English | Chinese Original

English translation of the Chinese original. This version is generated for international readers and may be refined over time.

Date: 2026-02-07

[toc]

Imagine if two of the world's smart brains announced that they had added an additional 200 to the Buff.

That's what happened on February 6, 2026. On that day, the two giants of artificial intelligence, Anthropic and OpenAi, each showed the latest ace: Claude Opus 4. 6 and GPT-5. 3-Codex.

It's not a simple version of an update, it's more like an "spring night" in the AI world, and both protagonists want to prove that their "intellectual evolution path" is the future. And the most interesting thing is, they're different in character and ability, and it's like a role from different science fiction films.

Two roles, two sets of philosophy

Anthropic and OpenAI, like two people of different origins, have different philosophy.

** Claude Opus 4. 6** looks like ** Life Professor of Ivy League**. It was born at the Anthropic Laboratory, known as "safe first", and was well-trained and reliable. It believes that good answers take time to develop, so when you ask complex questions, it builds arguments and looks for evidence like preparing academic papers.

In the face of a complex problem, it pushes the glasses and says, "Let me analyse the problem carefully..."

** GPT-5. 3-Codex** is like ** Silicon Valley Enterprise's technological genius**. It comes from OpenAI, which seeks great efficiency and believes that "action is better than perfection". If you tell it, "I want a website," it doesn't write 20 pages of needs analysis first, but goes straight to the code and asks, "Do you want red or blue on the front page?"

And then in the gap between your thinking, you've knocked out 10 lines of code.

Superpower: When the Master of Memory meets the Democrat

Claude's Life: The real "I can't forget"

Traditional AI has a fatal weakness -- poor memory. Give them a long story, read it back and forget the front.

Take as an example the problem of most developers with the context token, often you're out of vibe coding and you can only redo a new assignment. But Claude Opus 4. 6 solved the problem.

It can now** process over 1 million tokens at once** (equivalent to more than 700, 000 Hs., about 3 Red House Dreams). Worse still, its ability to find something in this "information ocean" has increased dramatically - imagine throwing a needle into the Pacific Ocean and getting it out accurately. In the actual tests,** its accuracy rate in such a "sea needle" mission rose sharply from 18. 5 per cent in the previous generation to 76 per cent**.

What does that mean?

• Counsel can throw it** all the documents of the whole merger** so that it can identify the risk clause.
• Researchers can upload ** dozens of academic papers** to enable them to summarize consensus and controversy.
• Writers can hand over ** all manuscripts and notes** to it, requesting proposals for structural optimization.

# Codex's killer: From "the proponent" to "the enforcer"

If Claude is the Master of Memory, Codex is the revolutionary in the field of implementation.

Before, AI was more like an "adviser" -- advice for you, but you have to do it yourself. GPT-5. 3-Codex changed the rules of the game and became ** a "digital worker" who can really work**. In the OSWorld test, which is dedicated to testing the operational capability of the real environment in AI,** its performance jumped from 38. 2 per cent to 64. 7 per cent in the previous generation.** This progress is tantamount to a shift from "sometimes" to "most of the time".

The coolest thing is its way of working: you can send it to a complex mission (e. g., "Improving the speed of loading our web site"), which ** will send you a "Progress Report" as human colleagues do:** "The photo compression has been completed, the code structure is being adjusted, and it is expected to take another 15 minutes. And by the way, I found a potential problem in the database.

The world of work

When Claude became your colleague:

** 9 a. m.** You opened the mailbox, and an in-depth analysis of the competitor's dynamics was lying still - it analyzed all the other's open messages for almost three months, even the CEO's statements at the industry forum.

** 10. 30 a. m.** You need to prepare the PPT for the afternoon board. Enter in PowerPoint: "A report with these three sets of data with a professional but not rigid style." Ten minutes later, a well-designed set of slides is ready, and even animated.

** At 2 p. m.**, a new project was launched. Claude was divided into four professional roles: one responsible for architecture design, one focused coding, one checking for potential errors and one writing technical document. Finally, it has seamlessly integrated all results, like experienced project managers.

When Codex joined your team:

** 3 a. m.** You're suddenly luminous: "Do a little program to record a dream!" An operational prototype with voice input and emotional analysis is ready at dawn.

** At 10 a. m.**, at the Daily Station. Codex volunteered to report: "Five attempted attacks were intercepted last night, optimizing database queries, and the average response time on the website was reduced by 23 per cent." More detailed than you thought.

** 3 p. m.** You suddenly think, "Let's make our online store more attractive to the Z generation." Codex does not ask "specifically what to do," but directly analyses data, studies trends, proposes complete solutions, and even begins to modify some pages.

Anyway, we're all happy for the entrepreneurs of one-man companies.

Philosophy contest: co-pilot or autopilot?

Behind this simultaneous publication is the fundamental difference between the two AI philosophies.

Anthropic takes the route of Human Empowerment * - AI should be a powerful tool to magnify human intelligence, but the key decision-making power is always in the hands of humanity. They're more like ** co-pilots*, ready to assist, but you always have the wheel.

OpenAI has chosen an "autonomous intelligence" orientation - AI should be able to perform its tasks independently and become a true digital colleague. Their AI is closer to** the auto-driving system** and, once the destination is set, it will handle most of the road itself.

Such differences are similar to differences in the concept of parenting: one side believes that guidance should be given but let the child try; the other side believes that children can grow up only with full autonomy.

AI sets

Some interesting little details are particularly striking in this competition:

** Self-evolution**: GPT-5. 3-Codex was developed,** using an earlier version to help calibrate and improve itself** - this may be the first AI to play a key role in its own creation, a bit like the science fiction of "self-born". We can finally explain the problem of having chickens before eggs.

-** Price philosophy**: Claude is like a fancy restaurant - chargeable on the basis of usage - Codex is like a buffet - buys tickets (subscribes) and uses them freely. Which is more cost-effective? It depends on your diet.

• ** Double-edged sword of cyber security**: OpenAi has itself classified Codex as a "high-capacity" model of cyber security - which means it is both an impenetrable shield and possibly an impenetrable spear. The tools themselves are not in the hands of good or evil.

Which one?

For those who struggle with "which one to choose", the answer is no longer self-defeating:** Why not both?**

One is the pocket of Doraemon, the other is the SEAL, which one do you think is appropriate?