GPT5.4 Released, Fixes Major Bug Cheaper than Claude

Waking up, GPT has been updated again. This industry is so competitive, every day you surpass me, I surpass you, and I don't know when this competition will end.

It reminds me of the famous slap contest, where everyone is gritting their teeth, just waiting to see who can’t hold on first.

No wonder the head of a certain company's large model department is leaving; it's tough to live in a state of constant readiness.

Back to the point, GPT has released version 5.4, with thinking and pro launched simultaneously, now available on web, client, and API.

What are the major updates?

• ① The code capabilities of GPT-5.3-Codex have been merged in.

• ② Tool invocation and agent workflow are stronger, and some people are already saying it is the best partner for crayfish.

• ③ Enhanced web search and deep research capabilities, especially for multi-round searches, multi-source integration, and finding needles in a haystack.

• ④ Significant improvements in Office tasks like PPT, documents, and spreadsheets.

• ⑤ Computer operation capabilities have greatly improved, with evaluations stating it is stronger than humans. (OSWord test → Human: 72.36% | GPT-5.4: 75%)

• ⑥ Longer context and higher token usage efficiency, up to 1M context, now you can really paste an entire code repository. However, be aware that longer context is indeed very expensive.

• ⑦ Supports making changes while working, you can interrupt it and modify the prompts. (I tried it, and you can only interrupt and modify while it is thinking; once it starts outputting, you cannot.)

Overview of GPT-5.4 Highlights

Overall, compared to several predecessors: 5.2 is a strong reasoning model, 5.3-Codex is a strong coding model, while this 5.4 is an all-around model after capability integration, aimed at real professional tasks.

Interestingly, the 5.3-instant version was just released for a day when OpenAI hurriedly launched 5.4. Why is that?

Someone pointed out the key insight, 5.4 successfully fixed the major bug of pricing being cheaper than Claude, hahaha.

Looking closely at this latest pricing, it is indeed that OpenAI has significantly adjusted the price with this release, aligning the standard version basically with Sonnet4.6, and the Pro version's price has already surpassed Claude's flagship model Opus4.6.

However, the coding experts in the community have given very positive feedback on this update, stating "AI really becomes stronger with investment; GPT Pro can basically get through coding in one go, much easier than Claude which tends to go off track."

It seems that after receiving a staggering $110 billion in financing, OpenAI feels more pressure to make money.

Note: The title is "borrowed" with permission from coding expert Kang Xiaoning's social media.