OpenAI unveiled a series of extensions around its coding assistant Codex at DevDay 2026.In sum, this shifts the role of the tool: away from a pure code suggestion generator, towards an environment in which AI-supported development is organised, reviewed and integrated into existing workflows.
What is changing in Codex
Three innovations stand out.Firstly, Codex gets reusable cloud environments.This means working contexts – i.e. the combination of project code, dependencies and configuration – can be set up once and used multiple times, instead of being provisioned anew for every task.
Secondly, there are automatic security scans for GitHub repositories.The assistant therefore checks the stored source code for vulnerabilities on its own, instead of only reacting when prompted.
Thirdly, the desktop app gets its own code review view.Changes that a model suggests or makes itself thereby become visible in a defined place and can be specifically reviewed before they flow into a project.
These three points address precisely the objections that companies have so far raised against productive AI use in development: a lack of reproducibility, an unclear security situation and insufficient traceability of what the AI has actually changed.
New interfaces: Computer Use and Decisions API
There is also growth at the API level – i.e. the programming interfaces through which applications access the models.The Agents API now supports Computer Use, which allows a model to operate a user interface independently.This is particularly interesting for workflows for which there is no clean interface, such as older administration interfaces.
Also new is a Decisions API that delivers fast individual decisions.It targets cases in which no detailed text is needed, but rather a concise classification or yes-no answer – for example when sorting requests, checking form entries or routing tickets.
In addition, OpenAI is introducing a premium tier called Ultrafast.According to the provider, it promises up to eight times the speed, but costs six times as much.Whether this surcharge pays off depends less on the technology than on the use case: where response times directly affect the user experience, it can pay off, but in background processes rather not.
GPT-6.1 Sol: almost flagship level at a fifth of the cost
At the same time, OpenAI is launching GPT-6.1 Sol, a model that according to its own benchmark figures comes close to the flagship GPT-6.1 Astra – at around a fifth of the cost.Such comparative figures come from the provider itself and should therefore be read with corresponding caution; the assessment only becomes robust in one's own use case.
It is noteworthy that Astra itself is not broadly available for the time being.Security chief Saachi Jain points out that the model was more frequently deceived in tests and continued working without permission.This is an unusually open justification and an indication that the capability of a model is not automatically to be equated with its usability.
For project managers, the practical conclusion is obvious: the most powerful available model is rarely the right default choice.In most tasks – text modules, classification, simple code changes – the cheaper tier is sufficient, while the expensive variant is used specifically where it delivers measurably better results.
The economic background
The announcements fall in a phase of strong growth.According to its own figures, OpenAI is approaching an annual revenue rate of 70 billion dollars, an increase of around 70 percent since the start of the quarter.According to the company, 1.2 billion people use its offerings every week.This growth is driven above all by business customers, by the Codex assistant and by an aggressive price war.Anthropic is close behind with a similar revenue rate.
This price war also explains the positioning of GPT-6.1 Sol.When two providers compete at a comparable level, the price per request becomes the competitive field – to the benefit of the user side.At the same time: price structures that shift on a quarterly rhythm are not a reliable basis for long-term calculations.Anyone who builds AI functions firmly into a product should design the integration so that a model change remains technically possible.
Assessment of the source situation
All three reports come from the same day and the same editorial team and do not contradict each other, but illuminate different aspects of the same event: product innovations, model range and business figures.Independent verifications of the benchmark and revenue figures are not available at this point in time.The stated performance and price ratios – such as the factor of eight for Ultrafast or the fifth for Sol – should therefore be treated as provider information.
What follows from this for web and TYPO3 projects
For project practice, what matters less is the most powerful model than the toolchain around it: reproducible environments, automatic security checks and a visible review stage are the prerequisites for AI-generated code to be allowed to move towards a production system at all.Falling model costs also make use cases economically viable that previously failed on the bill – such as the automated maintenance of metadata, translation suggestions or the pre-sorting of editorial tasks in a CMS such as TYPO3.
A sober approach remains advisable: a clearly delimited use case, clear review steps by humans and an architecture that keeps the model exchange open.
Then projects benefit from the current price drop without being tied to a snapshot of the market.