GPT-6 Astra and the “AGI Era”: What OpenAI Actually Shipped, and What It Means for Agentic AI Developers

Short version: On 3 September 2026 OpenAI released GPT-6 Astra and framed it as the beginning of something it has been promising for a decade. President Greg Brockman closed the press briefing with four words — “Welcome to the AGI era” — and the company says Astra is “the most intelligent and aligned model in the world.” The benchmarks are genuinely striking in places and genuinely oversold in others. This post separates the two, then gets to the part that actually changes your week: the model is built to operate a computer and run multi-step work end to end, which moves the valuable skill from prompting to agent engineering. That is exactly what our WSQ – Agentic AI Applications with Codex course teaches, with fees funded up to 70%.

What OpenAI announced

Astra is positioned less as a better chatbot than as a better worker. OpenAI calls it “the world's best computer use model”: it navigates software the way a person does — across browsers, spreadsheets, websites and desktop applications — produces finished documents and presentations rather than generic first drafts, and carries multi-step workflows to completion. The pitch is expense forms, CRM updates, financial models and engineering design, not clever paragraphs.

Three product details matter more than the marketing:

• It decides when to ask. OpenAI describes the model as choosing “when to ask you a question and when to proceed on sensible assumptions.” Anyone who has watched an agent stall on a trivial ambiguity, or barrel through a serious one, will recognise why this is the hard part.
• It keeps notes instead of compressing. In Codex, Astra maintains searchable notes across context windows rather than summarising history away. That is an architectural answer to long-running agent tasks, not a bigger context window.
• It is the first model OpenAI has rated “critical” on cybersecurity. Astra can potentially find and exploit previously unknown vulnerabilities in well-protected systems, so OpenAI deliberately built it to refuse advanced offensive-security requests and gated the strongest capability behind its Daybreak Access programme.

Availability and price

Astra rolled out first to a limited set of organisations under Daybreak Access, then to ChatGPT Plus, Pro, Business and Enterprise, the API (gpt-6-astra) and Amazon Bedrock. API pricing is USD $10 per million input tokens and $50 per million output tokens — a 2.5x increase over GPT-5.6 Sol's $4/$20. Brockman's defence of that jump is worth keeping: “What you actually want…is the price per task” rather than the price per token.

He has a point, and it is measurable. Artificial Analysis found Astra uses roughly one-third the tokens of GPT-5.6 Sol and about one-fifth of Claude Opus 5 on its Coding Agent Index, landing the same score for less than half the cost of Claude Fable 5. A more expensive model that thinks in fewer tokens can still be cheaper per finished job. It can also be more expensive: the same analysis puts Astra at roughly 75% more per task than its predecessor at max effort on the Intelligence Index. Which way it lands depends entirely on your workload.

The benchmarks, read honestly

Astra is state of the art on computer use, browsing, software engineering, cybersecurity, science and professional work. Here are the numbers that came out with the launch.

The OSWorld result is the one to sit with. The score moved from 65.7% to 72.6%, but the average time per task fell from about 75 minutes to about 40 — a 47% speed-up. For agentic work, wall-clock time per completed task is often the binding constraint, not the last few points of accuracy.

Three caveats the headline skips

Being useful about this launch means naming what is soft in it.

1. The ARC-AGI-3 number depends on the harness. The widely quoted ~99.9% comes from OpenAI's stateful adapter harness. Independent stateless API testing lands between roughly 17% and 63% depending on reasoning tier. That is not a rounding difference — it means a large part of the headline result belongs to the scaffolding around the model, not the model alone.
2. It is not uniformly ahead. Astra sits at 57.2% on Humanity's Last Exam with tools against Claude Fable 5.1's 65.0%, ties GPT-5.6 Sol on the Intelligence Index, and shows measurable regressions on GDPval-AA v2 and minor ones on SciCode and banking tasks. “Best in the world” is true on some axes and not others.
3. “AGI” is doing heavy lifting. OpenAI's own definition is “highly autonomous systems that outperform humans at most economically valuable work.” A model that scores 72.6% on desktop tasks is remarkable and is also failing more than a quarter of them. Brockman said Astra could eventually be seen as the arrival of AGI — a hedge that the headlines quietly dropped.

One improvement deserves more attention than it got: the measured hallucination rate fell from 92% to 51% on Artificial Analysis's evaluation while accuracy rose 4 points. That is still a coin flip on the hardest questions, which is precisely why verification loops belong in your agent design rather than in your hopes.

What this actually changes for developers

Strip the AGI framing away and a concrete shift remains. When a model can drive a browser, a terminal and a desktop reliably enough to finish a 40-minute task, the scarce skill stops being writing a good prompt and becomes designing a system an autonomous model can operate safely. Four things get more valuable, not less:

• Tool and context design. Astra's Codex notes-over-compression approach only pays off if the tools it calls are well-scoped and its context is curated. MCP has become the common way to expose tools to an agent without hard-wiring them to one vendor.
• Evaluation. A 51% hallucination rate on hard questions means you cannot ship on vibes. Task-level evals, golden datasets and regression suites are how you find out whether a model swap helped or quietly broke you.
• Guardrails and least privilege. The first model rated “critical” for cyber capability is also the first one you should not hand unrestricted credentials. Sandboxing, scoped tokens and human approval gates on irreversible actions are engineering requirements now.
• Cost architecture. At $10/$50 per million tokens, the difference between a well-designed agent loop and a naive one is a line item. Effort tiers, caching (Astra offers a 90% discount on cache reads) and knowing when a smaller model suffices are budget decisions made in code.

None of that is prompt engineering. It is software engineering with a probabilistic component, and it is learnable.

Where to build these skills

Our WSQ – Agentic AI Applications with Codex course (course code TGS-2023041081) is 2 days / 16 hours and covers exactly this ground: planning agentic applications with Codex, building reusable workflows with Skills and MCP tools, RAG and multi-agent development, and testing, evaluating and optimising what you build. You leave having built and evaluated an agent, not having watched a demo.

If your team works in a different assistant, the same engineering skills transfer — see WSQ – Agentic AI Applications with Claude Code. If your interest is operational rather than hands-on-code, WSQ – Agentic AI for Business Process Automation is the better fit. The full Agentic AI Series lists every course in the track.

The sober take

Whether 3 September 2026 turns out to be the start of the AGI era is a question for historians, and the people declaring it are the people selling it. What is not in dispute is narrower and more immediately useful: a frontier model can now operate a computer well enough, and fast enough, that the bottleneck has moved to the systems we build around it. Tools, evaluation, guardrails, cost. The organisations that get value out of Astra in the next year will be the ones with engineers who can build that scaffolding — not the ones with the best prompts.

Frequently asked questions

Sources: OpenAI — GPT-6 Astra: A new generation of intelligence; OpenAI Deployment Safety Hub — GPT-6 Astra System Card; Artificial Analysis — Benchmarking GPT-6 Astra; The New Stack — OpenAI launches GPT-6 Astra. Benchmark figures as published at launch, 3 September 2026.