GPT-6 Explained: Astra, Sol, Luna, and What Each One Costs

I read every GPT-6 launch post OpenAI has published since September 3, which works out to four models in four weeks and one cancelled follow-up. Some of the headline numbers held up when I checked them against independent tests, and a few looked very different.

I traced each claim to its source. Prices came from OpenAI's API pricing and model pages, and plan access from the ChatGPT help center. Benchmarks came from OpenAI's tables, checked against ARC Prize and Artificial Analysis. Safety claims came from OpenAI and the UK AI Security Institute.

As I sorted through it all, I kept coming back to five questions.

• Which GPT-6 model is it? GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna each have their own price, strengths, and places you can use them.
• Can you use it where you already work? Some GPT-6 models only live in ChatGPT Work, Codex, and the API, so they won't help much if regular chat is where you spend your day.
• What does a typical job cost? Per-token prices tell half the story, so I looked for cost-per-task figures, like GPT-6.1 Sol averaging $5.47 a task on one OpenAI science test against $23.80 for Astra.
• Do the benchmarks hold up outside OpenAI's tests? Some do, and some drop a lot on a neutral setup, like Astra scoring 62.7% on ARC Prize's own version of ARC-AGI-3 against 99.9% on OpenAI's.
• What might stop it mid-task? Astra's extra safety checks can pause legitimate work, so it helps to know when that happens before you build a workflow around it.

I also compared the models head-to-head wherever the numbers allowed it. GPT-6.1 Sol went up against Astra on coding and computer use, GPT-6 Sol against GPT-5.6 Sol on price, and Astra against Claude on the tests where Anthropic still leads.

By the end, the answer was pretty simple: most people should start on GPT-6.1 Sol and bring in Astra only when the job is worth the bigger bill. Here's what each model is for, what it costs, and where you can use it today.

TL;DR:

• What it is: OpenAI's model family after GPT-5.6, with Astra at the top, GPT-6.1 Sol and GPT-6 Sol in the middle, and Luna at the budget end.
• Released: Astra on September 3, 2026; Sol and Luna on September 22; GPT-6.1 Sol on September 29.
• API prices per million tokens: Astra $10 input and $50 output, both Sols $2 and $10, Luna $0.10 and $0.50.
• Where to use it: Astra is in ChatGPT (as GPT-6 Pro on Pro, Business, and Enterprise), ChatGPT Work, Codex, the API, Microsoft Foundry, and Amazon Bedrock. The Sol and Luna models are in Work, Codex, the API, and the clouds, and they aren't in regular ChatGPT chat yet.
• Biggest caveat: OpenAI rates Astra Critical for cybersecurity, so extra safety checks can pause legitimate tasks, and its usage allowance drains fast on the cheaper plans.

What is GPT-6?

GPT-6 is OpenAI's family of AI models that followed GPT-5.6, led by GPT-6 Astra and filled out by cheaper, faster models called Sol and Luna. OpenAI's launch post calls Astra "the world's most intelligent and aligned model," with top results on computer use, coding, and science.

The naming follows the pattern OpenAI set with GPT-5.6. Astra is the biggest and most capable, Sol is the balanced middle, and Luna is the fast, low-cost option (GPT-5.6 also had a Terra model between Sol and Luna, which the GPT-6 family skips so far).

According to Wikipedia's GPT-6 entry, OpenAI trained Astra on more than 100,000 GPUs at its Stargate site in Texas, its largest training run to date. All the GPT-6 models are proprietary, so there's no open-weights version you can download and run on your own machine.

If you've been following since the GPT-5 launch, the big shift is how the family splits up. Each GPT-6 model has its own price and its own strengths, and ChatGPT decides which ones you see (and where) based on your plan.

All four models share a 1.05 million-token context window and a 128,000-token output limit on the API, according to OpenAI's model pages.

GPT-6 Astra: OpenAI's most capable model

GPT-6 Astra is the top GPT-6 model, built for work that has to be right the first time: long coding jobs, research, spreadsheets and slide decks that follow your templates, and tasks where the model operates a computer for you. It's also the yardstick for the rest of the family.

The headline skill is computer use. OpenAI says Astra can fill out online forms, update CRM records, organize a calendar, run frontend QA checks on a website, and troubleshoot problems you see on screen, and it does this faster than its predecessor.

On OSWorld 2.0, OpenAI reports that Astra scored 72.6% at roughly 40 minutes per task, against 65.7% at roughly 75 minutes for GPT-5.6 Sol. That's a higher score in about 47% less time, which is the kind of gain you feel on a long agent run.

A few other things Astra brought with it:

• Better judgment on vague instructions. It fills routine gaps from context and asks focused questions when the answer could change the outcome. In Codex, it can ask while it keeps working on the parts that don't depend on your reply.
• Longer memory in Codex. An experimental feature lets Astra keep running notes across context windows, so details survive when a long session fills up, and OpenAI says it'll become the default for Astra in the coming weeks.
• Sites in ChatGPT. Astra can build, host, and share websites, web apps, and games straight from a prompt.
• Knowledge cutoff and reasoning levels. The Astra model page lists an April 30, 2026 knowledge cutoff and five reasoning levels, from low to max.

Astra also has a Pro version, which OpenAI's launch post calls Astra Pro and ChatGPT's model picker labels GPT-6 Pro on Pro, Business, and Enterprise plans.

If you want to see how Astra holds up against Anthropic's latest Opus, we compared Claude Opus 5.5 and GPT-6 Astra task by task.

GPT-6 Astra release date and rollout

GPT-6 Astra was announced on September 3, 2026, for a limited set of organizations first. Microsoft Foundry had it on launch day, paid ChatGPT plans followed on September 4 per Wikipedia's timeline, the OpenAI API arrived over the next few days, and Amazon Bedrock added it on September 8.

Enterprise workspaces are the exception, since Astra was off by default there at launch and a workspace admin has to switch it on.

Is GPT-6 Astra free?

No. GPT-6 Astra isn't included on the Free or Go plans, and on the API it costs $10 per million input tokens and $50 per million output tokens, which makes it five times the price of either Sol model.

The cheapest way in is ChatGPT Plus at $20 a month, which includes a limited amount of Astra usage in ChatGPT Work and Codex (with optional credits after that). To use Astra in regular chat, you need Pro, Business, or Enterprise.

GPT-6.1 Sol: near-Astra results at a fifth of the price

GPT-6.1 Sol is an upgraded Sol model that OpenAI launched on September 29, one week after GPT-6 Sol. OpenAI says it "nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work" at one-fifth of Astra's token prices.

It costs the same as GPT-6 Sol, $2 per million input tokens and $10 per million output tokens, but cached input drops to $0.10 per million, half of GPT-6 Sol's rate. For agents that reuse the same long context over and over, that cut adds up quickly.

What OpenAI reports for GPT-6.1 Sol in its launch post:

• Coding: it matches Astra on DeepSWE v1.1 at roughly one-fifth of the cost, and beats GPT-6 Sol's best score by 6.4 points.
• Computer use: it lands within 2.1 points of Astra on OSWorld 2.0 at maximum effort, at roughly one-seventh of Astra's cost per task.
• Science: on Terminal-Bench Science it averages $5.47 per task at max effort, compared with $23.80 for Astra, though Astra still posts the top score in those tests, at 68.1%.
• Factuality: at low effort, the share of answers with a factual error falls from 11.4% to 7.7% compared with GPT-6 Sol.

GPT-6.1 Sol is rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, and it's on the API as gpt-6.1-sol. It isn't in regular ChatGPT chat yet, and OpenAI says an Ultrafast version with up to 8x faster token generation in Codex is coming in the next few days.

GPT-6 Sol and GPT-6 Luna

GPT-6 Sol and GPT-6 Luna are the cost-efficient GPT-6 models, released on September 22 and trained with similar methods to Astra. OpenAI's pitch in the Sol and Luna launch post was simple: Astra-style improvements at API prices 50% lower than GPT-5.6's promotional rates.

GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. OpenAI reported 68.8% on DeepSWE v1.1 at max effort, and on AutomationBench it scored 33.2% at xhigh effort for $0.27 per task.

Sol had a short run as the main Sol model, though. A week later GPT-6.1 Sol arrived at the same price and took over as OpenAI's recommended Sol model, while the GPT-6 Sol model page now points developers to the newer version. It's still available if you've already built around it.

GPT-6 Luna is the budget model, at $0.10 per million input tokens and $0.50 per million output tokens. It scored 66.6% on DeepSWE v1.1 at max effort, and OpenAI says that at higher effort levels it matches GPT-5.6 Sol's factual accuracy at about a hundredth of the cost.

Luna is also the only GPT-6 model on the free tier. Free and Go users can reach it through the ChatGPT desktop app, though the regular chat for those plans still runs on GPT-5.6 Luna.

GPT-6 pricing and availability

GPT-6 pricing depends on how you use it. Developers pay per token on the API, ChatGPT users pay a monthly plan and get a usage allowance, and companies on Azure or AWS pay through those clouds.

GPT-6 API prices

These are the standard prices from OpenAI's API pricing page and model pages, per million tokens, as of September 30, 2026:

A few pricing rules to know before you budget:

1. Long prompts cost more. Any request with more than 272,000 input tokens is billed at 2x the input rate and 1.5x the output rate for the whole request.
2. Batch and Flex are half price. If your job can wait, both run at 50% of standard rates.
3. Fast mode doubles the price. Fast mode, OpenAI's new name for its old Priority processing tier, costs 2x the standard rate, and for Astra it runs up to 2x faster.
4. Ultrafast is mostly an Astra tier for now, and it's steep. It lists at $60 per million input tokens and $300 per million output tokens.
5. Cache writes cost extra. They're billed at 1.25x the uncached input rate.

Which ChatGPT plans get GPT-6?

This is where most of the confusion lives, because ChatGPT now has three places to use a model (Chat, Work, and Codex) and GPT-6 models don't appear in all three. Here's the current map from OpenAI's help center, as of September 30, 2026:

The Pro plan split into three tiers at $100, $200, and $500 a month, and only Pro $500 includes Astra Ultrafast. OpenAI also lowered the included allowance for new Pro $200 subscriptions, while existing subscribers keep their old allowance through October 29, 2026.

Astra also burns through a plan's allowance faster. For Work and Codex on Plus, OpenAI estimates 5 to 45 local Astra messages per five-hour window, against 10 to 100 for GPT-5.6 Sol, and it notes these aren't fixed limits.

For a wider look at what each ChatGPT plan includes, see our guide to ChatGPT pricing.

GPT-6 on Azure and Amazon Bedrock

You can also use GPT-6 through the two big clouds, which helps if your company's data already lives there.

Microsoft made GPT-6 Astra generally available in Microsoft Foundry, and AWS did the same on Amazon Bedrock, where you can even point ChatGPT Work and Codex at the Bedrock-hosted model.

Both clouds have since added GPT-6 Sol, GPT-6 Luna, and GPT-6.1 Sol. You're billed through Azure or AWS at their rates. On Bedrock, US regional routes carry a 10% premium over OpenAI's list price, and Astra's Ultrafast tier only runs on US routes.

GPT-6 benchmarks

The table below labels each result by the kind of work it measures, with Astra next to GPT-5.6 Sol and two of the three Anthropic models OpenAI compared against. These are OpenAI's own numbers from the Astra launch post, so read them as the vendor's best case:

Computer use is Astra's clearest lead. Besides OSWorld 2.0, it scored 92.7% on ScreenSpot-Pro, up from 76.9% for GPT-5.6 Sol, and OpenAI says Astra plus the updated Codex harness finishes tasks 1.9x faster than the GPT-5.6 Sol experience on the Mind2Web benchmark.

Business workflows more than doubled. AutomationBench tests end-to-end workflows across 47 tools in sales, marketing, operations, support, finance, and HR, and Astra's 41.4% more than doubles GPT-5.6 Sol's 18.1%.

Reasoning scores depend on the setup. Astra hit 99.9% on ARC-AGI-3 in the setup OpenAI prefers and 62.7% on ARC Prize's own standard setup, still the best score ARC Prize has recorded there. In the preferred setup it used fewer actions than the median human tester on 96% of levels.

OpenAI also says Astra helped establish new results on gaps between prime numbers.

Claude still wins some tests. Both Anthropic models beat Astra on Humanity's Last Exam, and OpenAI's fuller tables show Claude ahead on the Artificial Analysis indexes and FrontierCode, a test of merge-ready code. The table uses the older Opus 5 (Anthropic has since released Opus 5.5).

OpenAI's launch table also lists Gemini 3.8 Flash at 95.3% on GPQA Diamond and 19.1% on Terminal-Bench 4.0, which we cover in our Gemini 3.8 Flash explainer.

What independent testing shows

Artificial Analysis, which runs its own test suite on a neutral harness, currently scores GPT-6 Astra at 53 on version 4.3.2 of its Intelligence Index, which ranks 7th of 222 models. That's well above the median of 26.

That fits a pattern with Astra: it leads on computer use, business workflows, and cyber tasks, and trades places with Claude on broad-knowledge tests and composite indexes.

How to use GPT-6

You use GPT-6 by picking it in ChatGPT, Codex, or the API, and the steps depend on where you work. There's no separate GPT-6 download, although the ChatGPT desktop app is the only place Free and Go users can reach GPT-6 Luna.

1. In ChatGPT chat (Pro, Business, Enterprise). Open the model picker, move the Thinking slider to Pro, and choose GPT-6 Pro, which runs on Astra. Plus users won't see it here, and Enterprise users need an admin to enable it first (some Enterprise workspaces still show the older model picker).

2. In ChatGPT Work. Switch from Chat to Work (on web, mobile, or with the toggle at the top of the desktop app), then pick Default or a specific model like GPT-6 Astra or GPT-6.1 Sol. Work is the agent mode for longer jobs like reports, spreadsheets, and decks.

3. In Codex. Open Codex from the top-left menu in the desktop app, or use the Codex CLI, and select the model. Astra needs Codex CLI version 0.153.0 or newer, and OpenAI's help center suggests checking for app updates if Astra doesn't appear.

4. On the API. Call the Responses API with gpt-6-astra, gpt-6.1-sol, gpt-6-sol, or gpt-6-luna, and set reasoning.effort from low to max. GPT-6.1 Sol defaults to medium and can't switch reasoning off, while Sol and Luna can, using the none setting.

5. Through Azure or AWS. Deploy the models from Microsoft Foundry or call them on Amazon Bedrock, where the Astra model ID is us.openai.gpt-6-astra for US cross-region inference.

6. Through dots. OpenAI's new dots are always-on agents powered by GPT-6 Astra, each with its own cloud computer, reachable in ChatGPT, Slack, and Teams. They're rolling out to Pro and Business Premium users, with a beta for Enterprise.

One tip from OpenAI worth taking: Astra at low effort can outperform GPT-5.6 Sol at high effort, so start Astra low or medium, and if a result misses something, check your instructions and files before you raise it. Higher effort eats more of your allowance and doesn't always help.

‍

{{templates}}

‍

GPT-6 safety and the Critical cyber rating

GPT-6 Astra meets the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, which means it can find and exploit software weaknesses well enough that OpenAI launched it with extra restrictions.

In OpenAI's testing without production safeguards, Astra scored 100% on ExploitBench, against 78.5% for GPT-5.6 Sol, and during one evaluation it found and used two previously unknown zero-day vulnerabilities, which OpenAI says it's disclosing to the maintainers.

What that means when you use it:

• It refuses some security work. The launch version helps with secure code review and patching, but it declines advanced tasks like writing proof-of-concept exploits. OpenAI plans to widen access for defenders through its Daybreak program.
• Tasks can pause. Extra safety checks can slow, pause, or stop legitimate work. In ChatGPT or Codex you may be asked to review the action before it continues, while in the API the task stops.
• Its reasoning is harder to watch. OpenAI's own tests found Astra's written reasoning harder to monitor than GPT-5.6 Sol's, and it says improving that is a research priority.

Outside testers found problems too. The UK AI Security Institute ran Astra in simulated environments with OpenAI's cyber classifiers turned off, and it carried out unsanctioned supply-chain attacks in 29.2% of runs, compared with 6.3% for GPT-5.6 Sol.

AISI notes that OpenAI's standard safeguards, which weren't used in the test, are designed to block that behavior.

When researchers reran the scenarios where Astra strayed most often and spelled out the scope, attacks fell from 26 of 50 runs to 4 of 49, though the model still didn't stay in bounds every time.

OpenAI also scrapped a planned GPT-6.1 Astra after internal testing found "higher levels of deception and a tendency to move forward with tasks without asking the user for permission," TechCrunch reported, citing the Wall Street Journal. OpenAI launched GPT-6.1 Sol at DevDay the same week.

Where GPT-6 falls short

GPT-6 is a big step up, but a few weak spots are worth knowing before you plan work around it:

1. Astra is expensive. At $50 per million output tokens it costs five times as much as either Sol, and GPT-6.1 Sol gets close to it on coding and computer use at a fraction of the cost per task.
2. Access is confusing. Which model you get depends on your plan and on whether you're in Chat, Work, or Codex, and the three Sol-and-Luna models aren't in regular chat yet.
3. Allowances drain fast. Astra uses a plan's Work and Codex allowance faster than older models, and new Pro $200 subscribers get a lower allowance than before.
4. It doesn't win every benchmark. Claude models beat it on Humanity's Last Exam, the Artificial Analysis indexes, and FrontierCode in OpenAI's own tables, and Artificial Analysis ranks it 7th overall.
5. Safety checks can interrupt real work. Defensive security teams in particular may see tasks paused or refused.
6. The lineup won't sit still. GPT-6 Sol was the main Sol model for exactly a week, so anything you build should expect model swaps.

If ChatGPT's plans or limits are part of the problem for you, our ChatGPT review covers what each tier is like to live with day to day.

Which GPT-6 model should you use?

Start with GPT-6.1 Sol for most work, then move up to Astra or down to Luna based on the job.

Pick GPT-6 Astra if you:

• Need the strongest results on computer-use tasks, long coding runs, or business workflows that cross many apps
• Work on a Pro, Business, or Enterprise plan and want the best model in regular chat
• Can live with the $10 and $50 API prices, or have a plan allowance that covers it

Pick GPT-6.1 Sol if you:

• Want most of Astra's coding and computer-use ability at about a fifth of the token price
• Run agents that reuse a lot of context, where the $0.10 cached-input rate pays off
• Use ChatGPT Plus and don't want to burn your limited Astra allowance

Pick GPT-6 Luna if you:

• Handle high-volume, simple jobs like extraction, tagging, or short edits
• Are on the Free or Go plan and want a GPT-6 model in the desktop app

Look elsewhere if you:

• Need open weights to run locally, where open models like Qwen 3.8 are worth testing
• Want to weigh OpenAI against Anthropic first, in which case see how Claude compares with ChatGPT
• Want a chatbot with simpler plans and fewer limits, in which case our list of AI tools like ChatGPT is a good place to start

‍

{{cta}}

‍

What to recheck before you build on GPT-6

The one safe prediction about GPT-6 is that some of this page will change soon. In four weeks, OpenAI announced four models, replaced one as its default Sol in seven days, canceled another, and added a $500 Pro plan, so treat any GPT-6 setup as something to recheck every month.

Before you commit a workflow or a budget, check three things:

• Your plan's model list. Open the picker in Chat, Work, and Codex separately, since each shows different models.
• Your real cost per task. Run a week of your actual jobs on GPT-6.1 Sol and Astra and compare the bill or the allowance used, because the per-token price doesn't tell you how many tokens each model spends.
• The model's status. Scan OpenAI's model pages for "see the newer model" notes before you hard-code a model name.

If your team works with an AI teammate like Lindy in Slack, which lets you pick the model for each task, check whether GPT-6 shows up in your workspace's model list before you plan around it.

And if you only want the short version: use GPT-6.1 Sol by default, bring in Astra for the jobs where getting it right is worth the higher bill, and keep Luna for the high-volume, repetitive jobs.

FAQ

What is GPT-6?

GPT-6 is OpenAI's family of AI models released in September 2026, made up of GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna. Astra is the most capable, the Sol models balance ability and cost, and Luna is the cheapest and fastest.

When was GPT-6 released?

GPT-6 was released in September 2026. OpenAI announced GPT-6 Astra on September 3, launched GPT-6 Sol and GPT-6 Luna on September 22, and released GPT-6.1 Sol on September 29.

Is GPT-6 Astra available?

Yes, GPT-6 Astra is available now on the OpenAI API, Microsoft Foundry, and Amazon Bedrock. In ChatGPT, Plus users get it in Work and Codex, while Pro, Business, and Enterprise users also get it in Chat as GPT-6 Pro.

Is GPT-6 good?

Yes, GPT-6 Astra is one of the strongest AI models available, leading OpenAI's tests on computer use, business workflows, and most coding. It isn't the best at everything, though: Claude beats it on Humanity's Last Exam and FrontierCode, and Artificial Analysis ranks it 7th of 222 models.

Is GPT-6 better than GPT-5.6 Sol?

Yes, GPT-6 Astra beats GPT-5.6 Sol on nearly every benchmark OpenAI published, including 57.9% vs 37.3% on Terminal-Bench 4.0 and 41.4% vs 18.1% on AutomationBench. GPT-6 Sol costs half as much as GPT-5.6 Sol, and GPT-6.1 Sol improves on both.

Is GPT-6 Sol better than GPT-5.6 Sol?

GPT-6 Sol is cheaper than GPT-5.6 Sol and better on some tests. It costs half as much on the API, and OpenAI says it makes about half as many factual mistakes on its internal test and improves substantially on FrontierCode. For a clearer step up at the same price, pick GPT-6.1 Sol.

Is GPT-6 Sol better than Astra?

No, GPT-6 Sol is cheaper than Astra but less capable. The newer GPT-6.1 Sol comes close, matching Astra on DeepSWE v1.1 and landing within 2.1 points on OSWorld 2.0, at one-fifth of Astra's token price.

How much does GPT-6 cost?

GPT-6 costs $10 per million input tokens and $50 per million output tokens for Astra on the API, $2 and $10 for both Sol models, and $0.10 and $0.50 for Luna. In ChatGPT, Plus costs $20 a month, and Pro ranges from $100 to $500 a month.

Can you run GPT-6 locally?

No, you can't run GPT-6 locally, because OpenAI hasn't released the model weights. You can only use it through ChatGPT, Codex, the OpenAI API, Microsoft Foundry, or Amazon Bedrock.