GPT-6.2 Astra is OpenAI's new flagship, released on 17 November 2026. OpenAI calls it its most capable model for hard projects from start to finish — from code to documents and computer use. At OpenAI it is 20% cheaper than GPT-6 Astra, and on Kumo it costs another 40% below the official rate. Here is what the model can do, what independent measurements show and how to connect it.
Context
1.05M tokens
up to 128K tokens in one answer
Price on Kumo
$4.8 / $24
input / output per 1M tokens, 40% below OpenAI
Input
text and images
answers in text
Released
17 Nov 2026
on Kumo since 18 November
What GPT-6.2 Astra is
Astra is the top class in OpenAI's line-up, above the main Sol and the fast Terra and Luna. GPT-6.2 Astra replaces GPT-6 Astra, released in September, and is noticeably cheaper: $8 / $40 per million tokens instead of $10 / $50, and $0.4 instead of $1 for cache reads.
OpenAI focuses it on work documents: the model builds slides, spreadsheets and documents that follow your templates and field conventions — law, finance, accounting, sales — even from a screenshot of a sample. It also promises cleaner code, stronger frontends and more precise computer use, down to individual parts in CAD and engineering tools.
The knowledge cutoff is 18 May 2026. Over the API it is gpt-6.2-astra; reasoning effort is adjustable, and the model does not accept effort none.
GPT-6.2 Astra vs GPT-6 Astra and GPT-6.2 Sol
| Spec | GPT-6 Astra | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|---|
| Class | flagship | flagship | main |
| Context | 1.05M tokens | 1.05M tokens | 1.05M tokens |
| OpenAI price, input / output | $10 / $50 | $8 / $40 | $2 / $10 |
| Cache read, OpenAI | $1 | $0.4 | $0.1 |
| Kumo price, input / output | $6 / $30 | $4.8 / $24 | $1.1 / $5.5 |
What the measurements show
Per OpenAI, the model gains 5.2 points over GPT-6 Astra on Agents' Last Exam at under a third of the cost, 6.2 points on GDPval-v2 with 22% fewer output tokens, and 9 points on Terminal-Bench-Science.
Independent measurements confirm flagship level. On the Artificial Analysis Intelligence Index GPT-6.2 Astra scores 62.1 — second only to Claude Fable 5.2 (64.6) and above GPT-6 Astra (56.9). Vals AI measures 63.1% on Terminal-Bench 4.0, and a task in its index costs $10.68 against $21.28 for GPT-6 Astra. Epoch AI ties it with Fable 5.2 on its Capabilities Index.
Where to expect less: on agentic coding (the Artificial Analysis Coding Agent Index) Astra 6.2 matches GPT-6 Astra without a gain, and its hallucination rate on the very hard AA-Omniscience test is higher than its predecessor's. For documents and complex analysis it is a strong model; for code, compare it with its neighbours on your own tasks.
What GPT-6.2 Astra costs
| Where | Input | Output | Cache read |
|---|---|---|---|
| OpenAI | $8 | $40 | $0.4 |
| Kumo, paid from balance | $4.8 | $24 | $0.2 ($0.24) |
On Kumo GPT-6.2 Astra is 40% cheaper than at OpenAI itself, and in token packages it is charged with a ×1.4 multiplier — less than GPT-6 Astra (×1.8). Re-reading the same context costs a twentieth of regular input, so long projects with a shared history come out well below the list price.
GPT-6.2 Astra in the Kumo catalogue
The model card with its current price is below; it leads straight to setup.
What it is good for
- Work documents: reports, slides, spreadsheets and contracts that follow your templates.
- Complex analysis and research: long reasoning chains over a large context.
- Computer use: multi-step work across apps, engineering tools included.
- Hard end-to-end projects where a mistake costs more than the price gap to Sol.
What Kumo has verified
Before launch Kumo runs probe requests against every model. For GPT-6.2 Astra, plain answers, streaming and tool calling are confirmed over both OpenAI protocols — Chat Completions and Responses. The model then passed live requests through the api.kumorouter.com gateway, so it also works in Responses API clients such as Codex CLI.
How to connect GPT-6.2 Astra
- Sign up in the Kumo console and top up your balance.
- Create an API key and set a spending limit on it.
- Point any OpenAI-compatible client or SDK at the gateway https://api.kumorouter.com/v1.
- Set the model to gpt-6.2-astra and send your first request.
curl -s https://api.kumorouter.com/v1/chat/completions \
-H "Authorization: Bearer $KUMO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6.2-astra",
"messages": [
{"role": "user", "content": "Draft a migration plan for this service to the new database schema: ..."}
],
"max_tokens": 4096
}'
Python example
The official OpenAI SDK works as is — just change the base URL and the key. The example uses the Responses API; Chat Completions works too.
from openai import OpenAI
client = OpenAI(
base_url="https://api.kumorouter.com/v1",
api_key="KUMO_API_KEY",
)
response = client.responses.create(
model="gpt-6.2-astra",
input="Compare the two contracts and list the differences in a table: ...",
)
print(response.output_text)
FAQ
- How does GPT-6.2 Astra differ from GPT-6 Astra?
- It scores higher on the independent Artificial Analysis index (62.1 against 56.9), is stronger on work documents per OpenAI and is 20% cheaper at OpenAI. On agentic coding it matches GPT-6 Astra.
- Which to choose: GPT-6.2 Astra or GPT-6.2 Sol?
- Astra is for the hardest tasks, documents and analysis where every detail matters. Sol is a quarter of the price and close to Astra on many tasks, so it pays off for everyday development and high volume.
- Why is Kumo cheaper than OpenAI?
- That is Kumo's price for GPT-6.2 Astra: 40% below OpenAI's list price — $4.8 against $8 per million input tokens and $24 against $40 for output. You pay from your Kumo balance, with no separate OpenAI account.
- Does GPT-6.2 Astra work in Russia?
- Through Kumo, yes: the gateway works without a VPN, and the balance can be topped up with a Russian card, SBP or crypto.
Sources
- OpenAI docs: GPT-6.2 Sol and GPT-6.2 Astra, ChatGPT and Codex changelog, 17 November 2026.
- Launch coverage and OpenAI figures: The Decoder.
- Independent measurements: Artificial Analysis and Vals AI.
- Kumo prices and specs: models page.
- Epoch AI Capabilities Index — via The Decoder.