OpenAI’s new flagship lists at $10 for input and $50 for output per million tokens. The previous flagship lists at $5 and $30. Whether the extra money buys more depends on your tasks, and only a test of your own will show that.
- Astra’s API price is 2x the previous flagship on input and about 1.67x on output.
- The benchmark gains come from OpenAI. They do not tell you what a task in your business will cost.
- A 20-task test on your own work takes an afternoon and settles the question.
GPT-6 Astra pricing is the first thing a buyer should read, and it is not on the launch page’s headline. OpenAI released the model on 3 September to a limited set of organizations. The headline claims are about capability. The price sits one click down, and it is higher than what most teams pay today.
TL;DR
- Verified fact: OpenAI’s model page lists Astra at $10 per million input tokens and $50 per million output tokens, with cached input at $1.
- Verified fact: the previous flagship, GPT-5.6 Sol, lists at $5 and $30 on the API pricing page.
- Attributed claim: every benchmark gain, including 72.6% on OSWorld 2.0 and 57.9% on Terminal-Bench 4.0, comes from OpenAI itself.
- Analysis: on a typical mix of input and output, Astra must finish the same job with about 55% of the tokens to cost the same as Sol.
- Action: run 20 of your own tasks through both models and compare cost per finished job, not cost per token.
What does GPT-6 Astra cost compared with its predecessor?
OpenAI lists Astra at $10 per million input tokens and $50 per million output tokens. GPT-5.6 Sol lists at $5 and $30. That makes Astra 2x on input and about 1.67x on output. Batch and flex processing run at half the standard rate. The context window is about 1.05 million tokens, with up to 128,000 tokens of output.
ChatGPT access covers Plus, Pro, Business and Enterprise, with an extra Astra Pro tier on Pro. OpenAI’s release notes describe a limited rollout at launch, and a separate OpenAI page says Enterprise access is off by default, so an administrator has to switch it on.
What does the price change mean per task?
Take an invented but typical job, 20,000 tokens in and 4,000 tokens out. On Sol that costs $0.22. On Astra it costs $0.40. Run 1,000 of those a month and the bill moves from $220 to $400. This is an illustration built from list prices, not a measured result.
The break-even point follows from the same sum. Astra has to complete the job with about 55% of Sol’s tokens to cost the same. OpenAI says Astra is 47% faster at computer use. Faster is not the same as fewer tokens, and nothing checked here shows that Astra uses fewer.
The Limitation
Almost every number that makes Astra look strong comes from OpenAI. The company reports 72.6% on OSWorld 2.0 against 65.7% for Sol, and 57.9% on Terminal-Bench 4.0 against 37.3%. Those are vendor claims. Several of the other headline scores, including the ARC-AGI-3 results, are reported with different testing setups, so they cannot be compared with each other or with a rival’s.
The launch language is also stronger than the evidence. OpenAI’s own page calls the model a generational leap that gets closer to AGI. No agreed definition of AGI exists, and a benchmark does not measure the cost of a late invoice or a wrong customer email. The evidence supports one narrow claim: on OpenAI’s own tests of agent tasks, Astra scores higher than Sol. It does not show that your support queue, your quoting or your reporting gets better.
What should a Canadian buyer check before switching?
Run a test before changing a default, and ask about residency in writing. OpenAI documents Canada as a data-residency region for ChatGPT Enterprise at no extra cost, but the help page does not mention Astra. No source checked here confirms Canadian availability for Astra, or any price in Canadian dollars.
- Pick 20 real tasks from last month, such as a quote draft, a contract summary or a support reply. Remove client names first.
- Run each task on Sol and on Astra and record input tokens, output tokens and cost.
- Have the person who owns the work score each result as usable, fixable or wrong.
- Compare cost per usable result. Ignore cost per token.
- Ask your OpenAI contact whether Astra falls under your Canadian residency terms, and get the answer by email.
Our earlier pieces on what AI really costs a Canadian business and on deployment spending cover the other costs that sit around the model fee.
What does the sceptic say?
The best case for Astra is that cost per token is the wrong unit. If a model solves 57.9% of hard coding tasks where its predecessor solves 37.3%, as OpenAI reports, the cost per solved task could fall even as the price per token rises. For a team running coding or computer-use agents, that arithmetic could favour Astra.
That case holds only if OpenAI’s numbers survive independent testing, and the test above is how an individual company finds out in a day. The view here is narrow. Test first, because a price rise this size is justified only by results on your own work.
What to watch
- Astra appearing on OpenAI’s pricing page with a full price table, including any fast mode or long-context rates.
- Independent evaluations that use the same testing setup for Astra and its rivals.
- A written statement on whether Astra falls under Canadian data residency.
- A deprecation notice for Sol. None was found, and OpenAI says it gives at least six months for generally available models.
Frequently asked questions
How much does GPT-6 Astra cost?
OpenAI lists the API price at $10 per million input tokens and $50 per million output tokens, with cached input at $1. Batch and flex processing are half that. These are US dollar list prices from OpenAI’s model page, and no Canadian dollar price was found.
Is GPT-6 Astra available to Canadian businesses?
OpenAI lists availability on ChatGPT Plus, Pro, Business and Enterprise and on the API. No page checked says Canada is excluded, and none says Astra is covered by Canadian data residency. Ask your OpenAI contact and keep the answer in writing.
Should my team switch to Astra now?
Not on launch claims alone. Astra costs more per token, and the performance numbers come from OpenAI. Run 20 of your own tasks on both models and switch only if cost per usable result falls.
The decision in one line
Switch when 20 of your own tasks cost less per usable result, because a price rise this size has to be earned on your work and not on a launch page.
Written by Priya Chen, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Archive entry dated 7 September 2026, written and fact-checked on 7 October 2026. Sources are linked on the claims they support.
