Report: Next Claude Model Targets Root Causes as AI Labs Ready New Releases

Users who have tried the model say it is better at judging a person's latent intent and goals, according to the report. In one example, a user asks the model to fix a single bug; the model identifies the class of problem that produced the bug and asks whether it should fix the entire class, one person who used the model said.

The model also performs better at breaking out of conventional thinking instead of repeating common-sense advice or standard solutions, the report said. Asked for stock-investing advice, it would not stop at suggesting a review of the price-to-earnings ratio, according to the account; it would propose its own creative algorithm to predict future stock performance. It is not yet clear whether the new model is Fable or Opus, the report said.

Google, meanwhile, is developing a model codenamed Gemini 4 Carbon, according to the report. An early version, Argon, trailed Claude Opus 5 on some coding tasks. Carbon outperforms Argon on coding tasks and continues to do better on tasks that require visual understanding, such as analyzing video or testing front-end page functions, helped by training on YouTube data. Carbon was opened to employees for several days on Google's internal coding platform Jetski; employees who tried it said it 'felt comparable to Opus 5.5' but that more testing is needed.

Google tested a series of Gemini 4 models under the codenames Argon, Barium and Carbon, according to the report. The internal codename Barium-B was ultimately selected and released as Argon. After Carbon's internal test performance leaked, Google employee Vedant Misra posted, 'Have you heard of recursive self-improvement (RSI)?' It is not certain whether the post was meant to suggest that Carbon uses RSI.

At OpenAI, GPT-6.1 Terra has appeared in Microsoft Copilot code, according to the report. That undercuts speculation that 6.1 Sol was simply Terra released under a different name; instead, 6.1 Sol appears to be only a modest upgrade from 5.6 Sol. Some observers saw the Terra news as a sign that OpenAI is struggling, while others said a cheaper Terra model would matter because OpenAI plans to combine ChatGPT's Chat/Work and Codex modes and needs a lower-cost model for free users.

OpenAI also published an update saying its latest GPT models perform better in less common programming languages such as FORTRAN and COBOL. Many banks and utilities rely on these legacy systems, and few human programmers can maintain them. COBOL supports more than 40% of banking systems, 80% of face-to-face credit-card transactions and 95% of ATM transactions, while Fortran is used in weather forecasting and nuclear-reactor safety analysis. In a test converting COBOL programs to other languages, GPT-6.1 Sol reached 86.6% accuracy at the low setting, compared with 5.4% for GPT-5.6 Sol; at a higher reasoning setting, accuracy reached 97.9%, nearly completing the test.

On the Chinese side, DeepSeek has ended its quiet period, and new models from Kimi, GLM and MiniMax are expected to follow, according to the report. November is likely to remain busy for model releases.

Editor's Summary

The Information reported that a new Claude model can identify root causes and better infer user intent, though its exact name remains unclear. Google's Gemini 4 Carbon and OpenAI's GPT-6.1 Terra are also in testing or appearing in partner code, while OpenAI highlighted large gains on legacy programming languages. The expected releases point to continued competition in frontier models and coding tools.