What's New in GPT-6 Astra

What the AGI Era Claim Actually Means

When OpenAI released GPT-6 Astra on September 3, 2026, the announcement used the phrase "AGI era." That framing deserves unpacking before you build anything on top of this model.

The architectural change behind the claim is called recurrent depth. A standard Transformer model processes each token through its layers exactly once — a single forward pass from input to output. Recurrent depth changes that: the same layers run multiple times before a token is produced, letting the model iterate on its own reasoning internally before committing to an answer.

This isn't chain-of-thought reasoning, where the model writes out intermediate steps you can read. Recurrent depth works silently, inside the forward pass — which is partly why OpenAI acknowledges that Astra is less monitorable than prior models. You see the output; the iterative reasoning that produced it is not visible in the response.

The practical result is five controllable thinking tiers. You can dial reasoning intensity up for hard problems and down for fast, cheap answers — within the same model. That tunability is what makes Astra different from simply using a bigger model.

The Benchmarks Worth Knowing

Four benchmarks from the OpenAI release are worth understanding before you decide how to use Astra:

OSWorld 2.0 (computer use). Astra completed tasks at 72.6%, against 65.7% for GPT-5.6 Sol. The more significant number is speed: Astra averaged 40 minutes per task, compared with 75 minutes for Sol. For agent workflows involving real desktop or browser automation, that efficiency gap matters more than the raw completion rate.

SRE-Bench (agentic engineering). Astra solved 88% of tasks on the first attempt, compared with 55.9% for GPT-5.6 Sol. The gap is large enough to change what workflows are practical — tasks that required retries or human fallback under Sol become single-pass operations under Astra.

DeepSWE v1.1 (code editing). Astra scored 74.1% versus 72.7% for Sol — a narrower margin. On software engineering tasks specifically, the two models are much closer than the SRE-Bench gap suggests.

AutomationBench (end-to-end automation). The category where Astra shows the largest relative gain: 41.4% against a prior baseline of 18.1%. Automation workflows that weren't viable before become worth building now.

Who Gets Access and How

Astra is available across ChatGPT paid plans and through the API, but the level of access differs by tier:

• ChatGPT Plus — includes Astra in Work and Codex, within standard usage allowances.
• ChatGPT Pro ($200/month) — includes 200 Astra Pro messages per week.
• ChatGPT Pro ($100/month) — includes 50 Astra Pro messages per week.
• ChatGPT Business and Enterprise — includes Astra Pro with higher message allowances and data privacy controls.

There is no free-tier access to Astra. ChatGPT Free users remain on earlier models.

For developers, the API model ID is gpt-6-astra. Pricing is $10 per million input tokens and $50 per million output tokens — there is no free API tier. Azure OpenAI and AWS Bedrock are also supported for enterprise deployments.

One important constraint: Astra's knowledge cutoff is April 30, 2026. Events after that date are outside the model's training data. For tasks that rely on current information — recent software releases, live pricing, current events — you'll need to pass that context explicitly in the prompt or use tools.

How This Track Is Organized

This track runs in two parts, reflecting the two main contexts where you'll use Astra:

Part 1 (Lessons 1–8): ChatGPT features. These lessons cover what changed in the ChatGPT interface — the reasoning tier controls, Projects, Canvas, operator system prompts, and the computer-use features built on Astra's OSWorld performance. If you're a ChatGPT user who wants to use Astra effectively without touching the API, Part 1 is complete on its own.

Part 2 (Lessons 9–15): Responses API. These lessons cover building with Astra programmatically — the Responses API, tool calling, multi-turn agent loops, streaming, and structured outputs. If you're a developer building on top of Astra, Part 2 is where the practical implementation lives.

The two parts are independent. Skip ahead to Lesson 9 if you're here for the API. Come back to Part 1 if you find yourself wondering what the reasoning tiers actually look like in a real ChatGPT session.