GPT-6 Astra (2026): What It Costs, Who Actually Gets It, and GPT-6 vs Claude Opus 5 for a Business

GPT-6 Astra: the short answer

As of September 2026, GPT-6 Astra is OpenAI's top model, and it costs $10 per million input tokens and $50 per million output on the API: exactly twice Claude Opus 5, and the same as Claude Fable 5.1. It has a 1,050,000-token window, a price line at 272K input tokens that reprices the whole request, and on ChatGPT Plus it runs 5 to 45 Codex messages per five hours. Every new flagship launch produces the same two weeks of benchmark screenshots and "this changes everything" threads. It changes one thing for most businesses: the price of the hardest ten percent of your AI work. This page is the price, the catch, who actually gets it, and what an owner-led company should do, with every number read from OpenAI's and Anthropic's own pages on September 11, 2026.

What GPT-6 Astra costs on the API

From OpenAI's GPT-6 Astra model page: $10 per million input tokens, $1 per million cached input tokens, $12.50 per million cache writes, and $50 per million output tokens. Batch and Flex processing are priced at 50% of those rates ($5 in, $25 out), and Fast mode is priced at 2x. The model takes text and images, returns text, supports reasoning effort from low to max, and lists web search, file search, code interpreter, computer use, and MCP among its tools. Knowledge cutoff: April 30, 2026.

For context, the rest of OpenAI's current lineup on the same pricing page: GPT-5.6 Sol at $4 / $20, GPT-5.6 Terra at $2 / $12, and GPT-5.6 Luna at $0.20 / $1.20. Astra is 2.5x Sol and 50x Luna on input. The full Claude-versus-OpenAI API comparison, beyond the flagships, is on Claude API vs OpenAI API.

The 272K line: the math OpenAI's pricing page does not do for you

The window is 1,050,000 tokens (922,000 input, 128,000 output). The line that matters sits much lower. OpenAI's words: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request." Not the tokens over the line. The full request.

Run one document review both ways. A 270,000-token contract bundle with a 10,000-token answer: 270K at $10 is $2.70, plus 10K at $50 is $0.50, so $3.20. Add 30,000 tokens of exhibits and the same review is 300K at $20 ($6.00) plus 10K at $75 ($0.75), so $6.75. Eleven percent more input, a bill more than twice as large. The same 270K job on Claude Opus 5 at Anthropic's listed $5 / $25 is $1.60.

If you build anything that stuffs whole folders into a prompt, the first optimization is not the model. It is keeping each request under 272K, which usually means sending the three relevant documents instead of the thirty.

GPT-6 in ChatGPT and Codex: who actually gets it

OpenAI announced Astra on September 3, 2026, as reported by launch coverage, and rolled it out in stages: a limited set of organizations first, then ChatGPT Plus, Pro, Business, and Enterprise, plus the API. OpenAI's Codex pricing page is the clearest first-party map of who gets how much. Its estimates for GPT-6 Astra, in local messages per five-hour window: Plus ($20) 5 to 45, Pro 5x ($100) 25 to 225, Pro 20x ($200) 100 to 900, and Business the same as Plus. The Free and Go plans have no Astra row. Those are estimates, not caps; a big multi-file task uses far more of the window than a quick question, and weekly limits may apply on top. The mechanics of that window are on Codex usage limits, and the plan ladder itself is on Codex pricing.

Past the included limits, Codex bills in credits, and Astra is the expensive one: 250 credits per million input tokens and 1,250 per million output, with Fast mode in Codex at 2.5x Astra's standard rate. Five to 45 messages is a taste, not a workday. On Plus, Astra is the model you save for the hard problem and not the one you leave selected.

GPT-6 Astra vs GPT-5.6 Sol, Terra, and Luna

• Astra ($10 / $50): the hardest end-to-end work, in OpenAI's own description. Long research, complex code changes, computer use.
• Sol ($4 / $20): OpenAI's pick for complex reasoning and high-stakes decisions, and the model Codex cloud chats run on.
• Terra ($2 / $12): the everyday workhorse: production tasks, reporting, document analysis.
• Luna ($0.20 / $1.20): fast, high-volume work such as routing, classification, extraction, and support.

The practical split I set up: Luna or Terra for anything that runs a thousand times a day, Sol for the judgment calls, Astra only when a wrong answer costs more than the tokens.

GPT-6 vs Claude Opus 5 (and Fable 5.1) on price

Anthropic's pricing page on September 11, 2026: Claude Opus 5 at $5 / $25, Claude Sonnet 5 at $2 / $10, Claude Haiku 4.5 at $1 / $5, and Claude Fable 5.1 at $10 / $50, with cache hits at a tenth of the input rate. So at the top, Astra matches Fable token for token and costs twice Opus 5. One level down, Sonnet 5 and Terra are both $2 in, with Terra's output $2 higher. At the bottom, OpenAI wins outright: Luna is a fifth of Haiku 4.5's price. The per-model Anthropic rates, caching, and batch discounts are broken down on Anthropic API pricing.

Which one is better is a question for your own work, not a launch-week benchmark. Take five real tasks, run them on Astra and Opus 5, and compare the answers and the bills. If the flagship difference does not show up in your five tasks, the only place it will show up is the invoice.

Is GPT-6 free?

No. On the API every token is billed, and on OpenAI's own Codex table Astra starts at Plus. The free and $8 Go plans run the GPT-5.6 family. That is fine: the free tier's job is to get you to Plus, and Astra's job is to get Plus users to Pro.

What a business should actually do

Do not switch assistants because a new flagship launched. If your team is on ChatGPT Plus or Business, Astra shows up inside the plan you already pay for; tell people it exists and that it is for the hard problems. If your team is on Claude, Opus 5 does the same job at half the API price, and nothing about this launch is a reason to migrate a working setup. If you build on the API, keep requests under 272K, route volume to the cheap models, and give the flagship the few jobs that earn it. And if what you want is the assistant, the connectors, and the first three workflows installed rather than benchmarked, the Claude for Small Business install does that in about two weeks, from $4,500; the $397 Playbook is the do-it-yourself version. The model is the part that changes every quarter. The workflow is the part that pays.