Four AI Models Launched in One Week. Picking the Right One Is Now a Monthly Job.

Four credible frontier models now sit within a few points of each other on independent benchmarks, and each one is best at a different kind of work. If your team picked a default model six months ago and hasn't revisited it, you're overpaying on some tasks and under-serving others — and you probably don't know which is which.

What shipped and why it matters

In the span of seven days: OpenAI launched GPT-6.1 Terra (Oct 26) at $1/$5 per million tokens, half its predecessor's price, built around computer use and desktop automation. Google released Gemini 4 Carbon (Oct 28), which took the #1 spot on Artificial Analysis's Intelligence Index v5.0 — ahead of Claude Opus 5.5 and GPT-6 Astra — with a cybersecurity-first design. Mistral shipped full open weights for Mistral Large 4 (Oct 29), a trillion-parameter model with a 1-million-token context window that teams can download and run on their own servers. Combined with Claude Opus 5.5, which Anthropic released days earlier and which still leads on agentic coding, four frontier models now split the lead depending on what you ask them to do: Carbon on reasoning, Terra on price, Mistral on data control, Opus on coding agents.

Stratechery's Ben Thompson noted the same week that this density of capable options is new — for most of the last two years, one model was clearly ahead for most tasks. That's no longer true.

The skill that's left

"Which AI should my team use" used to be a decision you made once a year. It's now a decision you make per task, and the model that was right for your team in September might already be the expensive option for half of what you're doing in November. The fix isn't switching providers every time a new benchmark drops — it's separating your actual task types first, then checking which model fits each one, instead of running everything through whatever you signed up for last.

Monday morning, that's a 15-minute exercise: list the five AI tasks your team runs most often, and for each one, name what it actually needs — speed and low cost, deep reasoning, a model that can touch sensitive data without leaving your infrastructure, or strong coding/agent behavior. Most teams have never written this down.

Two teams, two different right answers

Priya runs marketing operations at a 40-person SaaS company. Her team's AI spend was almost entirely on one flagship model, used for everything from drafting campaign briefs to reformatting weekly performance reports. Most of that work is routine — it doesn't need frontier reasoning, it needs to be fast and cheap. Routing the reformatting and first-draft work through a lower-cost model like GPT-6.1 Terra freed up budget for the handful of tasks that genuinely benefit from a stronger model, like positioning work ahead of a product launch. She didn't run a bake-off between all four new models to get there — she ran the same five prompts through her current model and the cheaper candidate, side by side, and compared the outputs before switching anything.

Daniel leads operations at a 12-person law firm. His team's draw to Mistral Large 4 wasn't price — it was that client files can't legally leave the firm's own servers, and a self-hosted model solves a problem no cloud model can. But before anyone ran a client document through it, he had to answer a question that had nothing to do with benchmarks: who in the firm approved this as the sanctioned tool, versus someone just trying it on a personal account because it sounded promising. A model that satisfies data-sovereignty rules on paper is still a liability if it gets adopted ad hoc instead of through whoever owns that decision at the firm.

Four models launched in a week, and none of them is "the best one" — each is the right one for a specific job. The teams that will spend less and get more out of this aren't the ones chasing whichever benchmark is on top this month. They're the ones who did the five-minute work of naming what each task actually needs.