Anthropic Readies a Model That Reads Between the Lines as Gemini 4 Carbon and OpenAI Terra Stack Up

The next round of the frontier model race is taking shape — mostly behind closed doors. The Information reported Monday that Anthropic is preparing a new model it says "reads users' intentions better" than anything the company currently sells, while Google develops a successor called Gemini 4 Carbon and OpenAI readies a cheaper model in its Terra family for release in the coming weeks.

The Anthropic model's headline skill is intent inference. According to a person who has tested it, cited by The Information, the system does not just fix the software bug you describe — it identifies the broader class of defects behind it and asks whether you want the whole family repaired. Asked a generic question about stock investing, it reportedly skips the textbook price-to-earnings advice and tries to devise its own algorithm for predicting stock performance. The improvements also cover intuition and creative problem-solving. The model's name, series placement, timeline and pricing are all unknown.

Google's track, per the report, is Gemini 4 Carbon — a codename Business Insider previously reported — positioned as an upgrade to Gemini 4 Argon, which launched September 30 and widened access on October 13. Developers familiar with the project told The Information that Google's models remain particularly strong on visual tasks such as video analysis and website testing, which would make Carbon's differentiation as much about perception as reasoning.

OpenAI's entry is a new model in the Terra family expected within weeks, offering stronger performance at a lower price than GPT-6.1 Sol — which reportedly already approaches the company's flagship Astra. That is a direct shot at Anthropic's pricing tiers and part of a broader strategy to cover multiple performance levels rather than funnel everyone into one flagship. Separately, The Information has reported OpenAI is still resolving safety and security questions around its next major flagship.

There is one piece of public forensic evidence for the OpenAI track. An unannounced gpt-6.1-terra model id has already appeared in Microsoft's own eval configurations: a pull request in the dotnet/efcore repository, merged October 16, switched 18 evaluation configs from a gpt-5.6-terra judge to gpt-6.1-terra via GitHub's Copilot SDK. The change was spotted on X by @LuminaBench on Monday. An id in a config file is not a launch — it shows a Microsoft team planned to call the model, nothing more.

The cadence context matters: by The Information's count, the three labs have shipped more than a dozen new models in the past month. Competition is visibly shifting away from leaderboard intelligence scores toward cost, reliability and task-specific performance — the same dynamics on display in this week's MiniMax Flash launch, which undercut frontier pricing by an order of magnitude on agentic benchmarks.

Every detail here traces to unnamed sources, and none of the three companies has confirmed a name, date or price. Treat the report as a map of direction rather than a schedule. But the direction itself is clear: the next wave is being designed to understand what you meant rather than what you typed — and to charge less for it.