What GPT-6 Astra Is [September 2026]: Which ChatGPT Plans Include It, API Pricing, and How to Split Work with GPT-5.6
GPT-6 Astra, announced by OpenAI on 3 September 2026 (US time), is a flagship tuned for computer use and long, multi-step work, true to the pitch that it "can do anything a person can do on a computer." It scores 72.6% on OSWorld 2.0, and time per task drops 47% from GPT-5.6 Sol's roughly 75 minutes to roughly 40. This article checks what changed against the benchmarks, then lays out where each ChatGPT plan can use it (Pro at ¥16,800 / ¥30,000, Business and Enterprise get "GPT-6 Pro" in Chat; Plus at ¥3,000 gets it through ChatGPT Work and Codex; Free and Go do not), the caps such as 50 or 200 messages a week, API pricing ($10 input / $50 output, 2.5x Sol) and how to split work with Sol, Terra and Luna, and the behaviour changes on switching (more clarifying questions, more sensitivity to skill files), based on OpenAI's announcement, Help Center and model page.
On 3 September 2026 (US time), OpenAI announced its new flagship model, GPT-6 Astra. True to the pitch that it "can do anything a person can do on a computer," it is a model tuned for computer use and long, multi-step work. Within days of the announcement it began rolling out to ChatGPT Plus, Pro, Business and Enterprise, and to the API.
The first thing people trip over when they try to use it is "where does my plan actually get it?" On Pro and above it appears in the Chat model picker as GPT-6 Pro. On Plus it does not appear in the chat screen at all; it is delivered through ChatGPT Work and Codex. And the API price is 2.5x that of GPT-5.6 Sol.
This article walks through what changed with GPT-6 Astra, where each ChatGPT plan can use it and with what limits, and how to split work between Astra and GPT-5.6 Sol, Terra and Luna on the API, based on OpenAI's announcement, Help Center and model page.
The short version
GPT-6 Astra is less "smarter answers" and more "work finishes faster." It scores 72.6% on the OSWorld 2.0 computer-use benchmark (GPT-5.6 Sol: 65.7%), and time per task drops 47%, from about 75 minutes to about 40. Its target is multi-step work such as form filling, CRM updates, research and document creation.
In ChatGPT, Pro (¥16,800 / ¥30,000 a month), Business and Enterprise get it in Chat as "GPT-6 Pro". Plus (¥3,000) gets it through ChatGPT Work and Codex, and Free and Go do not get it at all. Each plan has a cap, such as 50 or 200 messages a week.
The API costs $10 per million input tokens and $50 per million output tokens, 2.5x GPT-5.6 Sol. Check first whether Luna (classification and extraction), Terra (routine work) or Sol (judgement-heavy work) is enough, and reserve Astra for jobs that use several tools and have to be carried through to the end.
What changed: speed more than intelligence
The striking thing in OpenAI's announcement is not the scores but the time taken. Here are the main figures alongside GPT-5.6 Sol.
| Benchmark | What it measures | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|
| OSWorld 2.0 | Computer-use completion rate | 72.6% | 65.7% |
| OSWorld 2.0 time | Per task | ~40 min | ~75 min |
| AutomationBench | Business automation | 41.4% | 18.1% |
| Agents' Last Exam | Overall agent ability | 59.3% | 53.6% |
| Terminal-Bench 4.0 | Terminal work | 57.9% | 37.3% |
| FrontierMath Tier 4 | Research-level maths | 97.6% | 83.0% |
| BenchCAD | CAD design | 95.9% | 83.3% |
The OSWorld 2.0 completion rate is 7 points higher, but the time is 47% shorter. On Mind2Web, task completion is said to be 1.9x faster. The improvement is in the direction of "finish the same job within the time a person is willing to wait."
The use cases OpenAI cites are filling in web forms, updating customer records in a CRM, web research, drafting emails and documents, analysing scientific data, building websites, and installing, testing and troubleshooting software. None of these return a single answer; all of them drive a screen through many stages.
Note that "operating a computer" here is a different layer from a product such as Grok Bot, which signs in to your apps and stays resident. Astra is a model. It works as the "brain" of agents that run through ChatGPT Work, Codex or the API.
Its safety positioning also changed. Astra is the first model to reach the "Critical" level for cybersecurity capability under OpenAI's Preparedness Framework. At the same time, on the ExploitGym honeypot, which tests whether a model crosses lines it should not, the crossing rate was 0% (GPT-5.6 Sol: 48.2%). OpenAI's framing is "it can do more, but it does not do what it must not."
Which ChatGPT plans can use it
This is the most-asked question. OpenAI's announcement says "all Plus, Pro, Business and Enterprise users", but the Help Center shows that availability differs between Chat, ChatGPT Work and Codex. Here it is alongside the monthly price in Japan (tax included, individual plans).
| Plan | Monthly | Chat (regular chat) | ChatGPT Work and Codex | Limit |
|---|---|---|---|---|
| Free | ¥0 | Not included (GPT-5.6 Luna) | Not included | — |
| Go | ¥1,400 | Not included (GPT-5.6 Luna) | Not included | — |
| Plus | ¥3,000 | Not shown (GPT-5.6 Sol at Medium / High) | GPT-6 Astra, rolling out | Limited allowance; extra credits can be bought |
| Pro $100 | ¥16,800 | GPT-6 Pro | GPT-6 Astra | 50 messages a week shared between GPT-6 Pro and GPT-5.6 Sol Pro |
| Pro $200 | ¥30,000 | GPT-6 Pro | GPT-6 Astra | GPT-6 Pro 200 a week (Sol Pro 170 a day; 200 a day for both combined) |
| Business Standard | For organisations | GPT-6 Pro | GPT-6 Astra (limited allowance) | 15 a month shared between GPT-6 Pro and Sol Pro |
| Business Premium | For organisations | GPT-6 Pro | GPT-6 Astra (full existing allowance) | 50 a week shared between GPT-6 Pro and Sol Pro |
| Enterprise and Edu | For organisations | GPT-6 Pro | GPT-6 Astra | Set by the workspace admin; off by default for the first two weeks |
"GPT-6 Pro" is Pro reasoning powered by GPT-6 Astra. The announcement calls it "GPT-6 Astra Pro"; the Help Center and the model picker call it "GPT-6 Pro".
ChatGPT | Chat needs Pro or above; Plus gets it via Work and Codex
This is where Plus users get confused. The Plus column on the ChatGPT pricing page says "advanced reasoning models with GPT-6 Astra and GPT-5.6" and "full ChatGPT Work access". The Help Center, however, says the following.
- In Chat (regular chat), Plus can choose GPT-5.6 Sol at Medium or High only. Extra High and Pro are not available
- GPT-6 Astra for Plus is "rolling out" in ChatGPT Work and Codex. The allowance is limited, and you can buy extra credits if you need more
- It states plainly that "rollout is gradual, and availability can differ between Chat, Work and Codex"
So when a Plus user does not see GPT-6 in the model picker, that is not a bug but the design. If you want it in Chat you need Pro or above; if you stay on Plus, the right door is open Work and switch the model there.
Since 4 September (Japan time), OpenAI has also been giving Plus, Pro, Business and Enterprise users who cannot yet access GPT-6 Astra one usage-reset credit for each day they cannot use it, as a bridge until the rollout completes.
What ChatGPT Work is
ChatGPT Work, the entry point for Plus, is described in the Help Center as "an agent designed for longer, multi-step work and finished deliverables." The division of roles between Chat, Work and Codex is as follows.
Conversation and everyday questions
- Fast responses
- Plus model is GPT-5.6 Sol
Research, analysis and deliverables
- Documents, spreadsheets, slides, reports, Sites
- Run once, on a schedule, or triggered by events in Gmail, Slack or GitHub
- GPT-6 Astra can be selected even on Plus
Software development
- Implementation and fixes against a codebase
- Shares its allowance with Work
Work and Codex draw from the same allowance. The Help Center says "Astra uses your allowance faster than GPT-5.6 Sol", and consumption varies with the task, input and output size, reasoning setting and Fast mode.
Codex | Astra needs CLI 0.153.0 or newer, and keeps notes across context windows
To use Astra in Codex you need Codex CLI 0.153.0 or newer and the latest desktop app. The announcement says that in Codex, Astra "can keep notes across context windows, preserving accumulated details without repeatedly compressing them into a single summary." It is a fix for the problem of assumptions dropping out during long sessions.
What happens at the limit
On Pro $200, once the 200 GPT-6 Pro messages a week are used up, ChatGPT automatically switches to GPT-5.6 Thinking (Medium). The subscription stays as it is; you wait for the reset or pick another model. The other plans work the same way: after the limit, switch model or wait. The Help Center also notes that "a fixed number of messages is not a reliable measure of remaining usage", which is why the Business guideline per five-hour window is a range of 5 to 45 messages for Astra.
API pricing: splitting work with GPT-5.6 Sol, Terra and Luna
On the API it is served as gpt-6-astra. Standard prices per million tokens, alongside the three GPT-5.6 models (Standard processing, USD).
| Model | Input | Cached input | Output | Input / output above 272K tokens | Positioning |
|---|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $20.00 / $75.00 | The hardest end-to-end work |
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 | $8.00 / $30.00 | Flagship for judgement-heavy professional work |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | $4.00 / $18.00 | Balance of intelligence and cost |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | $0.40 / $1.80 | Mechanical work such as classification and extraction |
Astra costs 2.5x Sol, 5x Terra and 50x Luna. For a request with 100K input tokens and 10K output tokens, the rough cost is Astra $1.50, Sol $0.60, Terra $0.32 and Luna $0.03.
Which model gets which job
01Jobs for Astra
Because the unit price is high, keep it for jobs where the time saving pays.
- Long processes across several tools — research, then formatting, then entry, then reporting, with a person waiting in the middle
- Computer use — driving a real screen or terminal through computer_use and hosted_shell
- One-shot hard reasoning — maths, design, large code changes, where a failed retry is expensive
02Jobs that can stay on GPT-5.6
Check whether these are enough before moving up to Astra.
- Sol — professional work that needs judgement and context. The flagship until now
- Terra — routine work that needs context: summaries, drafts, enquiry handling
- Luna — classification, extraction, tagging. One fiftieth of Astra's price
What changes when you switch: prompts need rework
In its model guidance for Astra, OpenAI states that it behaves differently from GPT-5.6 Sol and asks you to revisit your prompts. There are four main points.
- It asks clarifying questions more often. Where the previous model inferred intent and carried on, Astra may ask and stop. If you want it to proceed autonomously, say explicitly that it should infer the user's intent, bias towards action and carry the task to completion
- It is more sensitive to instructions. It is stronger with long instructions but also more sensitive to what is written in skill files and AGENTS.md, and a contradiction there can make it pause mid-task. OpenAI strongly recommends auditing the skills and files the model can read
- Output leans towards lists and tables. Long, formatted responses are the default, so specify the style you want
- It delegates to subagents less. If you want parallel work, you need an instruction that encourages delegation
The API migration items are also fixed: drop temperature, top_p and top_logprobs; replace reasoning none / minimal with low and compare; and move tool calling to the Responses API.
Three things are newly possible: async tool calling (reasoning and further tool calls continue while a slow tool runs), mid-turn steering (adding instructions over WebSocket while work is in progress), and configuration_update (changing the reasoning effort mid-conversation while keeping the cache). All three assume long agentic runs.
Other models at the same price: Claude Fable 5.1 and Gemini 3.7 Flash
Around the time of the Astra announcement, Anthropic and Google also released new models. Lining up the API prices makes the choices visible.
Claude | Fable 5.1 costs the same $10 / $50, with cache reads at $0.25
Claude Fable 5.1, released on 1 September, costs $10 input / $50 output, the same as Astra. The difference is the prompt-cache read price: Fable 5.1 is $0.25 (Astra: $1.00). In agentic work that re-reads the same long context again and again, that gap matters. It is the natural comparison if you develop in the Claude Code flow.
Gemini | 3.7 Flash is half price until year end at $0.75 / $3.75
Gemini 3.7 Flash, announced on 13 August, has its API price halved until 31 December 2026 to $0.75 input / $3.75 output ($1.50 / $7.50 from January 2027). Google positions it as "the most intelligent workhorse model for coding and agents." It is not in Astra's tier; it is a price band to consider as an alternative to Sol or Terra.
Should you switch now?
Fine to switch now
- Pro or Business and above, and "GPT-6 Pro" already shows in the Chat model picker
- Multi-step computer use or long agentic runs are your main use
- Time to completion matters more than output cost
Better to wait
- Plus, mainly using Chat — Chat stays on GPT-5.6 Sol. If you do not plan to use Work or Codex, nothing changes
- Mostly classification, extraction and summaries — Luna or Terra is enough; Astra only raises the price
- Existing prompts depend on AGENTS.md or skills — switch without auditing them and you will see mid-task pauses
Even if you are in the "better to wait" column, Pro and above can try "GPT-6 Pro" at no extra cost. The surest test is to run one of your usual jobs within the 50- or 200-a-week cap and see whether the time to finish shrinks.
FAQ
Summary
- GPT-6 Astra is a flagship tuned for computer use and long, multi-step work. 72.6% on OSWorld 2.0, with time per task down 47%
- In ChatGPT, Pro, Business and Enterprise get it in Chat as "GPT-6 Pro". Plus gets it through ChatGPT Work and Codex; Free and Go do not
- The caps are 50 a week on Pro $100 and 200 a week on Pro $200. Beyond that it switches to GPT-5.6
- The API is $10 / $50, 2.5x GPT-5.6 Sol. Check first whether Luna, Terra or Sol is enough
- Its behaviour changes: more clarifying questions, more sensitivity to skill files. Audit your prompts and AGENTS.md before switching
Related reading
- Claude Fable 5, GPT-5.5 and Gemini 3.1 Pro compared (July 2026) — the previous generation of the three models, by performance, price and use
- What Grok Bot is — a resident agent that signs in to your apps and works. How its layer differs from Astra
- 7 autonomous AI agents compared — choosing an agent by where the deliverable lands
- Generative AI (LLM) category — the latest listings
The information in this article is based on what OpenAI had published as of 8 September 2026 in its announcement, Help Center and model page. The rollout is gradual, and specifications, prices and limits may change, so please check the official site for the latest information.
AI tools featured in this article
Related articles
What Is Cloudflare OS? A Deep Dive into the Open-Source AI Agent Workspace (Architecture, Gatekeepers, Self-Hosting, Pricing)
Cloudflare open-sourced Cloudflare OS, an AI agent workspace, on August 5, 2026. This article covers its zero-permission Gatekeeper security, model selection and cost control via AI Gateway, Gadgets (small personal apps), how to deploy it into your own Cloudflare account, and pricing — based on the official blog and GitHub.
Claude Fable 5 vs. GPT-5.5 vs. Gemini 3.1 Pro: Which Should You Choose? A Deep Dive into Performance, Pricing, and Use Cases [July 2026]
A thorough comparison of the latest flagship models from the industry's three leading AI labs, all released in the first half of 2026: Claude Fable 5, GPT-5.5, and Gemini 3.1 Pro. Covers benchmark performance, API pricing, context windows, multimodal support, and how to use them on consumer plans like ChatGPT Plus, Claude Pro, and Google AI Pro — with recommendations by use case, based primarily on official data.
![Claude Fable 5 vs. GPT-5.5 vs. Gemini 3.1 Pro: Which Should You Choose? A Deep Dive into Performance, Pricing, and Use Cases [July 2026]](/_next/image?url=%2Fimages%2Farticles%2Fai-model-compare-claude-fable-5-gpt-5-5-gemini-3-1.png&w=3840&q=75 "Claude Fable 5 vs. GPT-5.5 vs. Gemini 3.1 Pro: Which Should You Choose? A Deep Dive into Performance, Pricing, and Use Cases [July 2026]")
How to Use Sakana AI Fugu: From API Key to Claude Code | Pricing, Free Tier, and Performance
A guide to Sakana AI's multi-agent platform Fugu: issuing an API key at console.sakana.ai, calling the OpenAI-compatible API, and wiring it into Codex and Claude Code. Covers choosing between Fugu, Fugu Ultra, and Fugu Cyber, pricing and whether a free tier exists, and benchmark performance.
What Is Claude Fable 5 | A Deep Dive into Anthropic's Latest Frontier Model: Performance, Pricing, and How to Use It
A comprehensive look at Claude Fable 5, Anthropic's latest model released in June 2026. Covers benchmark performance in coding, knowledge work, and science, how it compares to Mythos 5, API pricing, and usage guidance — all backed by official data.
Top 4 AI Coding Tools Compared | Choosing Between Claude Code, Cursor, Codex, and Antigravity
A side-by-side comparison of the leading AI coding tools — Claude Code, Cursor, Codex, and Antigravity. Guides you to the right choice based on features, strengths, and ideal use cases.
[2026 Edition] 20 AI Tools Transforming Developer Workflows | Coding Assistance, Automated Code Review, No-Code Web Development, and Workflow Automation
A curated look at the best AI tools for developers in 2026. Covers 18 tools across coding assistance, automated code review, no-code development, and workflow automation — all aimed at improving productivity.
![[2026 Edition] 20 AI Tools Transforming Developer Workflows | Coding Assistance, Automated Code Review, No-Code Web Development, and Workflow Automation](/_next/image?url=%2Fimages%2Farticles%2Fdeveloper-ai-programming-tools.png&w=3840&q=75 "[2026 Edition] 20 AI Tools Transforming Developer Workflows | Coding Assistance, Automated Code Review, No-Code Web Development, and Workflow Automation")