Three days in late October, two of the three biggest AI labs put out new frontier models — and they couldn’t be aimed at more different jobs. On October 26, 2026, OpenAI shipped GPT-6.1 Terra, a model built to be fast and cheap. On October 28, Google followed with Gemini 4 Carbon, a model built to be locked down and used almost nowhere yet. Here’s what each one actually does, with the real numbers behind the headlines.
GPT-6.1 Terra: half the price, not the top model
GPT-6.1 Terra rolled out October 26 across ChatGPT Work, Codex, and the API (as gpt-6.1-terra) for Plus, Pro, Business, Enterprise, and Edu plan holders. OpenAI isn’t pitching it as the smartest model it makes — that’s still GPT-6.1 Sol or GPT-6 Astra for the hardest coding work — it’s pitching it as the one you’ll actually want to call thousands of times a day. Terra costs $1 per million input tokens and $5 per million output tokens, half the price of both GPT-5.6 Terra (the model it replaces) and its own sibling, GPT-6.1 Sol.
The company’s own benchmarks back up the “cheap but not dumb” pitch. On OpenAI’s internal error-prompt test, Terra’s error rate at low reasoning effort dropped from 53.6% (GPT-5.6 Terra) to 16.7%. On Terminal-Bench 4.0 — a benchmark that scores how well a model handles real command-line and engineering tasks — Terra solved 50.3% of tasks at max reasoning effort, more than double GPT-5.6 Terra’s 21.5%. OpenAI is positioning Terra specifically around computer use, visual reasoning, coding, and the kind of everyday office tasks that don’t need a frontier-tier model, which is also why the company flagged it in its safety system card with a “Critical” cybersecurity capability rating under its Preparedness Framework — fast, affordable models that are good at code are also good at writing exploits.
Gemini 4 Carbon: Google’s strongest model, and almost nobody can use it yet
Two days later, Google took the opposite approach. Gemini 4 Carbon, announced October 28, is what Google is calling its strongest model yet — built specifically for complex, long-running engineering and security work, including vulnerability discovery, exploitation, and software patching. On Artificial Analysis’s Intelligence Index, a third-party benchmark that aggregates results across multiple reasoning and coding tests, Carbon ranked first with a score of 59. On Google’s own internal benchmarks, it generates secure code at more than double the success rate of Gemini 4 Argon, the model Google shipped just two weeks earlier.
But almost no one gets to use it yet. Carbon’s initial access is limited to a subset of cybersecurity defenders through Google’s Fairwind Program, part of a voluntary pre-release testing process that includes coordination with the U.S. government. Google says it’s still testing safeguards before expanding access to developers, businesses, and consumers, and hasn’t given a date for that broader rollout. When API access does widen, introductory pricing will match Gemini 4 Argon’s: $2 per million input tokens and $10 per million output tokens — four times Terra’s output-token price, for a model most people can’t call yet.
Why the gap matters
Put side by side, the two launches say something about where the two companies think the real competitive pressure is right now. OpenAI is racing to make its mid-tier model cheap enough that developers never think twice about calling it — Terra is a volume play, meant to run constantly inside ChatGPT Work and Codex without anyone worrying about the bill. Google is doing the opposite: treating its most capable model as something to hand to security researchers first, specifically because a model this good at finding and patching vulnerabilities is, by the same logic, a model that’s good at exploiting them. Neither company has said when the next tier up — GPT-6 Astra’s successor, or Gemini 4 Carbon’s public release — is coming. For now, the practical choice for most developers is straightforward: Terra for everyday, high-volume work; Carbon, once it’s actually available outside the Fairwind Program, for the jobs where security is the whole point.
