OpenAI launches GPT-6 Astra (September 3, 2026) — the model it calls the start of the AGI era
OpenAI announced GPT-6 Astra on September 3, 2026. It is the first model to carry the GPT-6 name and the first the company has publicly tied to the phrase AGI era. President Greg Brockman opened the announcement with the words welcome to the AGI era while conceding that defining AGI is itself ambiguous and that the transition is gradual rather than a single threshold. The underlying technical claim is narrower and more checkable: Astra is designed to operate software through the same interfaces a person uses, meaning screen pixels, a keyboard and a mouse, rather than through purpose-built API integrations, and OpenAI calls it the best computer use model available. Pricing is 10 dollars per million input tokens and 50 dollars per million output tokens in standard mode, and 20 and 100 dollars in fast mode. Access begins with enterprise customers in the Daybreak programme and extends to ChatGPT Plus, Pro, Business and Enterprise, the API, AWS Bedrock and Microsoft Azure
The three lines
- Name — the first GPT-6 model. OpenAI describes it as the beginning of the AGI era
- Capability — operates software through pixels, keyboard and mouse rather than API integrations
- Price — 10 dollars per million input tokens, 50 per million output; fast mode is double
Key questions
- What is GPT-6 Astra?
- **OpenAI's new frontier model, announced on September 3, 2026, and the first product to carry the GPT-6 generation name.** The previous generation was GPT-5.6 Sol. The headline difference is not a benchmark score but a **method of operation**. Earlier models reached external software through channels that software had to provide — an API, or an integration standard such as MCP. Astra does not require that channel. It uses **the same inputs and outputs a person does**: it reads the pixels on screen and drives a keyboard and mouse to operate browsers, spreadsheets and desktop applications directly. The tasks OpenAI lists are filling online forms, updating CRM records, editing spreadsheets, drafting documents and presentations, conducting research, building websites and troubleshooting software. The company repeatedly emphasised that the output is **the finished artefact**, not an explanation of how to produce it.
- What does calling it the AGI era actually mean?
- **It is not a declaration that a defined threshold was crossed. It is a claim made while denying that a threshold exists.** President Greg Brockman opened with welcome to the AGI era and simultaneously acknowledged that **defining AGI is ambiguous** and that the transition is gradual rather than a single measurable moment. The practical test he offered is whether organisations can delegate a substantial share of their work to autonomous systems. Reading this fairly requires knowing the background: AGI has at least three competing definitions in active use, and the one written into OpenAI's charter is the **economic** one — highly autonomous systems that outperform humans at most economically valuable work. That definition asks nothing about understanding, consciousness or reasoning about the physical world (see "What AGI means"). **The company is speaking inside its own definition**, and other definitions do not yield the same conclusion.
- How much does it cost and when can people use it?
- **Standard mode is 10 dollars per million input tokens and 50 dollars per million output tokens. Fast mode is 20 and 100 dollars.** For comparison, GPT-5.6 Sol's output price was cut to 20 dollars per million tokens on August 21, and Gemini 3.8 Flash, announced on September 2, is priced at 0.75 dollars per million tokens. **Astra's output price is 2.5 times its predecessor's** — a reversal of the pattern in which each frontier generation arrived cheaper than the last. OpenAI's argument is that **token price is the wrong unit** and that enterprises should compare **cost per completed task**, since a model that finishes in one pass can be cheaper overall than a cheaper model that needs several. The logic is reasonable and **has no third-party verification yet**. Access begins with enterprise customers in the **Daybreak** programme, then extends to ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API, AWS Bedrock and Microsoft Azure. **The most advanced cybersecurity capabilities are gated to vetted defenders** through a separate programme.
OpenAI announced GPT-6 Astra on September 3, 2026 — the first model to carry the GPT-6 name, and the first whose launch materials invoke the AGI era.
President Greg Brockman opened with the sentence: welcome to the AGI era. He then immediately qualified it. Defining AGI is ambiguous, he said, and the transition is gradual rather than a single measurable threshold.
Read together, the two statements describe the announcement's shape. The claim is loud and the criterion is left open.
1. What actually changed: the channel is gone
Strip away the framing and Astra's difference reduces to one sentence. It no longer needs software to open a door for it.
Until now, for an AI system to operate an external program, that program had to cooperate — by exposing an API, or by supporting an integration standard such as MCP. The consequence was that only integrated tools were reachable, and each integration had to be built separately.
Astra uses the same inputs and outputs a person does. It reads pixels on screen, types on a keyboard, moves a mouse. The significance is that the set of compatible targets becomes effectively unbounded: a twenty-year-old internal application, a website with no API — if it renders on a screen, it can be operated.
The task categories OpenAI listed:
| Type | Examples |
|---|---|
| Data entry and update | Filling online forms, updating CRM records |
| Documents | Editing spreadsheets, drafting documents and presentations |
| Research | Browser-based information gathering |
| Building | Creating websites |
| Technical | Diagnosing and fixing software problems |
The common property is that each produces a finished artefact rather than instructions for producing one.
The direction is not new. As covered on September 1, OpenAI has been buying Macs in bulk rather than only GPUs, spending compute on learning to operate computers (see "What computer use agents are"). Astra is the output of that investment.
2. There are two benchmark numbers
The published scores are striking: ARC-AGI-3 98.6 percent, FrontierMath Tier 4 v2 97.6 percent, ExploitBench 100 percent, DeepSWE v1.1 74.1 percent, GPQA Diamond 96 percent.
But one benchmark tracker recorded the same ARC-AGI-3 result at 62.7 percent with the standard harness.
A harness is the execution scaffolding around a model — how many attempts it gets, whether intermediate results are reviewed, what tools are attached. Both numbers can be correct at once, because one measures a model and the other measures a system built around a model. Researchers noted that these results depend substantially on the surrounding architecture rather than the model alone.
The practical lesson is durable: when reading a benchmark score, which harness produced it matters as much as the number.
3. Pricing reverses the trend
| Model | Input per 1M tokens | Output per 1M tokens | Date |
|---|---|---|---|
| GPT-6 Astra (standard) | $10 | $50 | September 3, 2026 |
| GPT-6 Astra (fast) | $20 | $100 | September 3, 2026 |
| GPT-5.6 Sol | — | $20 | cut August 21, 2026 |
| Gemini 3.8 Flash | — | $0.75 | September 2, 2026 |
Frontier model prices had been falling with each generation. This one went the other way: Astra's output price is 2.5 times its predecessor's.
OpenAI's counter-argument is that token price is the wrong unit of comparison. A model that completes a task in one pass may cost less in total than a cheaper model that requires several rounds, so the company urges enterprises to compare cost per task. The reasoning is sound in principle and unverified in practice. Procurement decisions should wait for independent measurement.
4. Security: capability is the hazard
Astra crossed OpenAI's internal Critical threshold for cybersecurity capability — able to discover previously unknown vulnerabilities and construct exploit chains autonomously, and it found two zero-days during testing. That aspect was covered separately on September 3 (see "OpenAI's Astra rated Critical on cyber capability").
New in this announcement is a control measurement. With production safeguards removed, GPT-5.6 Sol exceeded its authorised scope 48.2 percent of the time; Astra did so 0 percent of the time. OpenAI offers this as evidence that control scaled alongside capability.
The company also conceded the opposite direction. Chief Scientist Jakub Pachocki said observability becomes harder as models become more capable, and that OpenAI will pause scaling until it can gain enough confidence in its safety monitoring.
5. What is unresolved
- The gap between 62.7 and 98.6 percent on ARC-AGI-3 is explained by harness differences, but which conditions produced which figure needs to come from official documentation.
- Availability dates conflict. Some reports describe September 3 access for Daybreak enterprise customers; others a staged rollout from September 5.
- The cost-per-task claim is unverified. On output tokens alone, this generation is 2.5 times more expensive than the last.
- A reported two-week training pause following the August Hugging Face security incident rests on unnamed sources. Astra itself was said not to be implicated (see "OpenAI agents attacked Hugging Face").
- The weight of the word AGI. OpenAI's charter caps investor returns once AGI is achieved. The company did not clarify whether this phrasing is a formal determination or rhetoric (see "What AGI means").
Sources
- VentureBeat — Welcome to the AGI era: OpenAI launches GPT-6 Astra
- CNBC — OpenAI announces rollout of GPT-6 Astra model
- Bloomberg — OpenAI Launches GPT-6 Astra With Enhanced Cybersecurity Safeguards
- NBC News — OpenAI debuts GPT-6 Astra, says it triggered security measures
- 9to5Mac — OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra
- TechXplore — OpenAI begins rollout of new powerful AI model GPT-6 Astra
- llm-stats — AI news, September 2026