OpenAI has released GPT-6 Astra to its first groups of paying ChatGPT users. For now, the model mainly targets people and organizations with higher-tier subscriptions, along with users accessing it through external cloud services. The release matters not only because of Astra’s new capabilities, but also because of unusually tight allowances and conflicting benchmark results.
Availability and costs at a glance
According to the report on Astra’s initial rollout, the model is available in ChatGPT Work and Codex for Pro, Enterprise, and Business Premium subscriptions. It can also be accessed through the OpenAI application programming interface (API), a technical connection used by other services, as well as Microsoft Azure and AWS Bedrock. Plus and additional Business accounts were expected to follow in the next few days.
People using ChatGPT on the Free or Go plans do not get access to GPT-6 Astra or GPT-5.6 Sol. The report does not list a separate chat price for Astra; usage comes out of existing subscription allowances. The $200 Pro subscription includes 200 GPT-6 Pro messages per week, while the $100 Pro subscription provides 50 per week. Business Premium also receives 50 messages per week, whereas Business Standard gets 15 per month.
The labels in the source are not entirely consistent. The rollout concerns GPT-6 Astra, while some of the specific table entries are described as GPT-6 Pro allowances. The article also says Business users will follow later even though it already lists a Business Standard allowance. You should therefore treat these details as a snapshot rather than a fully unambiguous pricing plan.
Astra is estimated to allow roughly half as many messages in each five-hour period as GPT-5.6 Sol. Plus users are listed at an estimated 5 to 45 messages rather than 10 to 100, while Pro 20x gets 100 to 900 instead of 200 to 2,000. According to OpenAI, these are not fixed limits: model selection, conversation length, reasoning effort, tool use, and caching can all affect actual consumption.
The sources provide no Switzerland-specific terms. They do not state prices in Swiss francs, assess the quality of German-language output, or explain whether data is processed in Switzerland or the European Union. General availability through ChatGPT and the named cloud services therefore does not amount to a specific Swiss commitment on access or privacy.
What Astra demonstrates in practice
The most striking practical example involves the computer game Portal. A developer known as cozyblaze gave GPT-6 Astra an initial goal and then allowed it to navigate the entire game without further human intervention. The AI agent, meaning a system that evaluates observations and independently selects its next actions, reached the end credits after about 23 hours and 43 minutes.
According to the account of the Portal run, the model received screenshots, the player’s position, and the camera angle. It selected inputs while the game paused during its reasoning periods. The experiment demonstrates long-term task planning and visual navigation, but not real-time processing: removing the pauses from the published video can make the run look smoother than it actually was.
The developer said the usage represented at least $570 at Astra’s listed rates, although the experiment used a $200 Codex subscription. That makes the run an impressive demonstration, but not yet a typical everyday use case. The report also acknowledges that many problems remain without quantifying them, so the result does not prove that Astra can reliably complete every extended task.
A second example concerns visual and spatial design. Simon Willison reports greater attention to detail and stronger understanding of prompts, particularly when creating sophisticated 3D models. His observed examples include renderings of gardens, shipyards, animals, cityscapes, and Dyson spheres. This is a qualitative assessment rather than a controlled comparison, but it indicates the kinds of creative work where Astra may be strongest.
Why benchmarks tell only part of the story
A benchmark is a standardized test used to compare models on a defined set of tasks. For GPT-6 Astra, these tests initially produced no consistent verdict. Epoch AI ranked the model first among 267 models with 169 points across more than 50 benchmarks, while ARC-AGI-3 showed a large or very large improvement depending on the measurement. Artificial Analysis initially placed Astra only around the level of its predecessor.
Following criticism, Artificial Analysis released version 4.2 of its Intelligence Index. In the revised benchmark assessment, Astra is four points ahead of its predecessor but remains behind Anthropic’s Claude Fable 5.1. Meta ranks third. The gap between the different assessments has narrowed, but the disagreement has not disappeared.
Artificial Analysis also changed the index itself. It added practical knowledge-work tasks and document analysis spanning multiple PDF pages, while removing a test that models had already solved. In addition, 40 percent of the weighting now comes from private test data, which is intended to make deliberate optimization for known questions more difficult.
For you, the ranking is therefore a useful signal rather than a purchasing decision. Results depend on which tasks an index selects and how it weights them. A model may perform strongly in document analysis, visual design, or long-term navigation while proving less convincing with your own writing tasks. Revising an index after criticism is not automatically a flaw, but it does show how fluid these comparisons remain.
How better prompting gets more from Astra
Prompting is the process of giving an AI model instructions and relevant context. According to OpenAI, Astra asks follow-up questions more often than GPT-5.6 Sol instead of independently filling in missing information. This can reduce mistakes caused by unsupported assumptions, but it can also stop tasks early when you expect the model to keep working on its own.
OpenAI’s published prompting guidance recommends explicitly defining the level of initiative you want. Phrases such as “Can you…,” “I want…,” or “Help me…” should be treated as instructions to act, not merely as reasons to ask another question. For longer tasks, you can also tell Astra to infer your intent from the conversation and continue until it has produced a result you can verify.
For beginners: Start with a clearly limited task and describe the desired output. You can tell Astra to complete the instruction, identify any necessary assumptions, and ask a question only when it encounters a genuine obstacle. Review the result before assigning a larger follow-up task, because tight message allowances make aimless experimentation rather expensive.
For advanced users: Define which instruction takes priority when different parts of the context conflict, and specify when Astra should request approval. OpenAI recommends asking for permission only after the model has prepared a concrete, verifiable result. If Astra pauses unexpectedly or changes direction, you can ask it to identify the exact contextual instruction that triggered the behavior.
OpenAI says Astra follows longer instructions better than its predecessors but is also more sensitive to conflicting information in its context. In complex work environments, older rules or so-called skill files, meaning documents that contain persistent working instructions, can block the model or send it in an unintended direction. More context is therefore not automatically better; consistent instructions with clear priorities are more useful than the largest possible stack of rules.
Pros and Cons of GPT-6 Astra
Pros:
- Extended tasks – The Portal experiment shows that Astra can pursue a multistage objective independently over many hours.
- Visual design – Observations involving gardens, cityscapes, and other 3D models suggest stronger detail in spatial outputs.
- Efficient token use – Artificial Analysis rates Astra as more efficient than other leading models in its use of tokens, the small text units processed by a model.
- More controllable collaboration – Follow-up questions and explicit rules for initiative can reduce unsupported assumptions.
Cons:
- Tight allowances – Depending on the subscription, Astra permits substantially fewer messages than GPT-5.6 Sol.
- Expensive experiments – The developer valued the Portal run at a minimum of $570 at listed rates, making it an impractical casual test.
- Context sensitivity – Conflicting or poorly prioritized instructions can stop work or push it in an unexpected direction.
- Uncertain comparisons – Diverging benchmarks and a later index revision make definitive performance claims difficult.
GPT-6 Astra combines notable abilities in extended, visual, and complex tasks with a restrictive initial rollout. It may be worth testing for paying users whose specific use case justifies the limited message allowance, but the available demonstrations do not replace trials with your own work. The main unresolved risks are reliability beyond selected experiments and the lack of specific information about language quality, availability, and privacy conditions in Switzerland.
Sources
- OpenAI schaltet GPT-6 Astra für erste zahlende Nutzer frei mit reduzierten Kontingenten – Unknown, 2026-09-05
- KI-Agent steuert sich durch Portal: GPT-6 Astra meistert das komplette Spiel ohne menschliche Hilfe – Unknown, 2026-09-07
- Introducing GPT-6 Astra for developers – Unknown, 2026-09-05
- Artificial Analysis korrigiert eigenen KI-Benchmark nach Kritik an GPT-6-Astra-Bewertung – Unknown, 2026-09-05
- KI-Slop vermeiden: OpenAI teilt Prompting-Tipps für GPT-6 Astra – Unknown, 2026-09-05
Image: RDNE Stock project via Pexels