GPT-6 Astra release: price, availability, and what it can do

OpenAI's newest model runs $10 per million input tokens and $50 per million output tokens on the API, costs nothing extra inside ChatGPT Plus ($20/mo) or Pro ($200/mo), enforces weekly message caps in regular chat, and has no free tier at all. Figures as of September 5, 2026.

Most launch coverage answers "what can it do" and skips the question I'd actually ask before paying anyone: on the plan I already have, how many times can I run this thing this week? That number is where the story gets uncomfortable.

GPT-6 Astra release: price, availability, and what it can do — $10 input and $50 output per million tokens, launched September 3, 2026 (AI-generated image)

How much does GPT-6 Astra cost?

GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens on the OpenAI API as of September 5, 2026. Cached input drops to $1 per million, cache writes run $12.50 per million, and Batch or Flex processing bills at half the standard rate. ChatGPT subscribers pay no separate model fee.

Put next to the rest of OpenAI's lineup, that's a steep step up.

Two footnotes will show up on your bill before you notice them. Astra's context window holds 1,050,000 tokens and takes up to 922,000 tokens of input, but any request over 272,000 tokens gets billed at the higher long-context rate — roughly $20 input and $75 output per million — for the entire request, not just the overflow. Feed it one oversized repository dump and the whole call reprices.

The second: a "Fast mode" doubles throughput for double the price, around $20 and $100 per million, according to CloudZero's pricing roundup. That figure comes from a single source, since OpenAI's own pricing page was blocked at the time of writing. The cheap direction is better documented — cached input at $1 per million, and Batch or Flex at half rate for work that can wait a few hours.

GPT-6 Astra release: price, availability, and what it can do on your plan

Astra reached Pro, Enterprise, and Business Premium users inside ChatGPT Work and Codex on September 4, 2026, along with the API. Plus and Business Standard users were told the rollout "might take a few days" longer. Regular chat access runs on weekly message caps, and the Free plan gets zero.

OpenAI's help documentation was unreachable while this piece was written, so the per-plan caps below come from two independent user captures that matched each other rather than from a published official table. Close, but treat them as unconfirmed.

Read the Plus row again. Zero. A $20 subscription buys you Astra in Work and Codex, not in the chat box you open every morning — and if I were paying for Plus specifically to try the new model, I'd want to know that before the charge, not after. Enterprise admins have a similar trap in reverse: Astra is switched off by default and someone has to enable it for the workspace.

What Astra does better than Sol

The gains sit in computer use, not conversation. On the OSWorld 2.0 benchmark, Astra scores 72.6% at about 40 minutes per task, against GPT-5.6 Sol's 65.7% at roughly 75 minutes — more accurate, in close to half the time, per figures reported by VentureBeat at launch.

OSWorld 2.0 benchmark: GPT-6 Astra 72.6% versus GPT-5.6 Sol 65.7% (AI-generated image)

What that translates to in practice: clicking through browsers, spreadsheets, websites and desktop apps; producing finished documents and presentations; and running multistep jobs like form filling, CRM record updates, calendar cleanup, online research with a drafted summary, data analysis with generated plots, and basic website building with frontend QA checks.

One early tester wrote that Astra finished a task a person needs about 30 minutes for in 5 minutes 27 seconds, and used it inside Blender to build a detailed 3D model of a house from one photo. That's a personal account rather than a measured benchmark, so file it as promising, not proven.

The part the price tag hides

OpenAI shipped Astra first to a limited set of organizations in its cybersecurity-focused Daybreak program on September 3, 2026, before the wider rollout — a detail CNBC covered on launch day. The staged start is a hint about who the model was built for.

There's a colder read too. One independent analysis, from eesel AI, argues that Astra's general intelligence lands roughly where GPT-5.6 Sol already was despite the 2.5x price premium, with the improvement concentrated in agentic and computer-use work, and that the model's monitorability — how well OpenAI can observe its reasoning — has gone down compared with Sol. Single source, and it's an opinion piece, so weigh it accordingly.

My own read of the pricing table is simpler. If your work is short questions and drafting, Sol at $4 and $20 does that for 40% of the money until the promotional rate expires on November 21, 2026. If your work is a long agent run that clicks through twelve screens without you watching, the price gap starts making sense. The full rate card sits in OpenAI's API docs.

At a glance

That's the GPT-6 Astra release: price, availability, and what it can do, as of September 5, 2026 — and the weekly cap, not the sticker price, is the number worth checking before you upgrade.