A context window of 1,050,000 tokens, priced at $10 per million input tokens and $50 per million output tokens: those are the official numbers for GPT-6 Astra, the model OpenAI announced on Thursday, September 3, 2026, and opened to API access on Friday, September 4. The company describes it as its "flagship model for demanding end-to-end work" and "a new frontier on computer and browser use" — the first sixth-generation release to reach developers with a context window of this scale.
This explainer gathers everything officially documented about the model: the exact specifications, pricing across all serving tiers, how to access it from anywhere in the world, the working demos OpenAI itself has published, and why the launch carries a safety debate that is far from over. It was written two days after launch, so expect a complete explainer rather than a breaking-news flash.
Source: Official model page on OpenAI's platform
Spec Card: The Official Numbers
| Item | Official value |
|---|---|
| Model name | GPT-6 Astra |
| Launch | Announced September 3, 2026; API availability September 4, 2026 |
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Input pricing (standard tier) | $10 per million tokens |
| Output pricing (standard tier) | $50 per million tokens |
| Supported inputs | Text, images, files |
| Outputs | Text |
| Announced providers | OpenAI and Azure (US) with automatic failover |
| Stated use cases | Advanced analysis, software engineering, deep research, scientific work, document creation, long-horizon agentic tasks |
This card is drawn from the model's official documentation page and published price lists — every link is in the sources section at the end. Anything not in this table is unconfirmed, and we flag it as such.
What Is GPT-6 Astra and Why Does It Matter?
Astra is the newest member of the GPT family, and OpenAI has declared it its latest and most powerful model yet — a claim that belongs to the company, not to neutral benchmarking, since independent evaluations are still in their early days. What genuinely distinguishes it in the official documentation is the explicit targeting of "long-horizon agentic tasks" involving computer and browser use: assigning the model an entire chain of steps (opening pages, filling forms, operating tools, compiling results) instead of answering one question and stopping.
The model also supports function calling and JSON-schema structured outputs — the two hooks developers use to wire models into their own systems. In practice, Astra is designed as the engine of a "digital worker" that completes a job end to end, not merely a chat engine.
The second striking number is the context window: 1,050,000 tokens is enough for a large stack of documents, a mid-sized codebase, or hours of conversation history in a single session. That alone changes how applications get built, since developers previously had to split work into batches to fit smaller context limits.
Agents in the Field: Two Official Demos
OpenAI's developer-blog team published two working showcases that illustrate the model's end-to-end style:
- Architectural visualization: a designer used Astra to design a house, refine its everyday details, and then explore the result through Blender camera tours and an Unreal Engine 5 walkthrough — from concept to a walkable 3D model.
- Game building: a second walkthrough documents building a procedural space-exploration game inside Codex, from concept art to planets you can land on, through extended direction rather than line-by-line coding.
Source: OpenAI developer blog
Source: OpenAI developer blog
The message from both demos is deliberate: the company wants developers to see the model's value in "end-to-end work" — the exact phrase its official page uses.
Pricing in Detail: What Will You Actually Pay?
The published standard tier is $10 per million input tokens and $50 per million output tokens, with cached-input reads at $1 per million tokens, cache writes at $12.50 per million, and web search at $10 per 1,000 calls. Aggregator listings show two additional tiers: a cheaper Flex tier at $5 input / $25 output, and a faster Fast tier at $20 input / $100 output per million tokens.
Let us convert those abstract rates into real project costs — the part most coverage skips:
| Workload | Standard tier | Flex tier | Fast tier |
|---|---|---|---|
| Ten-document analysis (100K input + 5K output tokens) | ~$1.25 | ~$0.63 | ~$2.50 |
| Mid-sized coding project (500K input + 30K output) | ~$20 | ~$10 | ~$40 |
| Intensive research agent (2M input + 100K output) | ~$70 | ~$35 | ~$140 |
| Same project with 75% cache-effective input (standard) | ~$47.50 | not published for this tier | not published for this tier |
The figures are approximate and derived from the announced prices, but they surface two golden rules: Flex roughly halves the bill when top speed is unnecessary, and caching becomes your best friend in agentic projects that repeatedly re-read the same instructions and context.
How Do You Access GPT-6 Astra?
Three documented routes exist today, all aimed at developers and businesses rather than everyday chat users:
- OpenAI's API directly: through a platform account and API keys — the primary official route. Check availability and billing terms for your country before committing, since regional policies differ and change.
- Azure: the model is listed among the announced US providers via Azure, the route enterprises prefer for contractual billing, data commitments, and compliance.
- Aggregators such as OpenRouter: these expose the model on the Flex and Fast tiers mentioned above with one unified bill across many models — a comfortable entry point for indie developers comparing several models before committing.
What has not been officially announced at the time of writing is availability inside the consumer ChatGPT interface. Do not act on any claim that it has "landed in ChatGPT" ahead of an official statement. If you want economical day-to-day access to the GPT family instead, our guide on the shared ChatGPT Plus option remains the practical reference, and our Claude vs ChatGPT vs Gemini comparison helps you match the tool to the job.
Why Is the Launch Controversial?
The word "controversial" in the coverage was not journalistic flavor. One day before launch, OpenAI published its documented Path to Astra report, acknowledging that the model reached what the company itself calls a "critical threshold" in cyber capabilities, alongside a technical debate about recurrent depth and named safety-expert responses. We covered that file in detail in two pieces: the Path to Astra report explained and Astra hitting the critical cybersecurity threshold — the full background for anyone wondering why safety researchers watched this launch more closely than developers did.
The irony worth sitting with: the model promising the strongest agentic computer-and-browser capabilities is the same one that triggered the loudest debate about governing those capabilities. That is not a reason to avoid it; it is a reason to build with a least-privilege mindset — give the agent the minimum permissions possible, log every step, and keep sensitive data out of its paths.
Quick Comparison: Where Does It Sit Within the GPT Family?
| Criterion | GPT-6 Astra | Previous consumer-available generation (e.g., GPT-5.6 Sol and Luna) |
|---|---|---|
| Primary audience | Developers and companies via the API | Everyday ChatGPT users |
| Announced context window | 1,050,000 tokens | Much smaller per official pages |
| Focus | End-to-end agentic work on computer and browser | Chat and text, with Luna's free unlimited text as we covered earlier |
| Pricing model | Per-token API pricing with Flex and Fast tiers | Consumer monthly subscriptions |
| Official documentation | Full spec and pricing page published | Collected in our earlier coverage |
The table declares no winner — they serve different purposes. Anyone building a software product looks at Astra; anyone using AI for daily writing and study gets more value from consumer plans at a fraction of the cost.
What We Do Not Know Yet
Professional honesty requires fixing the boundaries of current knowledge: there is no official announcement about when Astra enters ChatGPT or consumer plans, nor about any regional pricing. Independent head-to-head benchmarks against Claude Fable 5.1 or Gemini 3.8 remain preliminary, and we will not publish numbers from them before they mature — see our coverage of Fable 5.1 and Mythos 5.1 and Gemini 3.8 Flash for the current competitive landscape. Finally, prices can change; the governing number is always the official pricing page.
What Does This Mean for You?
If you are a developer or a small team anywhere from Lagos to Lahore to São Paulo: the giant context window removes one of the biggest obstacles to building knowledge assistants over entire document archives, and the Flex tier makes experimentation possible on a small budget — start with one narrowly scoped project this week and evaluate on your own data. If you are a student or researcher: long-horizon agents mean an assistant that gathers references and prepares drafts in a single session, but human review of the output remains a condition of academic quality. If you manage an organization: the three questions before any adoption are where your data is processed, who owns the audit trail of the agent's steps, and which permissions it holds — the groundwork for that mindset is in our free AI tools on Truescho.
Frequently Asked Questions
How much does GPT-6 Astra cost?
The announced standard price is $10 per million input tokens and $50 per million output tokens, with a cheaper Flex tier at $5/$25 and a faster Fast tier at $20/$100. Separate rates exist for cached reads and web-search calls, detailed on the official pricing page, and we have computed worked cost examples in a table inside this article.
Do I need a ChatGPT subscription to use GPT-6 Astra?
No. Astra is available primarily through the API for developers and companies at the usage prices described. OpenAI has not announced its arrival in consumer ChatGPT plans as of publication. Everyday ChatGPT users need to do nothing right now beyond following official announcements.
How large is the context window?
The official context window is 1,050,000 tokens with a maximum single-response output of 128,000 tokens. Practically, that allows analyzing large document bundles or entire code projects in one session without splitting the work — one of the model's headline selling points.
How do I use Astra from outside the US?
The three documented routes are a developer account on OpenAI's platform with API keys, enterprise access through Azure, or aggregators like OpenRouter with a unified bill. Review the official availability and billing terms for your country, since regional policies change, and avoid unofficial channels selling "discounted access" — they are a data risk.
Why is the launch described as controversial?
Because, according to OpenAI's own pre-launch Path to Astra report, the model reached a "critical threshold" in cyber capabilities, sparking a documented debate among safety experts about governing agentic computer-and-browser skills. We detailed the full story in our two reports; the balanced summary is that the capability is real and documented, and the open question is the engineering of its guardrails, not its existence.
Our Verdict: Is It Worth Trying?
If you build applications, agents, or need wide-scale document analysis, the answer is practical, not promotional: yes — trial it this week on the Flex tier with one small, well-scoped project, measure results on your own data before expanding, and keep a running price comparison. If your use is purely conversational, waiting for the model to reach ordinary consumer plans is cheaper and smarter than paying API costs you will not exploit. Either way, a safe starting point lives in our free tools on Truescho, and we will keep this article updated the moment any official announcement lands.
Sources: Official GPT-6 Astra model page on OpenAI's platform, OpenAI's developer blog showcases, OpenRouter, and TechCrunch coverage, September 3, 2026.