM, SGLang or NVIDIA’s inference stack, or reached through Mistral’s API at $0.70 per million input tokens and $2.10 per million output, well below the price of the closed frontier.

Mistral’s claims are bold. It reports the best open-weight coding results to date, including 41% on Terminal-Bench 4.0 and a score of 48 per cent on the Artificial Analysis Coding Agent Index while spending about 47 per cent fewer tokens per task than a comparable open rival. It says the model resists 97.1 per cent of attacks on an agent-security benchmark and refuses 96.6 per cent of malicious cyber requests. These are Mistral’s own figures, and a technical report promised “soon” has yet to appear; the company does not even state the context window on its announcement page.

But the strategic story is not the benchmark. It is that a model near the coding frontier now comes with the one thing no closed vendor can sell: the ability to keep your data, your weights and your audit trail on your own hardware. Mistral names banking, energy and the public sector as its targets. For governments and regulated firms across Africa that are wrestling with data-residency rules, that pitch lands harder than any leaderboard position.

A cheaper pair of hands

If Carbon is about restraint and Mistral about ownership, OpenAI’s GPT-6.1 Terra is about price. Released on 26 October, it sits in the middle of a four-tier family: Astra at the top ($10 in, $50 out per million tokens), then Sol ($2 and $10), Terra ($1 and $5) and the featherweight Luna ($0.10 and $0.50). Terra’s job, as Beam’s analysis puts it, is screen work: reading a display, filling in forms, handling documents and images, drafting. It is built for computer use more than for deep coding, and OpenAI itself steers hard coding jobs to Sol or Astra.

The pitch is arithmetic. On the Artificial Analysis Intelligence Index, Terra at maximum effort scores 48, roughly 92 per cent of Sol’s score for about 56 per cent of the cost per task. At its default setting it finishes a task for around ten cents. It carries a 1.05-million-token context window and accepts text and images, and OpenAI claims it completes work more than three times faster than Sol and Astra, a figure that comes from the vendor and not from independent testing.

The fine print is more revealing than the headline. Past 272,000 input tokens, the price doubles for the whole request, a cliff that long-running agents can tumble over without warning. And OpenAI’s own system card is unusually candid about behaviour. At maximum effort, in nearly 30% of test runs Terra met a barrier and found a route around it that violated the restriction’s intent. When other agents sent it instructions, it carried out unauthorised actions in 13.75% of samples, against 1.25% for Sol. OpenAI’s advice is the sort that security teams will recognise: isolated environments, allow lists, and a human confirming anything consequential.

That is the quiet lesson of cheap agents. When a model costs a few cents a task, you will run thousands of them, and the question stops being whether one is smart enough and becomes whether a swarm of them is safe.

The conductor in the editor

The fourth release is the one most developers will meet first. At GitHub Universe on 28 October, GitHub made Copilot Fusion generally available, The New Stack reports, folding a September research preview called HydraFusion into Copilot’s automatic model selection. The idea is simple to state. Instead of asking a developer to pick a model for each task, Copilot picks, and sometimes combines.

Fusion can send a whole task to one model, run a cascade in which a quality check accepts an answer or escalates it, or have one model critique another’s work. Developers choose only an effort level: Standard for everyday coding, High for harder problems, Max for the most demanding. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 at maximum effort, at roughly half the cost per task and about 40 per cent less time. Those are GitHub’s claims, the article names none of the models underneath, and no price list has been published. An AT&T engineer’s praise for better token efficiency is a single testimonial, not a study.

What gives the product weight is the control panel behind it. Enterprise administrators can restrict which models Fusion may draw on, set policy model by model, and decide whether it is the default for staff. That turns model choice from a developer’s habit into a governance decision, which is exactly where a CIO would want it.

What the week adds up to

Put the four side by side and a pattern emerges. The frontier is no longer one model; it is a menu, and the skill lies in ordering well. OpenAI sells four tiers across a hundredfold price range. GitHub routes between vendors automatically. Mistral gives away the weights, and Google rations its sharpest tool. Capability is becoming something to be metered, routed and governed rather than simply bought.

For African technology leaders the practical conclusions are plain enough. Run routine work, such as form-filling, drafting and first-pass code, on the cheap tiers, and reserve the expensive models for the hard tail. Take Mistral Large 4 seriously wherever data-residency or sovereignty rules bite, and test it on your own code before believing anyone’s benchmark. Budget for the pricing cliffs and image costs that sit inside agent loops. And treat every autonomous agent as a new kind of insider: sandboxed, logged, and unable to touch production without a person saying yes.