Openai launches a preview of Codex Ai Swe Agent

a surprise! Just days after Reports showed that Openai was buying Windsurf, starting white codingIt seems that the previous company launches its competitors’ service as a research inspection under the name Codex brand name, and is heading to face against Windsurf and Cursor, and the increasing list of artificial intelligence coding tools offered by startups and large technology companies including Microsoft and Amazon.

Unlike the former AI Codex model from Openai, the new version is a phrase O3 openai thinking model Multiple development tasks can be carried out in parallel.

Starting today, it will be available for Chatgpt Pro, Enterprise and Team users, with support for Plus and EDU users soon.

Evolution of the manuscript: from a model to an independent artificial coding factor

This version represents an important step forward in the development of Codex. the The original manuscript appeared for the first time in 2021 As a model for translating the natural language into a symbol available through the emerging application programming interface in Openai.

The engine was behind CopilotThe famous coding assistant similar to the automatic completion designed to work within IDES such as Visual Studio Code.

This initial repetition focused on generating the code and completing it, the trainer on billions of general source code lines.

However, the early version came with restrictions. He was VulnerableSuggestions of icon is unsafe, and the biases included in their training data. Codex sometimes suggested a correct, superficial code, and in some cases, problematic links were made based on claims.

Despite these defects, it has shown an adequate promise to create artificial intelligence coding tools as a rapidly growing product category. Since then this original model has been neglected and turned into the name of a new collection of products, according to an openai spokesman.

GitHub Copilot officially moved on the Openai manuscript model in March 2023, and I adopted GPT-4 as part of Copilot X to enable the deepest IDE integrationChat capabilities, and more perceived code suggestions in the context.

Agents visions

The new manuscript exceeds its predecessor. It is now designed to work independently at longer periods, CODEX can write features, fix errors, answer questions for the code base, operating tests, and suggest withdrawal requests-every task working in a safe and insulated cloud sand box.

The design reflects the broader Openai’s ambition to overcome fast answers and to cooperative work.

“We are thinking about AI systems that can work on your behalf for a longer period of time to complete the large work pieces by interacting with the real world,” Josh Tubin, who leads the Openai agents research team, said during a recent press conference. Codex directly fits in this definition. He added: “Our vision is that Chatgpt will become like a virtual workshop – not just answering quick questions, but cooperating in great work through a set of tasks.”

The numbers released by Openai show that the new Agent Codex-1 SWE surpasses all thinking models in Openai on SWE’s internal tasks.

New possibilities, new interface, new workflow

Codex tasks are started through the interface of the sidebar in ChatGPT, allowing users to demand the agent to tasks or questions.

The agent treats each request in a closed air environment loaded with the user’s repository and its formation to reflect the development setting. It records its actions, indicates the test outputs, and summarizes the changes – which makes its work can be followed and reviewed.

Alexander Egrikos, head of the Openai’s Desktop & Agents (and former CEO and co -founder of The start of the Screenshare Collaboration Multi For a group that was not disclosed last year) in a briefing with reporters that “the manuscript agent is the software engineering agent based on the group of costumes who can work in many tasks in parallel, with his computer to operate it safely and independently.”

He said that the engineers are already using it “like the morning tasks list-to extract tasks to Codex and return to a set of draft solutions ready for review or merging.”

Codex also supports the composition through Agents.MD-Evidence at the level of projects that the agent teaches how to navigate at the base base base, run specific tests, and follow home coding patterns.

“We have trained our model to read the code and deduce the style – such as if the Oxford Oxis should be used – because the code style matters as much as it matters to the right,” said positive.

Security and practical use

Codex carries out tasks without accessing the Internet, and only drawing on the software instructions provided by the user and its consequences. This design guarantees safe operation and reduces potential misuse.

“This is more than just a model programming interface,” said Embiricos. “Since it works in a closed air environment with human review, we can give the model a lot of freedom safely.”

Openai also reports of early external use. Cisco evaluates Codex to accelerate engineering work through its production lines. It is used as time to run background tasks such as correcting errors and writing the test. Superhuman works to take advantage of the manuscript to improve test coverage and enable non -engineers to suggest lightweight symbol changes. Kodiak, an independent complex company, improving the code reliability and gaining visions about the unusual stack components.

Openai also put up updates to Codex Cli, the light -minute party agent for local development. CLI now uses a smaller model-Codex-MINI-Latest- It is identified for the low editing of the assignment and Q&A.

Pricing is determined at $ 1.50 per million input codes and $ 6 per million output symbols, with 75 % interim storage discount. Codex is currently free for use during the operating period, with price limits and pricing options on demand.

Does this mean that Openai does not buy Windsurf? *Thinking about the face of emojis*

The Codex version comes amid the increase in competition in the area of ​​artificial intelligence coding tools – and the signals that Openai intends to build, instead of buying, the next stage of products.

according to Modern data is similarThe traffic to AI, which focuses on developers, has increased by 75 % over the past 12 weeks, which emphasizes the increasing demand for coding assistants as a basic infrastructure instead of experimental additional functions.

Reports from Techcrunch and Bloomberg Openai suggested acquisitions with a fast -growing AI Dev Tool indicator and Windsurf. It is claimed that the indicator has moved away from the table. According to Windsurf, in principle, it was obtained by Openai at $ 3 billion, although no formal deal is not formally confirmed either from Openai or Windsurf.

Only yesterday, in fact, Windsurf appeared for the first time in the family-focused basic models family, SWE-1, It is designed for this purpose to support the life cycle of full software engineering, from correcting errors to long -term project maintenance. Swe-1 models are reported, and they are fully trained at home using a new serial data model designed on the progress of the real world development work.

Many things may happen behind the scenes between the two companies, but for me, the timing of Windsurf, which launches a model for its own coding-instead of its strategy to the use of Llama variables and giving users the option of the opening in Openai and anthropor models-followed by Openai a day after its opponent is launched from its Windsurf, it seems that it does not indicate that the two do not They compete soon.

But on the other hand, the fact that the new Codex Ai SWE agent in “Search Section” to start may be a form of Openai to pressure Windsurf, an indicator or anyone else to come to the bargaining table and reach a deal. When asked about the possibility of Windsurf and reports of one of them, an Openai Venturebeat spokesman told that they had nothing to share in that front.

Either way, IMRIERICOS develops the nutritional constitution with more than just a tool or assistant.

He said: “We are about to undergo a seismic transformation in how developers work with agents – not only to associate with them in the actual time, but to delegate the tasks completely.” “The first experiences were just thinking models with peripheral access. The experience was magic – they started doing things for us.”

Built

Codex is designed with the position of professional developers in mind, but Embiricos indicated that even product managers found it useful to suggest or verify the validity of changes before withdrawing in human SWES. This diversity reflects Openai’s strategy of building tools that increase productivity through technical teams.

Train, an engineering company in the project, summarized the broader ambition behind Codex: “This is a transformational change in how software engineers interact with artificial intelligence and computers in general. It inflated everyone’s potential.”

Openai Codex is envisioned by a new development workflow as engineers assign high -level tasks for agents and cooperate with them in an unequivocal way. The company builds a deeper integration via GitHub, Chatgpt Desktop, TricSers Trackors and CI. The long -term goal is to mix the actual time and delegate long -horizontal tasks in a smooth development experience.

Josh Tobin said, “Coding supports many useful things throughout the economy. The acceleration of coding is a high -level way to distribute the benefits of artificial intelligence to humanity, including ourselves.”

Whether Openaii closed deals for competitors or not, the message is clear: codex here, and Openai is betting on its agents to lead the next chapter in the developer productivity.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.