AI ExplainersNov 7, 202610 min read

GPT-6 Sol vs Astra vs Luna: Which GPT-6 Do You Get?

GPT-6 Sol vs Astra vs Luna: Astra costs 100 times Luna per token, ChatGPT's paid plans chat on Sol, and free users get Luna. Which tier fits which job.

On this page
  1. What changed since GPT-6 launched?
  2. What is the difference between GPT-6 Sol, Astra and Luna?
  3. Where do GPT-6.1 Sol and GPT-6.1 Terra fit?
  4. Which GPT-6 does ChatGPT give you on each plan?
  5. How much does each GPT-6 model cost in the API?
  6. Which GPT-6 should you use for coding?
  7. Which GPT-6 should you use for high-volume work?
  8. Where does each GPT-6 model fall short?
  9. Can you run GPT-6 with data kept in Canada?
  10. FAQ
  11. Sources

GPT-6 is OpenAI's current model family, sold in three main tiers. Astra is the most capable and costs $10 per million input tokens, Sol is the middle tier at $2, and Luna is the cheapest at $0.10. ChatGPT's paid plans chat on Sol, and the Free and low-cost Go plans get Luna. On 5 November 2026 OpenAI added a fourth model, GPT-6 Cyber, for approved security customers only.

Paying for ChatGPT? You are chatting with Sol. Astra is a separate pick for heavier work.

Building on the API and want the best results? Start with Astra, as OpenAI's own model guide advises.

For most of Astra at a fifth of the price, use GPT-6.1 Sol.

Running millions of small, repeatable calls? Use Luna. It costs a hundredth of Astra per token.

GPT-6 AstraGPT-6.1 SolGPT-6.1 TerraGPT-6 Luna
Released3 Sep 202629 Sep 2026 (GPT-6 Sol: 22 Sep)26 Oct 202622 Sep 2026
API namegpt-6-astragpt-6.1-solgpt-6.1-terragpt-6-luna
Input, per million tokens$10.00$2.00$1.00$0.10
Output, per million tokens$50.00$10.00$5.00$0.50
Reads at onceAbout 1.05M tokensAbout 1.05M tokensAbout 1.05M tokensAbout 1.05M tokens
OpenAI's own labelMost capableNear-Astra, lower costFast computer useHigh-volume tasks
In ChatGPT ChatNo (Work and Codex)Paid plans run GPT-6 SolNo (Work and Codex)Free and Go

Prices are from OpenAI's API pricing page as of 2 November 2026, and limits from its models page as of 7 November 2026. A token is about three quarters of a word.

What changed since GPT-6 launched?

OpenAI has shipped six GPT-6 releases in nine weeks, and the newest one is a cybersecurity model most buyers cannot use yet.

Date (2026)Release
3 SepGPT-6 Astra, first to a limited set of organizations, then ChatGPT paid plans, the API, Microsoft Azure and AWS Bedrock
22 SepGPT-6 Sol and GPT-6 Luna, at half the price of their GPT-5.6 namesakes
29 SepGPT-6.1 Sol, at the same $2 and $10 as Sol
7 OctGPT-6 in ChatGPT's Chat tab for every plan, with Intelligent UI
26 OctGPT-6.1 Terra, a tier between Sol and Luna
5 NovGPT-6 Cyber, for approved security customers in OpenAI's Daybreak programme

Intelligent UI lets ChatGPT answer with charts, buttons and small tools instead of text alone. OpenAI's API changelog also marked the Decisions API, which runs on Luna, generally available on 22 October.

What is the difference between GPT-6 Sol, Astra and Luna?

Sol, Astra and Luna are one generation at three price points: Astra trades cost for capability, Luna trades capability for cost, and Sol sits between. OpenAI says it trained Sol and Luna "with similar methods" to Astra. The family name comes from Generative Pretrained Transformer (GPT), the kind of model OpenAI builds.

Astra is the flagship, which its launch page calls OpenAI's most intelligent model. The page leads with computer use, meaning the model operates apps and websites on screen the way a person would. Pro, Business and Enterprise plans also get a heavier version, Astra Pro.

Astra comes with extra safety checks, because OpenAI rates its hacking ability at the top of its own risk scale. In the API, those checks can stop a task outright.

Sol is the everyday workhorse. It answers paid users in ChatGPT's Chat tab, and in the API it costs a fifth of Astra.

Luna is the volume tier. The models page calls it "our most efficient model for focused, high-volume tasks". It is also the only model behind OpenAI's Decisions API, the endpoint that picks one answer from a list you supply.

All four models on the current API price list read up to about 1.05 million tokens at once and write up to 128,000. They differ in price, speed and how well they handle hard, long jobs.

Where do GPT-6.1 Sol and GPT-6.1 Terra fit?

GPT-6.1 Sol is the current Sol, and GPT-6.1 Terra is a new tier at half its price. The API price list no longer shows the original Sol.

OpenAI describes 6.1 Sol as "near-Astra intelligence" at one-fifth of Astra's token prices. Text the model has already seen in an earlier request is billed at a cached rate of $0.10 per million tokens, half the original Sol's.

Terra costs $1 in and $5 out. OpenAI pitches it for repetitive screen work such as reading incoming invoices and entering them into accounting software. Its launch page says Terra finishes tasks more than three times faster than Astra or 6.1 Sol, and that the hardest coding work still belongs to those two.

Which GPT-6 does ChatGPT give you on each plan?

In the Chat tab, Plus, Pro, Business and Enterprise get Sol, and Free and Go get Luna. OpenAI's 7 October announcement says both are "tuned for everyday conversation".

The Work tab and Codex are a separate menu:

ChatGPT surfaceFree and GoPlusPro, Business, Enterprise
Chat tabGPT-6 LunaGPT-6 SolGPT-6 Sol
Work tab and CodexLuna in the desktop appAstra, 6.1 Sol, 6.1 Terra, LunaAstra, Astra Pro, 6.1 Sol, 6.1 Terra, Luna

Sources: the Astra, Sol and Luna, 6.1 Sol and 6.1 Terra launch pages. Enterprise admins must switch Astra on for their workspace, and the Terra page lists Plus, Pro, Enterprise and Edu but not Business.

How much does each GPT-6 model cost in the API?

Astra costs 5 times 6.1 Sol, 10 times Terra and 100 times Luna, on both input and output.

ModelInputCached inputOutput
GPT-6 Astra$10.00$1.00$50.00
GPT-6.1 Sol$2.00$0.10$10.00
GPT-6.1 Terra$1.00$0.05$5.00
GPT-6 Luna$0.10$0.01$0.50

Per million tokens, standard speed, from OpenAI's pricing page as of 2 November 2026. Fast mode on Astra runs up to twice as fast at twice the price. Regional data residency endpoints add 10%.

Take a job of 10,000 requests, each reading 2,000 tokens and writing 500. That is 20 million tokens in and 5 million out. At list price, with no caching:

ModelCost of the job
GPT-6 Astra$450.00
GPT-6.1 Sol$90.00
GPT-6.1 Terra$45.00
GPT-6 Luna$4.50

Token prices are not the whole bill. A model that thinks longer before it answers writes more output tokens per answer, and OpenAI's own charts compare cost per task, not per token.

Which GPT-6 should you use for coding?

Use Astra or 6.1 Sol. On OpenAI's numbers, 6.1 Sol matches Astra on long software tasks at about a fifth of the cost. Astra has the highest terminal score OpenAI has published for the family.

Every score below is OpenAI's own claim, at the effort setting it chose, on the same test version.

ModelLong, multi-file software tasksMulti-step jobs in a terminal
GPT-6 Astra74.1%57.9%
GPT-6.1 Sol"Matches Astra" (no figure given)Not published
GPT-6 Sol (max effort)68.8%Not published
GPT-6 Luna (max effort)66.6%Not published
GPT-6.1 Terra (max effort)Not published50.3%

Read plainly: on the long tasks, Luna at full effort finished about two in three, and Astra about three in four. A gap of a few points can come from settings alone. Moonshot's model card reports Astra at 58.2% on the same terminal test where OpenAI reports 57.9%.

For a team, the practical split is 6.1 Sol as the default and Astra for the jobs Sol gets wrong.

Which GPT-6 should you use for high-volume work?

Luna, if the job is short and repeatable: sorting, routing, pulling out fields, picking one answer from a list. It costs a tenth of Terra, the next tier up.

We have measured Luna on one kind of job. On 8 October 2026 we ran it through OpenAI's Decisions API on DecideBench, our test of 400 business decisions such as whether a refund falls inside a 30-day window. Luna got 97.2% right, and a million decisions cost $52.82 when each question came with worked examples.

Luna was close to the best, not the best. TypeSafe's Jev, a model built only for decisions, got 98.0% for $32.26 per million. OpenAI Decisions API vs Jev has the head-to-head, and what is the OpenAI Decisions API explains the endpoint. The test covers decisions only, so it says nothing about Luna's coding or writing.

For repetitive screen work, such as copying invoice fields into an accounting app, OpenAI points to Terra instead.

Where does each GPT-6 model fall short?

Every GPT-6 benchmark in this post comes from OpenAI, and we have not tested any GPT-6 model on coding.

Luna is cheap because it does less. In a test built to provoke failures, OpenAI checked whether each model admits that its search tool is broken. Luna failed to say so in 28.7% of cases, against 2.1% for 6.1 Sol and 1.5% for Astra, and everyday rates will be lower for all three because the tasks were chosen to trip the models up.

Astra does not lead everywhere. On very hard expert questions with tools, OpenAI's own table gives Astra 57.2% and Claude Fable 5.1 65.0%.

Astra's safety checks can halt API work. OpenAI says a paused task in the API "will stop", and that this can catch legitimate security work.

ChatGPT's Sol is tuned for conversation. OpenAI notes that its scores come from its research environment or API, and production ChatGPT can behave differently.

Can you run GPT-6 with data kept in Canada?

No. GPT-6 runs only on infrastructure that OpenAI or its cloud partners control, and none of those routes keep processing inside Canada today.

  • OpenAI API. OpenAI's data controls page lists Canada (ca.api.openai.com) as storage yes, processing no. Your stored data stays in Canada, and the model reads your prompt elsewhere.
  • Azure OpenAI. Microsoft's region table, as of 5 November 2026, lists regional deployments in Canada East and Canada Central up to gpt-5.2. GPT-6 models in Canada are Global-only, so they may be processed in any Azure region.
  • Amazon Bedrock. Bedrock's Canada profile keeps inference in Canada for Claude Haiku 5.5 only. OpenAI models on Bedrock go through profiles that can send requests to US regions.

For inference in Canada at this level, the route is an open-weight model, one you can download and run yourself. It runs on dedicated GPUs in a Canadian cloud region, such as AWS ca-central-1 (Montreal) or ca-west-1 (Calgary), or on hardware you own.

The closest open-weight match to Astra is Moonshot's Kimi K3.1 Pro, whose weights went up on Hugging Face on 2 November. Its model card, last changed on 5 November, lists 2.8 trillion parameters, 104 billion of them active per token, and reads about one million tokens at once. The card gives no hardware guidance, so the number of GPUs depends on your workload.

On that card, Moonshot puts Kimi ahead of Astra on the terminal test, 65.7% to 58.2%. It also shows Kimi well behind on scientific research in a terminal, 38.1% to Astra's 68.1%. Both are Moonshot's numbers. Kimi K3.1 Pro vs Kimi K3 covers what changed, and Kimi K3.1 Pro vs Claude Opus 5.5 goes further.

The Kimi K3.1 Pro License is permissive and leaves internal use free of conditions. A business that sells access to the model needs a separate agreement with Moonshot once its revenue passes $20 million in any 12 months. Products with over 100 million monthly users, or $20 million in monthly revenue, must show the model's name.

Running it yourself costs something. Large GPUs are scarcer in Canadian regions than in US ones, you pay for GPUs while they sit idle, and you own the serving.

XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account, or on-premises. Prompts and outputs stay in the country, and we build the workflow on top. How we deploy covers the build, and the Canadian residency guide covers every provider route.

If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.

FAQ

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

Yes, on the tests OpenAI published, and Sol costs 20 times as much per token. On long software tasks at full effort, the reported scores are 68.8% for Sol and 66.6% for Luna. On an honesty test, Luna failed to report a broken search tool in 28.7% of cases, against 4.9% for the original Sol. Pick Luna when the job is short and the volume is high.

Which GPT-6 model does ChatGPT Free use?

GPT-6 Luna. The 7 October 2026 announcement says ChatGPT's Chat tab runs on Luna for the Free and Go plans, and on Sol for Plus, Pro, Business and Enterprise. Free and Go users can also pick Luna in the ChatGPT desktop app's Work tab. Astra is not offered on the Free plan.

What is GPT-6.1 Sol, and does it replace GPT-6 Sol?

GPT-6.1 Sol is an upgrade to GPT-6 Sol released on 29 September 2026, at the same $2 input and $10 output per million tokens. The API price list shows only 6.1 Sol, which OpenAI says nearly matches Astra on coding, computer use and professional work at a fifth of Astra's price. In ChatGPT's Chat tab, paid plans still run GPT-6 Sol.

Can I use GPT-6 Astra through Azure or Amazon Bedrock?

Yes. Astra's launch page lists three routes: the OpenAI API as gpt-6-astra, Microsoft Azure and Amazon Bedrock. Those platforms bill you directly at their own rates. In Canada, neither Azure nor Bedrock keeps Astra's processing in the country, according to Microsoft's region table as of 5 November 2026 and our residency guide.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…