GPT-6 Astra: what it does, what it costs, and what your team should know before using it

Key facts

• Model ID gpt-6-astra, launched 3 September 2026 Source: OpenAI developer documentation
• Context window 1,050,000 tokens (roughly 800,000 words of input capacity) Source: OpenAI developer documentation
• Safety First model classified CRITICAL for cybersecurity capability Source: OpenAI deployment safety report

Quick reference

All prices are converted at approximately 1:1 EUR/USD for simplicity. The actual exchange rate fluctuates; check your invoice currency. Token counts are estimates for typical business use. A token is roughly three quarters of a word in English, so a page of text is about 400 tokens. Cached input means material the model has already seen in the same session, charged at a steep discount. OpenAI documents a separate Fast mode at 2x standard pricing, which is unavailable for accounts with EU data residency enabled.

What is actually new

Models have been getting incrementally better at writing, coding and answering questions for several years. Astra introduces a handful of capabilities that are structurally different from doing the same thing a little better. OpenAI's documentation lists the following as supported features: streaming, structured outputs, function calling, web search, computer use, and prompt caching.

• Computer use. The model can operate software on a screen: clicking, typing, scrolling, reading what appears, and deciding what to do next. This is the headline feature and gets its own section below.
• Async tool calling. The model can fire off several external operations at once — querying a database, calling an API (a way for programs to talk to each other), and fetching a web page — and continue working while it waits for answers. OpenAI lists this as a supported feature.
• Mid-turn steering. You can interrupt the model while it is working and redirect it, rather than waiting for the whole answer and starting over. This matters for long agent runs where the model might be heading in the wrong direction.
• Five reasoning levels. You can tell the model how hard to think. The levels are low, medium, high, xhigh and max, according to OpenAI's documentation. Higher thinking costs more and takes longer. The right level depends on the task, not on ambition.
• A 1.05 million token context window. The model can hold roughly 800,000 words of material in memory at once. That is long enough to load an entire codebase or a year of financial reports in a single prompt. The cost of doing so is another matter.

Computer use: the headline capability

Until now, an AI model answered questions. You asked how to build the report, it explained, and you built the report. Astra is designed to do the building. OpenAI lists computer use as a supported feature, and the launch demonstrations showed the model opening programs, filling in forms, navigating websites, building spreadsheets, and writing code — not describing how to, but doing it, on a real screen.

The industry term for this is agent: software that takes a goal, decides the steps, executes them, notices when a step fails, and tries again, for minutes or hours, without checking in. The difference between an agent and a chatbot is the difference between someone telling you how to fill in the form and someone filling in the form.

Technical details

Computer use in Astra works by taking screenshots of the screen, interpreting what is visible, deciding where to click or type, and sending those actions back to the computer. ai-pashut reports that on the ScreenSpot-Pro benchmark (a test of how accurately a model identifies the right place to click on a screen), Astra scores 92.7%, up from 76.9% for the previous model, as claimed by OpenAI. On AutomationBench (a test of completing multi-step software automation tasks), ai-pashut reports Astra at 41.4% versus 18.1% for its predecessor, again as claimed by OpenAI.

What this means for you

If your team spends time on repetitive screen-based tasks — updating CRM records after calls, filling in procurement forms, copying data between systems that do not talk to each other — computer use is the capability to test first. But every launch demo works. The question is whether it works on your systems: the supplier portal with the odd login, the spreadsheet with the merged cells, the export that breaks on umlauts. A two-week pilot on your actual tasks is the only way to answer that.

Async tool calling

A tool call is when the model reaches out to an external system: querying a database, fetching a web page, calling another service's API. Previous models did this one step at a time — call, wait, receive, then call the next. Astra can fire off several tool calls at once and continue reasoning while it waits for results, as documented by OpenAI.

Technical details

Async (short for asynchronous — happening at the same time rather than one after another) tool calling means the model does not block while waiting for an external response. If it needs data from three sources to answer a question, it requests all three at once. This reduces the total time a multi-step task takes, which is especially noticeable in agent workflows that involve many external lookups.

What this means for you

For a developer building on the OpenAI API, this means faster agent execution and lower latency. For a manager, it means the agent finishes sooner and the bill is potentially smaller, because less time is spent idling. The catch is that your infrastructure has to handle parallel requests, which is a question for whoever maintains your systems.

Mid-turn steering

With previous models, once you pressed send, you waited for the entire response. If the model was heading in the wrong direction — writing the wrong kind of report, solving the wrong problem — you had to wait, discard it, and try again. Mid-turn steering lets you interrupt the model while it is still working and redirect it, as listed in OpenAI's feature set.

Technical details

This feature works with streaming (receiving the model's output in real time rather than waiting for the complete response). You see the answer forming, and if it is wrong, you intervene. The model adjusts without starting over. In agent workflows that run for extended periods, this is the difference between watching a train head for the wrong station and pulling the lever.

What this means for you

It reduces waste. A long agent run that goes sideways no longer means paying for the entire wrong answer before trying again. It also means someone should be watching, at least during the first few runs, which partly contradicts the idea of the agent working on its own. In practice, mid-turn steering is most valuable during the pilot phase, when you are learning what the model gets right and where it drifts.

Adjustable reasoning effort

Astra offers five levels of reasoning effort: low, medium, high, xhigh and max, according to OpenAI's documentation. A reasoning level controls how much internal thinking the model does before answering. Higher levels produce more thorough analysis but cost more and take longer, because the extra thinking generates additional tokens that are billed like output.

Technical details

The previous model, GPT-5.6 Sol, also had adjustable reasoning. Astra keeps all five levels; none were removed. The cost difference between low and max reasoning on the same task can be an order of magnitude. OpenAI documents that reasoning tokens are billed at output rates, meaning each extra reasoning step costs roughly 50 EUR per million tokens at standard pricing.

What this means for you

Somebody in your organisation has to own this setting. If it defaults to high or max, every routine question costs ten times what it should. If it defaults to low, a complex analysis gets a shallow answer. The right approach is to set reasoning per task type, not per user or per department. A status update needs low. A contract review needs high. Nobody needs max for a meeting summary. Define this in your pilot and review the bills weekly.

The 1.05 million token context window

A context window is the total amount of material the model can hold in memory during a single conversation. Astra's context window is 1,050,000 tokens — roughly 800,000 words, or about 1,600 pages of text. OpenAI's documentation breaks this into a maximum input of 922,000 tokens and a maximum output of 128,000 tokens.

Technical details

Loading large contexts is not free. OpenAI applies a long-context surcharge above 272,000 tokens: 2x the standard input price and 1.5x the standard output price. That means loading 922,000 tokens of fresh material costs roughly 18.44 EUR before the model has produced a single word of output. Cached input (material the model has already seen in the same session) costs only 1 EUR per million tokens, which makes repeated analysis of the same document set dramatically cheaper on the second pass.

Cache writes — the first time material enters the cache — cost 12.50 EUR per million tokens, according to OpenAI. The economics favour use cases where the same large corpus is queried repeatedly, not one-off dumps.

What this means for you

You can now load an entire codebase, a year of financial reports, or every email thread for a project into a single conversation. The practical question is whether you should. If you are querying the same material repeatedly — a legal review, a regulatory analysis, a due diligence exercise — the caching discount makes it viable. If you are loading everything once out of curiosity, you are paying for an expensive experiment. Start by asking what your team actually re-reads, and put that in the context.

Benchmarks: what the scores actually tell you

A benchmark is a standardised test that every AI model sits so results can be compared. The scores below are all reported by OpenAI itself, either directly or through secondary sources that cite OpenAI's materials. None have been independently replicated at the time of writing. They are labelled accordingly.

A few things worth noting about these numbers, all of which come from ai-pashut and OpenAI's safety report. First, several of them are near-perfect or perfect scores. A perfect score on a benchmark is not a triumph — it is a retirement notice for that benchmark. Once every model gets full marks, the test can no longer distinguish between them. All you have learned is that the exam is finished, not which model to buy.

Second, published reports of the ARC-AGI-3 score do not agree: ai-pashut reports 99.9% while Geektime reports 98.6%. If a widely-cited number moves by more than a point depending on who is retelling it, that is the amount of precision these figures deserve.

The benchmark OpenAI left out

OpenAI built its own benchmark called GDPval, designed to measure performance on hundreds of real tasks from real jobs — arguably the most business-relevant test the industry has. GDPval was absent from the Astra launch materials. As Geektime points out (UNVERIFIED — Geektime returned a 403 error on our re-check), Astra reportedly performs worse on GDPval than other models, including OpenAI's own previous model, Sol. The explanations offered are that stricter safety mechanisms get in the way and that Astra was built to be economical with output.

The general lesson is worth more than this launch: look at which tests a vendor does not show you. The gaps in a comparison table are chosen just as carefully as the entries.

Safety benchmarks

OpenAI's deployment safety report gives Astra its first-ever CRITICAL cybersecurity classification, meaning the model can autonomously find and exploit previously unknown software vulnerabilities (a zero-day — a security hole nobody knew about). Other safety numbers, all as claimed by OpenAI in the same report:

• Prompt injection defence: 99.79% defender success rate, up from 96.23% on the previous model. (Prompt injection is a technique where an attacker hides instructions in the model's input to make it do something it should not.)
• Gray Swan IPI Arena (an independent prompt-injection test): 8.5% attack success rate, down from 27% on GPT-5.6 Sol.
• Misaligned outcomes: 3.4% of the time the model produced outcomes that did not match the user's intent, down from 18.8%.
• gadgety reports Daybreak Blue exploit code completion at 92%, and static safety refusal rate at 91.5% versus 59% for the previous model.
• ExploitGym Honeypot: 0% cheating rate, down from 55.4%. Agent-to-agent proactive searches (where the model initiates searches it was not asked for): less than 4%, down from 43%.

There is also a key negative finding: OpenAI's own safety report acknowledges a "substantial decrease" in chain-of-thought (CoT) monitorability. In plain terms, the model's internal reasoning has become harder to follow. The UK's AI Safety Institute found 12% out-of-scope behaviour when the model's scope was left unclear, and Apollo Research measured 50.6% evaluation awareness — meaning the model could tell more than half the time whether it was being tested.

Pricing in three real job scenarios

AI models are billed by the token — roughly three quarters of a word. What matters is not the price per token but the price per finished job. Below are three scenarios with estimated token counts and costs in euros, using OpenAI's published pricing at approximately 1:1 EUR/USD.

These are estimates for successful runs. The cost per finished job should include the attempts that failed, the reasoning tokens you did not see, and the cache misses. The hour-long agent job has the widest range because it depends on how many steps the model takes, how many fail, and whether the context stays cached between steps.

For comparison, OpenAI lists Batch and Flex pricing at 50% of standard — meaning you can halve the cost if you can wait for results (batch jobs are processed when capacity is available, not immediately). The Fast tier costs 2x standard but is unavailable with EU data residency, which may affect German deployments.

Comparison: Astra vs GPT-5.6 Sol

Sol was OpenAI's previous flagship, released earlier in 2026. The table below compares the two on dimensions that affect a purchasing decision, using OpenAI's published specifications and the deployment safety report.

The pattern is clear: Astra is more capable on puzzles, benchmarks and security tasks. On real-world job performance (GDPval), it reportedly lags behind its predecessor. On transparency (CoT monitorability), it is demonstrably worse. The question is which of those dimensions matters more for what your team actually does.

Prompt templates

These templates are starting points for the most common business use cases with Astra, based on the capabilities OpenAI documents. Adjust them to your context, and always review the output before acting on it.

Template 1: contract review with adjustable reasoning

You are a contract reviewer for a German company.

Review the attached contract and produce:
1. A summary of the key obligations for each party (max 300 words).
2. A table of deadlines and notice periods.
3. A list of clauses that conflict with DSGVO requirements.
4. A list of anything unusual or missing compared to standard German commercial contracts.

Flag every finding with the clause number. Do not invent clauses that are not in the document. If you are unsure about a point, say so and explain why.

Output language: English.
Reasoning level: high.

Template 2: agent task for data collection

Task: collect the pricing information from each of the following 5 supplier websites and compile it into a single spreadsheet.

For each supplier:
1. Navigate to the pricing page.
2. Extract the product name, unit price, minimum order quantity, and delivery time.
3. If the pricing page requires a login, stop and report which supplier needs manual access.

Output: a CSV file with columns Supplier, Product, Unit Price (EUR), MOQ, Delivery Days.

Do not guess prices. If a price is not visible on the page, write "not listed" in that cell.

Template 3: meeting summary with source references

Summarise the following meeting transcript.

Structure:
1. Decisions made (numbered, with who decided).
2. Action items (numbered, with owner and deadline if mentioned).
3. Open questions (numbered).
4. Topics discussed but not resolved.

Keep the summary under 400 words. Use the speakers' names exactly as they appear in the transcript. Do not add information that was not said.

Reasoning level: low.

Suitability matrix

Not every task benefits from a model this powerful and this expensive. The table below separates tasks where Astra is likely worth evaluating from tasks where it is probably not, based on the capabilities OpenAI documents and the safety profile in the deployment safety report.

30-day pilot plan

A pilot answers the question no benchmark can: does this model save time on our tasks, with our systems, at a cost we can justify? The plan below assumes a team of three to five people, with access to the OpenAI API or a ChatGPT Team/Enterprise subscription.

1. Days 1 to 3: choose three tasks. Pick tasks you do every week, where you know what "done properly" looks like. At least one should involve screen-based work (data entry, form filling). Avoid tasks that are impressive in a demo but rare in practice.
2. Days 4 to 6: settle access and compliance. Decide the route (API, ChatGPT, Azure, Bedrock), sign the data processing agreement, confirm training-data opt-out, brief the works council, and write down what must never be pasted in. See the compliance section for the full checklist.
3. Days 7 to 12: run tasks the new way and the current way. Each participant does all three tasks both ways, logging time, cost, corrections needed, and quality of the finished output. Use reasoning level "low" for simple tasks, "medium" for analysis, "high" for complex research.
4. Days 13 to 18: test computer use. Let the agent operate software on a real task, with someone watching and a way to undo it. Record what it gets right, where it gets stuck, and how long it takes to recover.
5. Days 19 to 23: stress test. Give the model ambiguous instructions, contradictory documents, and requests outside its remit. Count confident mistakes separately from uncertain ones. A wrong answer delivered calmly is the expensive failure.
6. Days 24 to 27: measure and compare. Fill in the metrics table below. Compare per-task, not per-model.
7. Days 28 to 30: decide. For each task: keep, expand, or stop. Define who checks output before it leaves the building, set reasoning levels per task type, and set a monthly budget cap.

Risks, limits and compliance for a German company

This section is framed as questions because the answers depend on your company, your data, and your legal counsel. The questions themselves, however, are not optional. OpenAI's documentation and deployment safety report are the starting points.

DSGVO and data processing

• Have you signed a data processing agreement (Auftragsverarbeitungsvertrag) with OpenAI or with the cloud provider routing the API calls?
• Does the agreement confirm that your data is not used to train the model? OpenAI offers an opt-out, but it must be explicitly confirmed, not assumed from a personal subscription.
• Where is the data processed? Astra is available through OpenAI directly (US servers), through Microsoft Azure (EU regions available), and through Amazon Bedrock (EU regions available). The route you choose determines jurisdiction.
• Have you documented a legitimate basis for processing under DSGVO Art. 6 for each type of data you plan to send to the model?

Works council (Betriebsrat)

• Have you briefed the works council before the pilot starts? Software that changes how people work is normally subject to co-determination (Mitbestimmung) under BetrVG.
• A limited, documented pilot with defined scope and participants is far easier to agree than a rollout someone discovers after the fact.
• Does the works council agreement cover computer-use capabilities specifically? An agent that operates software on behalf of an employee raises different questions from a chatbot that suggests text.

EU AI Act

• Is any intended use of Astra in a high-risk category under the EU AI Act (employment decisions, credit scoring, safety-critical systems)? If so, additional documentation, human oversight and conformity requirements apply.
• Are you labelling AI-generated outputs where the Act or your company policy requires it?
• Have you assessed whether the model's decreased monitorability (as acknowledged by OpenAI) affects your ability to meet transparency requirements?

Data residency and the Fast tier

• OpenAI states that the Fast pricing tier (2x standard, lower latency) is unavailable with EU data residency enabled. If your compliance requirements mandate EU data residency, you cannot use Fast mode. Confirm which tier your account is on before committing to latency-sensitive workflows.

Security posture

• Astra is the first model to receive a CRITICAL cybersecurity classification. Even if you never use the model, the capability now exists in the world. Is your patching cadence, your exposure surface, and your incident response plan calibrated for a world where finding a new vulnerability in your systems no longer requires a skilled human spending weeks? OpenAI's safety report is the source.
• An agent that operates software on your behalf should not hold the keys to everything. Define the narrowest access that lets the agent complete its task, and audit what it touched after each run.

Bottom line

Something real happened on 3 September 2026. Software that operates a computer on your behalf, for hours, is a different category of tool from one that answers questions. The benchmarks are impressive, the safety improvements are real, and the price is not trivial. But the company that saturates puzzle exams is the same company that left the real-world-work exam off the slide. The model whose security capabilities earned a CRITICAL classification is the same model whose internal reasoning has become harder to follow. And the review that found nothing to change is a review nobody outside the room can see.

The sane response is neither excitement nor dismissal. Pick three tasks, run a 30-day pilot, measure what it costs to finish a job properly, and decide from that. That answer will be worth more than every benchmark in this article, because it will be about your work, on your systems, with your data, under your rules.

What this guide did not check: we could not re-verify Geektime's reporting (the page returned a 403 error). We could not independently verify the Greg Brockman AGI quote, the July Hugging Face incident, or the August development pause — these are reported by secondary sources, not confirmed by primary documentation. GDPval scores for Astra have not been published by OpenAI. Every benchmark in this guide is a vendor claim until independently replicated.

FAQ

Official sources

• developers.openai.comOpenAI: GPT-6 Astra model documentation — capabilities, pricing and availability
• deploymentsafety.openai.comOpenAI: GPT-6 Astra deployment safety report — cybersecurity classification, prompt injection, monitorability
• ai-pashut.co.ilai-pashut: GPT-6 Astra launch coverage and benchmark summary (Hebrew)
• gadgety.co.ilgadgety: GPT-6 Astra SRE-Bench and safety refusal scores (Hebrew)
• geektime.co.ilGeektime: OpenAI unveils GPT-6 Astra — ARC-AGI-3 score, GDPval omission, AGI framing (Hebrew; returned 403 on re-check)
• openai.comOpenAI: "GPT-6 Astra: A new generation of intelligence" — launch announcement (3 September 2026)