Three reports are currently available on GPT-6 Astra, which together paint an illuminating picture: a prompting guide from the manufacturer, an internal productivity promise and a subsequently corrected benchmark assessment.For companies that want to use AI support in software development, these are three very different types of information — and they should also be weighted differently.
The prompting guide addresses concrete annoyances
OpenAI has published a detailed prompting guide for GPT-6 Astra.Prompting refers to the way a task is formulated for a language model — including framework conditions, role description and specifications for the output format.
Three points in this guide are particularly noteworthy because they relate precisely to the points of friction that stand out in everyday use:
- More initiative: The model can be instructed to complete tasks more independently instead of asking back at every intermediate step.
- Fewer typical AI phrases: The manufacturer describes how the familiar stock phrases can be suppressed that inflate results without adding substance — often referred to online as AI slop.
- Less sprawling test behaviour: In programming tasks, the model apparently tends to generate an excessive number of tests.According to the guide, this can also be limited through specifications.
These notes are the most practically useful part of the current reporting situation.They also confirm an experience from projects: A model does not become useful because it is powerful, but because the task is framed narrowly enough.Anyone who lets a language model work without specifications on code style, test scope and output format gets results whose post-processing eats up the time gained again.
The fact that the guide explicitly addresses scaling back test behaviour deserves a note.Tests are fundamentally desirable.It becomes problematic when a model produces large quantities of superficial test cases that have to be maintained and still say little about the correctness of the application.The decision about where the sensible limit lies remains a professional one and cannot be delegated to the model.
Six months ahead: an internal experience report, not a measurement
OpenAI developer Thibault Sottiaux describes Astra as the greatest competitive advantage as long as the tool was not publicly available.Internally, it increased productivity so much that plans were brought forward by six months.
This figure is impressive and should nevertheless be read with caution.It is a statement by the manufacturer about its own development work — not an independently verified measurement and not a figure that can be transferred to arbitrary project contexts.
The context of an AI lab differs clearly from that of a mid-sized web or software project in several respects:
- The teams know the model and its peculiarities better than any external user.
- Codebases, toolchains and processes are tailored to this way of working.
- A large proportion of the tasks are technical and internal — without coordination with specialist departments, without approval loops, without legacy systems that have grown over time.
The last point in particular is decisive.In typical client projects, a considerable part of the runtime is not spent writing code, but in requirements clarification, data migration, coordination, approvals and editorial work.If AI support accelerates code production, initially only one section of the chain is shortened.Total runtime only falls noticeably once the surrounding steps are also taken into account.
The benchmark dispute shows how shaky rankings are
The third development is the most instructive in its sobriety.Artificial Analysis has published version 4.2 of its Intelligence Index — apparently also in response to criticism that the previous ranking did not correctly reflect the progress of GPT-6 Astra.In the new version, Astra gains four points compared with the previous model, but remains behind Anthropic's Claude Fable 5.1.
What is remarkable here is not the ranking but the mechanism: The assessment of a model has changed without the model itself changing.What changed was the measurement method.Anyone who bases investment decisions on benchmark positions is therefore basing them on a quantity that depends on definitional questions of the test provider.
In practice, this leads to a simple consequence: Benchmarks are useful as guidance for which models belong in the shortlist at all.They are not suitable as a basis for deciding which model delivers the better results in a concrete project.This question can only be answered with your own, repeatable tasks from your own context — for example a typical TYPO3 extension, a data migration or a refactoring in your own codebase.
How to turn this into a robust approach
From the three reports, a pragmatic approach can be derived that does not require forecasts about future model generations:
- Treat prompting as a project artefact. Specifications on code style, test scope, output format and tone belong documented and versioned — not in the heads of individual developers.
- Build your own reference tasks. A handful of real tasks from your own project against which new model versions are tested is more valuable than any external ranking.
- Do not reduce review. Higher speed in code creation shifts effort into review.Whoever cuts there shifts costs into operations.
- Read manufacturer specifications as a hypothesis. Internal productivity figures are an indication of what may be possible — not of what will occur in a different environment.
For web and software projects, this means: the realistic gain from AI-assisted coding today lies less in spectacular leaps in time than in constant relief from routine work — provided that specifications and review processes are clearly defined. Benchmark positions should play no role in budget planning, because they can change without anything changing in the model. Whoever relies on their own reference tasks and documented prompting standards instead of rankings makes decisions that will outlast the next model generation.