Recommendation
OpenAI's new inference models o3 and o4-mini have brought revolutionary breakthroughs in both performance and price.
Core content:
1. The outstanding performance of o3 and o4-mini in key benchmarks
2. The test score improvement after the new model integrates ChatGPT function
3. The high cost-effectiveness of o4-mini and its impact on the intelligent startup market
Yang Fangxian
Founder of 53A/Most Valuable Expert of Tencent Cloud (TVP)
「
OpenAI sounded the clarion call for market cleansing.
On April 17, OpenAI released new reasoning models o3 and o4-mini.It was really surprising! Because the new version of the large model GPT-4.1 was just released the day before yesterday, the industry thought that OpenAI would not release new models in the short term. Previously, Altman clearly stated that the o3 model would not be released separately, and it would be integrated into GPT 5 and provided to users. Unexpectedly, this time there was a big news after another big news!
In addition to making history and topping key benchmarks such as Codeforces, SWE-bench, and MMMU, o3 and o4-mini also achieved impressive test scores after integrating ChatGPT's online search, Python environment, local search, structured data, and other functions into the new model.
The two new inference models also introduced image reasoning for the first time. Actual tests by netizens showed that the new model can analyze the exact geographic location of photos taken by users based solely on the content of the pictures.What’s even more exciting is that the smaller, faster and cheaper o4-mini has benchmark data that is just as good as o3!The price is even more explosive!The input price of o4-mini million tokens is 1.1 USD, the output is 4.4 USD, and the cache hit is 0.28, which is cheaper than GPT-4.1!We really have to thank DeepSeek for pushing OpenAI to this extent!But on the other hand, OpenAI’s new model has actually cleansed the existing GenAI large model startup market! It can be said to be a bloody cleansing!When GPT-4.1 was released, the market still thought that its programming ability was average and not worth worrying about. However, o3 and o4-mini not only took over programming, but also their in-model capabilities plus ChatGPT integration services directly took over the general AI Agent market.- OpenAI has leading basic model technology and various types of reinforcement learning technology, which can accurately allocate which parts of the intelligent agent services use in-model services and which refer to external process technologies.
- This time, o3 and o4-mini use their own python environment to directly solve a large number of mathematical, scientific and programming intelligent agent needs. The AiPy just released by Knowsec last week also follows this idea, using the powerful software library ecosystem of the python programming language to build a dedicated python use for large models to expand the capabilities.
- However, it now seems that o3 and o4-mini have directly internalized a large number of Python software libraries into services within the model, and can efficiently complete related services without having to load a large number of related libraries locally.
Regarding the potential of python use and code interpreters, interested readers can click to read " Domestic AiPy promotes the hidden functions of OpenAI, and open source local deployment can also make old computers usable ."In addition, in order to further extend its code capabilities, OpenAI directly came up with an open source local code agent: Codex CLI. Please note the keywords: open source, local deployment, command line tool, lightweight!Codex CLI can help you complete existing programming tasks with extremely low local resource usage!This is also the goal of AiPy, and AiPy currently exists in the form of a command line tool. Judging from the experience of AiPy, this goal has been achieved very well, and it can be used on old computers with extremely low configurations.The hunting mode has been activated. How can the intelligent startup team survive?In fact, OpenAI has launched a large number of AI Agent services, and has been steadily developing towards simpler and more complete developer services, lower local resource usage, and more efficient in-model services.Currently, it includes ChatGPT tasks, Operator, Deep Research, and the open source agent orchestration framework Agents SDK. In fact, the earlier GPTs can also be regarded as an attempt at the agent form.It now seems that in addition to the computer use computer control function of Operator, which still has value, the Agents SDK can customize local deployment services for the enterprise market. Other services are on the verge of being eliminated in front of o3 and o4-mini.When AI epiphany emerges, we are even more looking forward to GPT 5 and A-SWE all-round programming agents.