OpenAI has released a new mid-range model with GPT-6.1 Terra.Compared with the predecessor GPT-5.6 Terra, the API prices are halved according to the provider, i.e. the costs for programmatic access to the model from within your own applications.At the same time, accuracy is set to increase.
The analysis firm Artificial Analysis sees this as a new benchmark for the ratio of intelligence to cost.The model is particularly suitable for everyday computer work and especially for AI-assisted computer operation – i.e. scenarios in which a model not only generates text but carries out work steps in interfaces and tools.
Bigger is not automatically better
Almost at the same time, OpenAI employee Ted Sanders explained the logic behind the model family.His key statement: the smaller Terra outperforms the larger Astra model in technical communication – and is significantly cheaper in the process.In most other areas, however, Terra lags behind Astra.
This differentiation is the actually interesting point for project managers.The widespread assumption that the largest available model also delivers the best result does not hold true in such a blanket form.Sanders explicitly points out that model size, reasoning – i.e. the ability to think through intermediate steps before an answer is produced – and multi-agent systems cannot be reduced to a single comparative value in the sense of an "IQ".
Multi-agent systems refer to setups in which several model instances work together in a division of labour: one instance plans, another executes, a third checks the result.How well such a setup works depends not only on the performance of the individual model but on the interplay.
What the two announcements show together
The product announcement and Sanders' classification complement each other.One describes that a certain performance class is becoming noticeably cheaper.The other explains why this performance class is even ahead of the larger sibling model in some disciplines – and in others it is not.
Together, this produces a picture that moves away from the pure ranking list.Instead of an overall winner, a portfolio of models with different profiles emerges, from which a selection is made per task.In the list cited by Sanders, alongside Astra and Terra, the names Sol and Luna also appear; detailed performance statements on these are not available from the information provided.
Consequence for model selection in the project
Anyone integrating AI into development and automation processes should therefore treat model selection as a project decision in its own right – comparable to choosing a database or a caching method.A simple structure helps:
- Sort tasks: Which work steps are repetitive and well-defined, which require broad world knowledge or complex conclusions?
- Assign a profile: For technical communication and everyday computer work, the information available speaks strongly in favour of the cheaper mid-range model.For tasks outside these strengths, the larger model remains in the running.
- Calculate cost per process: What decides is not the list price but the effort per completed work step – including repetitions in the case of faulty results.
- Measure quality: Without your own test cases from real day-to-day project work, every model choice remains a gut feeling.
Halved prices change the cost-effectiveness calculation
A halving of API prices shifts the threshold at which automation pays off.Use cases that were previously marginally uneconomical – such as the automatic pre-checking of content, the generation of technical documentation or support with recurring migration steps – can thus move into the plausible range.
At the same time, caution is advised in planning: price information and performance profiles of models change at short intervals.Architectures should therefore be built in such a way that a model change does not force a new development.In practice, this means: access to the model lies behind its own interface layer, prompts and evaluation logic are versioned, and a test collection exists against which a new model can be checked before it goes into production.
AI operation of the computer as a development line
Noteworthy is the explicit suitability for AI-assisted computer operation.This application area targets tasks that today often run via click sequences in browser interfaces and specialist applications and are difficult to automate via classic interfaces.If models reliably handle such processes at falling costs, a bridge to legacy systems that offer no modern API is created.
The fact that a mid-range model in particular is recommended for this fits the logic of such tasks: they usually consist of many small, clearly defined steps.Here, reliable repeatability at low unit costs counts for more than maximum depth in a single answer.
Classification for web and software projects
For projects, this means two things: model selection becomes a configuration parameter that should be justified per use case and regularly reviewed – using the largest model across the board costs money without a guaranteed gain in quality.At the same time, the architecture should remain interchangeable so that falling prices and new model generations can be used without rebuilding.
Anyone who builds a small collection of their own test cases today will have a solid basis for decision-making instead of guesswork at the next model change.