Skip to content
Main Site News Console

New Claude Model Revealed! Thinks Outside the Box to Fix Bugs at the Root

· 量子位
国内AI

Scan QR code to follow QbitAI

Image 1

< img id=“wx_img” src=“https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png” width=“400” height=“400”>

Image 2Meng Chen 2026-10-20 09:09:43 Source: QbitAI

OpenAI Only Delivers a 6.1 Terra…

Meng Chen reporting from Office
QbitAI | WeChat Official Account: QbitAI

The funniest joke circulating in the AI community lately:

Over the next month, the “Big Three” are all releasing new models.

The new Claude is being hailed as a “major leap forward”—when fixing a bug, it addresses the root cause, identifying broader underlying issues and tackling them together.

Gemini is releasing Carbon, an upgrade to Argon, with internal employee feedback claiming it “can take on Opus 5.5.”

Meanwhile, ChatGPT is serving up just a medium-tier 6.1 Terra…

Image 3

Seriously? And here everyone thought the current 6.1 Sol was just Terra under a different name.

Image 4

Claude’s New Model Can Think Outside the Box

According to reports from The Information:

If you thought the pace of model releases over the past month was rapid, don’t expect it to slow down anytime soon.

Beta testers say the new Claude performs significantly better in reasoning and intuition, excelling particularly at understanding a user’s underlying intent and objectives.

For instance, a user might ask the model to patch a single bug, but the model could identify the root class of problems causing the bug and ask if it should fix the entire issue category, according to someone who tested the model.

Image 5

The new Claude also excels at thinking outside the box, rather than simply regurgitating generic, conventional advice or boilerplate solutions.

If a user asks for stock investment advice, instead of offering cliché tips like checking for high P/E ratios, it might propose its own creative algorithm to forecast future stock performance.

It remains unclear whether the new model is Fable or Opus—either way, it is substantially more capable.

Gemini 4 Carbon May Have Used RSI

Meanwhile, Google is also developing a new model codenamed “Gemini 4 Carbon.”

Early versions of Argon lagged behind Claude Opus 5 on certain coding tasks.

Carbon outperforms Argon in coding tasks and continues to excel at benchmarks requiring visual understanding, such as analyzing video content or testing front-end web functionality (benefiting from training on YouTube data).

Carbon was made available to employees for several days on Jetski, Google’s internal coding platform. Staff who tried it noted that it “feels on par with Opus 5.5,” though more testing is needed.

Google tested a lineup of Gemini 4 models, codenamed Argon, Barium, and Carbon.

Among them, the model internally codenamed “Barium-B” was ultimately selected and launched as the Argon we see today.

After Carbon’s internal test performance was leaked, Google employee Vedant Misra posted: “Have you heard of recursive self-improvement (RSI)?”

It is unclear whether this was a hint that Carbon leverages RSI techniques.

Image 6

Is OpenAI Running Out of Steam?

GPT-6.1 Terra has already surfaced inside Microsoft Copilot’s codebase.

Image 7

So it turns out that many people’s suspicion—that 6.1 Sol was just a rebranded Terra—was mistaken. As it happens, 6.1 Sol was simply not that much of an upgrade over 5.6 Sol.

Upon seeing the news that 6.1 Terra is coming, many reacted with a sense that OpenAI has thrown in the towel and is just phoning it in.

Image 8

However, some argue that the Terra upgrade is genuinely important for OpenAI.

Because they plan to merge ChatGPT’s Chat/Work and Codex modes, they will need a more cost-effective model to serve free-tier users.

Image 9

At the same time, OpenAI announced a new milestone.

The latest GPT models demonstrate significantly stronger performance in niche legacy programming languages (such as FORTRAN and COBOL).

Many banks and public utility systems still rely on this legacy software, which few human programmers can maintain anymore.

Image 10

COBOL powers over 40% of banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions, while Fortran is widely used in weather forecasting and nuclear reactor safety analysis.

In tests migrating COBOL codebases to other languages, GPT-6.1 Sol achieved an 86.6% accuracy rate at the low reasoning tier, compared to just 5.4% for GPT-5.6 Sol. At higher reasoning settings, accuracy reached 97.9%, virtually acing the benchmark.

Image 11

Finally, on the domestic Chinese model front, DeepSeek has resurfaced first, with upcoming new models from Kimi, GLM, and MiniMax set to roll out in close succession.

In short, this coming November promises to be just as thrilling.

References:
[1]
https://x.com/kimmonismus/status/2112242007062233122?s=20
[2]
https://x.com/maxxrubin_/status/2112256163761016977?s=20
[3]
https://openai.com/index/low-resource-programming-languages/

All rights reserved. Reproduction or use in any form without authorization is strictly prohibited.

Claude

Meng Chen

Scan QR code to share to Moments Image 13: Scan me!

Led by Yao Class Alum: Claude Conquers First Complete Formal Proof of Fermat’s Last Theorem

Saved in the end by Harness

Meng Yao 2026-09-05

Claude

Google Bets Again on OpenAI Rival! $450 Million in New Funding Secured, Total Reaches $1.45 Billion

Led by Spark Capital

Kresi 2023-05-25

Fable 5 Returns to a Flood of Bad Reviews in 24 Hours! Benchmark Drops, Refusal to Answer, and Secretly Insulting Users

Counting the letter ‘r’ is also a safety risk?!

Kresi 2026-07-03

LLMs Join the Reasoning Collapse Debate! From “Illusions of Thought” to “Illusions of Illusions” and “Illusions of Illusions of Illusions”

Wen Le 2025-06-20

Claude-Powered Manus Built in Just 10 Days, Entire Codebase Written by AI! Netizens: Zuck’s $14B Acquisition Looks Like a Sucker Deal

Humans are still needed for planning, design, and iterative AI trials

Meng Chen 2026-01-14

Claude

Stitching DeepSeek with Claude Yields Better Results Than R1 or o1 Alone! 3k Stars on GitHub

The DeepClaude Open-Source Project Goes Viral

Xi Feng 2025-02-14

Breaking! DeepSeek Releases New Model, Super Cost-Performance Obliterates the Competition 2026-10-17

Scan QR code to follow QbitAI Image 24

Tracking new AI trends, reporting breakthroughs in the tech industry

QbitAI All Rights Reserved © Beijing Geek Partner Technology Co., Ltd. Beijing ICP Filing No. 17005886-1