GPT-6 Astra is the flagship model OpenAI introduced on 3 September 2026. It is built for the heaviest work: code, research and tasks where the model operates a computer. In the API its identifier is gpt-6-astra. In ChatGPT and Codex it opens on paid plans only, with no free access.
What GPT-6 Astra is
Astra sits at the top of the GPT-6 line. In the model documentation OpenAI calls it its most capable model for the most demanding tasks. The parameters come from the same page: a 1,050,000-token context, up to 922,000 tokens of input and up to 128,000 of output, knowledge through 30 April 2026. Requests go through the Responses API and Chat Completions, and a batch mode exists. Reasoning depth is set by reasoning effort, which has five values: low, medium, high, xhigh, max.
The naming is confusing. "ChatGPT 6 Astra" is not a separate product: ChatGPT is the app, GPT-6 is the model family, and Astra, Sol and Luna are three tiers inside that family. A search for "chat gpt 6 astra" usually means the Astra model inside ChatGPT or Codex.
How GPT-6 Astra differs from Sol and Luna
OpenAI split GPT-6 into three steps. Astra takes the hardest work. Sol is meant for complex code and professional tasks. Luna covers narrow, high-volume operations: field extraction, summaries, classification. Sol and Luna came out on 22 September 2026, 19 days after Astra (TechCrunch). On 29 September GPT-6.1 Sol joined them, which OpenAI describes as nearly Astra-level for less money.
| Model | What it is for (OpenAI's description) | OpenAI price, input / output per 1M | codexkey catalog, input / output per 1M |
|---|---|---|---|
gpt-6-astra | the hardest work | $10 / $50 | $0.68 / $3.3 |
gpt-6.1-sol | complex code, computer use, professional tasks | $2 / $10 | $0.28 / $1.42 |
gpt-6-sol | complex code and agent scenarios | $2 / $10 | $0.14 / $0.68 |
gpt-6-luna | narrow high-volume tasks | $0.10 / $0.50 | $0.04 / $0.22 |
The OpenAI prices come from its pricing page and apply to requests of up to 272 thousand input tokens; longer ones are billed at other rates. There is also GPT-6.1 Terra ($1 and $5), but the codexkey catalog did not list it on 29 October 2026.
At OpenAI's prices Astra costs exactly five times GPT-6.1 Sol, on input and on output. In the codexkey catalog the gap is smaller: about 2.4 times on input and 2.3 on output. We do not repeat OpenAI's benchmarks, because they are the vendor's own reports. Compare on your tasks. A general selection table for the whole line is in GPT-6 and GPT-5.6 for the API, and how the roles split inside Codex is in Codex models: Sol, Astra, Luna.
When you need Astra and when Sol is enough
The rule is simple: Astra pays off where a mistake costs more than the tokens.
| Task | Where to start | Why |
|---|---|---|
| Fixing a function, tests, reviewing a small PR | gpt-6.1-sol or gpt-6-sol | a few steps are enough, price and limits are lower |
| An intermittent failure across several services | gpt-6-astra | an engineer's hour costs more than the token difference |
| Refactoring a module with non-obvious dependencies | gpt-6.1-sol, then gpt-6-astra if it fails | cheap attempt first |
| Tags, summaries, fields from thousands of documents | gpt-6-luna | volume matters more than depth |
| A long autonomous agent session | gpt-6.1-sol | Astra eats the subscription limit fast |
On limits the numbers are clear. On the Codex pricing page the estimates for Plus are: Astra 5-45 local messages per five hours, GPT-6.1 Sol 15-160, GPT-6 Luna 350-3000. These are estimates, not a guarantee, but the order of the gap is plain: Astra burns through the subscription window many times faster than the others.
We would not make Astra the default. Keep Sol as the default and switch Astra on for one session when Sol has failed the task twice in a row. Checking this is easy: take ten closed tasks from your repository, run them on both models with one prompt, and compare not the price of a request but the price of a solved task.
How much GPT-6 Astra costs
In the OpenAI API on 29 October 2026: $10 per 1M input tokens, $1 per 1M cached input, $50 per 1M output. In ChatGPT you pay for a plan, not for tokens. Plus at $20 a month includes Astra in ChatGPT Work and Codex, Pro starts at $100, and on Free and Go only GPT-6 Luna opens in Codex.
What each ChatGPT plan gets, per the Codex pricing page of 29 October 2026:
| Plan | Price per month | Astra |
|---|---|---|
| Free | $0 | no, Codex offers GPT-6 Luna |
| Go | $8 | no, Codex offers GPT-6 Luna |
| Plus | $20 | yes, estimated at 5-45 local messages per five hours |
| Pro | from $100 (tiers $100, $200, $500) | yes, the five-hour limit does not currently apply to Pro |
| Business | $20 per user on an annual contract, $25 month to month | yes, same estimate as Plus for the standard plan |
In the codexkey catalog gpt-6-astra costs $0.68 per 1M input tokens, $3.3 per 1M output and $0.08 for a cache read, with a 1000K context. By the catalog's description the model accepts images and supports reasoning. For a rough estimate: a request with 100 000 input and 10 000 output tokens comes to $0.101 (0.068 + 0.033). We do not convert prices to rubles, because the exchange rate floats. Fresh prices are on the models page.
How to use GPT-6 Astra
In ChatGPT and Codex from OpenAI
Astra opens on paid plans in ChatGPT Work and Codex. The app and the web have Power, Smarter and Faster presets, and a specific model is picked under Advanced; in the CLI the /model command works (this is what OpenAI's model documentation says). For a reader in Russia this route does not fit: Russia is not on the list of countries where OpenAI supports the API, and ChatGPT has its own separate list. Details are in ChatGPT 6 in Russia.
Through the OpenAI SDK and codexkey
You need a sk_cdx_... key from your account. Payment is by a Russian bank card, SBP (the Russian Faster Payments System) or crypto, and no OpenAI account or VPN is required. The address of the OpenAI-compatible API is https://api.codexkey.online/v1.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.codexkey.online/v1",
api_key=os.environ["CODEXKEY_API_KEY"],
)
reply = client.chat.completions.create(
model="gpt-6-astra",
messages=[{"role": "user", "content": "Find why the test fails and suggest a fix"}],
)
print(reply.choices[0].message.content)
The same thing as a single request without the SDK:
curl https://api.codexkey.online/v1/chat/completions \
-H "Authorization: Bearer $CODEXKEY_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-6-astra", "messages": [{"role": "user", "content": "Explain what this regexp does"}]}'
Current IDs come from GET https://api.codexkey.online/v1/models. If the model you want is missing from the response, check the spelling: IDs are copied from the catalog exactly.
Which reasoning effort to set
There are five levels, and the top ones are not needed on every request. This split is ours, not an OpenAI requirement:
| Level | When |
|---|---|
low | pull a field out of text, a mechanical edit |
medium | plan a feature, compare two approaches |
high | debugging with an unclear cause, reading someone else's code |
xhigh, max | when high fell short and the gain is worth the extra time and tokens |
The higher the level, the longer the chain of reasoning and the more the answer costs. Start with medium and climb only if a specific task failed.
In Codex through codexkey
The provider is declared in ~/.codex/config.toml following the codexkey guide: a base_url ending in /v1 and wire_api = "responses". After that one command is enough:
codex --model gpt-6-astra
Signing in with an API key instead of a ChatGPT subscription is covered in Codex CLI with an API key.
What to check before you launch
Budgets usually burn on three things. First, reasoning tokens on models with reasoning are most often billed at the output price, so max on every request inflates the bill. Second, a long context costs more than a short one, so avoid sending a whole repository into the window without need. Third, the gateway has its own request-rate limits, described in the limits section.
For critical scenarios, put tests and a human review after the model. A strong model does not cancel CI. If you are choosing between a subscription and paying per token, a breakdown by scenario is in GPT-6 Astra: free or paid.
FAQ
What is GPT-6 Astra in plain words?
It is the most capable OpenAI model in the GPT-6 line, introduced on 3 September 2026. People reach for it for complex code, research and tasks where the model operates a computer. For simple high-volume operations there are the cheaper Sol and Luna.
How much does GPT-6 Astra cost?
In the OpenAI API: $10 per 1M input and $50 per 1M output tokens. In the codexkey catalog: $0.68 per 1M input and $3.3 per 1M output tokens (price list of 29 October 2026). In ChatGPT the model comes with paid plans, and Plus costs $20 a month.
How does GPT-6 Astra differ from GPT-6 Sol?
Astra is more capable and, at OpenAI's prices, five times as expensive as GPT-6.1 Sol. Sol suits everyday development, while Astra is for when Sol cannot cope and a mistake is costly. Astra spends subscription limits noticeably faster.
How do I use GPT-6 Astra?
In ChatGPT and Codex you pick the model in the selector on a paid plan. Through the API it is enough to set model="gpt-6-astra". At codexkey you need a sk_cdx_... key and the address https://api.codexkey.online/v1.
Is chat GPT 6 Astra in free ChatGPT?
No. The Codex pricing page lists only GPT-6 Luna for Free and Go. Astra opens with paid plans, and without a subscription you can use it through the API with per-token billing.
Does GPT-6 Astra work in Russia?
Officially no: Russia is not on the list of countries where OpenAI supports the API, so direct sign-up and payment are unreliable. The working route for code and applications is the codexkey OpenAI-compatible API, paid in rubles and with no VPN.