Reviewed by Jonathan West · Updated Sep 3, 2026

ChatGPT 6 Features

ChatGPT 6 shipped as GPT-6 Astra, and the specs are now published.

Reviewed by Jonathan West · Updated Sep 3, 2026

ChatGPT 6 shipped on September 3, 2026 as GPT-6 Astra. Plus, Pro, Business and Enterprise accounts get it over the days that follow, and developers reach it through the OpenAI API and AWS.

The budget numbers are published. Astra costs $10 per million input tokens and $50 per million output, so writing a long document costs five times what reading one does. It holds 1,050,000 tokens of context, writes up to 128,000 tokens in a call, and knows nothing after April 30, 2026.

Below we separate what OpenAI stated from what pre-launch coverage assumed. For the full capability picture, visit the OpenAI Astra explained hub.


Is GPT-6 Out?

Yes. OpenAI released GPT-6 Astra on September 3, 2026 to a limited set of organisations, with the general release on September 4.

It succeeds GPT-5.6, which reached general availability in July 2026 in three tiers: Sol, Terra, and Luna. Those tiers are still sold, and they remain the cheaper option for routine work.

The name settled at launch. Astra was OpenAI's designation for the generation, GPT-6 is the number it shipped under, and the GPT-5.7 label raised in pre-launch reporting was never used.

  • GPT-6 Astra is out, so there is a version to try and a system card to read.
  • The API id is gpt-6-astra, which is the string an integration has to match.
  • It reaches ChatGPT Plus, Pro, Business and Enterprise, so most paying seats get it without a new contract.
  • It costs $10 per million input tokens and $50 per million output tokens, which is roughly double GPT-5.6 Sol on input.
GPT-6 Astra is the model the ChatGPT 6 searches were looking for. It shipped on September 3, 2026, and the GPT-5.6 tiers stay on sale beneath it.

ChatGPT 6 has arrived, so the decision is now which work moves and which stays on GPT-5.6. We can scope that against your actual workflows rather than a benchmark table.

Book a Consultation

How the Naming Resolved

OpenAI shipped the model as GPT-6 Astra, keeping both the generation number and the designation it had used since August. That settles a naming question that ran for five weeks and produced most of the conflicting coverage.

GPT-5.7 was the alternative label reporting had raised, and it would have placed the model inside the existing GPT-5 series rather than opening a new one. A new generation number is what lets OpenAI retire GPT-5 models on their own schedule and price GPT-6 separately, so the label decided whether Astra inherited the GPT-5 line's deprecation dates or started its own.

Other codenames circulated in Google search suggestions, including Doug. OpenAI used none of them, and anything still describing the model under one of those names predates the launch.

  • GPT-6 Astra — the shipped product name, combining the generation number and OpenAI's own designation.
  • gpt-6-astra — the API model id, which is what code and vendor integrations have to reference.
  • GPT-5.7 — never used. Anything describing Astra as an addition to the GPT-5 series is out of date.
  • Sol, Terra, and Luna are real names, but they belong to the GPT-5.6 tiers that Astra succeeds.
  • Doug and similar codenames were never used by OpenAI in a publication.

What Is Sourced and What Is Rumour

Four claims about the next OpenAI model can be traced to OpenAI or to an outlet that spoke to OpenAI. Everything else circulating is inference.

The sourced set is narrow and it is all capability and safety, with nothing about the product. There is a research result, a risk classification, a training pause, and a statement about limited access at launch.

The rumoured set holds the product details, which is the opposite of what a buyer needs. Prices, context windows, and dates all lack a source.

  • Sourced — an internal Astra version solved ten open problems in mathematics and theoretical computer science, published August 1, 2026.
  • Sourced — a safety evaluation on August 7, 2026 placed Astra at the Critical cybersecurity capability threshold.
  • Sourced — OpenAI stopped two weeks of deployment-focused reinforcement-learning training and held its largest planned frontier run (Axios, August 18, 2026).
  • Sourced — on September 2, 2026 OpenAI said advanced capabilities will go to selected early testers rather than to everyone, and the September 3 launch followed that shape.
  • Sourced — at launch OpenAI published a price of $10 and $50 per million input and output tokens, a 1,050,000-token context window, a 128,000-token output ceiling, and an April 30, 2026 knowledge cutoff.
  • Sourced — OpenAI reports FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%.
The four fields every pre-launch ChatGPT 6 spec sheet guessed at are now published: price, context window, output limit, and date.

What OpenAI Says GPT-6 Astra Does

GPT-6 Astra is aimed at work that runs across many steps. OpenAI says it beats its earlier models on computer use, browsing, software engineering, cybersecurity, science and professional work, and that it holds up better over long workflows. The practical effect is fewer handoffs back to a person mid-job.

The published scores are FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%. Those describe mathematics, abstract reasoning and exploit challenges rather than your documents, so they set an upper bound on capability rather than a prediction about your workload.

One capability delayed the launch by five weeks. OpenAI's own evaluation put Astra ahead of GPT-5.6 Sol at finding vulnerabilities and writing exploits, on fewer tokens. That result earned Astra the Critical cybersecurity rating, and both shipped versions now hold back their strongest cybersecurity capabilities, so a security team evaluating Astra is testing a deliberately limited model.


ChatGPT 6 Price and Access

GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens through the OpenAI API. There is no separate ChatGPT 6 subscription tier, because the model arrives inside the existing Plus, Pro, Business and Enterprise plans rather than as a new product.

Access rolled out in order rather than all at once. A limited set of organisations got it on September 3, the general release followed on September 4, and business early testers reached it first through OpenAI's Daybreak programme.

The GPT-5.6 tiers stay cheaper and stay on sale. Sol runs $5 per million input tokens and $30 per million output, Terra is $2 and $12, and Luna is $0.20 and $1.20 after the July 30, 2026 cuts, so routine work that already runs on Luna has no cost reason to move. Current figures are on OpenAI's API pricing page.

  • $10 per million input tokens and $50 per million output tokens through the API, which is the figure a team plans against.
  • No separate ChatGPT 6 subscription, since the model ships inside Plus, Pro, Business and Enterprise.
  • Available on AWS as well as the OpenAI API, so an existing AWS commitment can cover it.
  • GPT-5.6 Luna is 50 times cheaper on input, which is why tier routing decides most real bills.

How OpenAI Named Its Last Release

OpenAI's most recent naming pattern splits a generation into named tiers rather than shipping one model for every job. GPT-5.6 arrived as three of them: Sol, Terra, and Luna. Each targets different work at a different price.

Astra did not repeat that shape. It shipped as a single model rather than a tiered family, so the choice is between Astra and a GPT-5.6 tier rather than between Astra tiers, and that keeps routing decisions in the older generation for now.

GPT-5.6 shipped with a model card and per-token pricing on day one, and so did Astra. What differed was access, which arrived in stages rather than open to everyone at once.

  • GPT-5.6 arrived as three named tiers. GPT-6 Astra arrived as one model, so cost routing still runs through the GPT-5.6 line.
  • GPT-5.6 tier pricing spans a 25-fold range on input tokens, which is why most bills are decided by which tier a request lands on.
  • GPT-5.6 launched with no waitlist. Astra launched to a limited set of organisations first and reached general release the next day.
  • OpenAI cut Luna by 80% and Terra by 20% on July 30, 2026, the same month the family reached general availability, and left Sol unchanged. Launch Astra rates may not be the lasting ones either.

Frequently Asked Questions

  • GPT-6 Astra holds a 1,050,000-token context window, so a full contract set goes in as one request, and writes up to 128,000 tokens in one call. Prompt caching bills repeated context at a lower rate, which matters when a long document is sent on every turn, and function calling and file search let it act on your systems rather than only describe them. OpenAI says it performs better than its earlier models on computer use, browsing, software engineering, cybersecurity, science and professional work.
  • Yes. OpenAI released it as GPT-6 Astra on September 3, 2026 to a limited set of organisations, with the general release on September 4. It reaches ChatGPT Plus, Pro, Business and Enterprise accounts over the following days.
  • Yes. OpenAI shipped it as GPT-6 Astra, keeping both the generation number and its own designation. The GPT-5.7 label that reporting had also raised was never used, and the API id is gpt-6-astra.
  • GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens through the OpenAI API. There is no separate ChatGPT 6 subscription, because the model arrives inside the existing Plus, Pro, Business and Enterprise plans.
  • OpenAI has not announced a free tier for GPT-6 Astra. The rollout covers ChatGPT Plus, Pro, Business and Enterprise, so a paid plan is what gets access, and the GPT-5.6 tiers remain the cheaper option through the API.
  • Doug is an unconfirmed codename that surfaced in Google search suggestions before the launch. OpenAI never used it in a publication. The model shipped as GPT-6 Astra.

Working out whether GPT-6 Astra is worth the switch?

We can map your workflow against GPT-6 Astra and the GPT-5.6 tiers, and show you which jobs justify the higher output rate and which are cheaper left alone.

Book a Consultation