Which ChatGPT model is best: choosing a GPT version for the task
Which ChatGPT model is best depends on the task and the plan: OpenAI splits its models into GPT-6 Astra for hard work, GPT-6.1 Sol for a balance of price and quality and GPT-6 Luna for high-volume tasks. On the open ranking, ChatGPT holds 21st place in Arena Text as of 16 October 2026, and the open gpt-oss holds 215th place in Arena Text as of 16 October 2026. Below is a review by criteria: quality, tasks, price and access.
In short
- Complex reasoning and code: GPT-6 Astra
- Quality close to Astra at $2 per 1 million input tokens: GPT-6.1 Sol
- A flow of simple tasks at $0.1 per 1 million input tokens: GPT-6 Luna
- Free ChatGPT plan: GPT-6 Luna in the Work and Codex modes
- Your own model on a company server: gpt-oss, Apache 2.0 license
The list
Order: Arena Text, checked 16 October 2026.
- #1
ChatGPT
OpenAI, USA
- What it is good for
- OpenAI’s assistant with three classes of models, Astra, Sol and Luna, plus the Codex agent for code.
- Why it is on the list
- Place 21 on Arena Text as of 16 October 2026, model gpt-5.6-sol-xhigh
- Price
- free plan; paid from $8 per month (Go), Plus $20 per month (16 October 2026) plans
- Russian language
- yes per the site
- Access from Russia
- not officially supported official page
- Official site
- chatgpt.com
- Glossary term
- ChatGPT
- #2
gpt-oss
OpenAI, USA
- What it is good for
- OpenAI’s open models gpt-oss-120b and gpt-oss-20b to run on a company server.
- Why it is on the list
- Place 215 on Arena Text as of 16 October 2026, model gpt-oss-120b
- Price
- free (open weights) (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- openai.com/index/introducing-gpt-oss
In short: what to choose and when
OpenAI recommends GPT-6 Astra for complex reasoning and code, GPT-6.1 Sol for a balance of quality and price, and GPT-6 Luna for large flows of similar tasks. In a subscription, the plan sets the choice: Free and Go have Luna, and Plus adds Sol. For code, OpenAI has a separate agent, Codex: on the Plus plan it works in the web version, in the terminal, in a code editor extension and on iOS.
If data must not go to the OpenAI cloud, look at gpt-oss: the weights are open, and the model installs on your server. If you are comparing ChatGPT with other assistants, start with our review of how to choose AI for a task.
How we compared
Four criteria: the place of the versions in open rankings, the tasks for which OpenAI recommends each model, the price per tokentoken. A token is a piece of text that a language model works with: a word, part of a word, or a symbol. and the plans as of 16 October 2026, and access from Russia. Versions and prices come from OpenAI’s API models page and the ChatGPT Work and Codex pricing page on the check date. OpenAI’s help article with the list of models in the regular chat did not open on the check date.
The places come from Arena. Arena runs rankings based on blind paired comparisons: a user sees two answers without model names and picks the better one. We look at the overall text ranking and the WebDev board for code. How to read such boards is covered in our language modellarge language model. A large language model (LLM) is a neural network trained on a huge amount of text. It writes, summarizes and analyzes text on request. ranking.
Quality by rankings on the date
On the overall board, ChatGPT holds 21st place in Arena Text as of 16 October 2026, and gpt-oss holds 215th place in Arena Text as of 16 October 2026. In the snapshot of 16 October 2026, ChatGPT’s place comes from gpt-5.6-sol-xhigh, which is not among the flagship models on OpenAI’s page. Right after it stands gpt-6.1-sol-max, and below are gpt-6-astra-max and gpt-6-luna-max.
On the code board the order is reversed: in Arena WebDev, gpt-6-astra-max ranks above gpt-6.1-sol-max. For code, the extra cost of Astra shows on the board, while in the overall text ranking Sol ranks above Astra at a lower price.
The place of gpt-oss comes from gpt-oss-120b. The open modelopen weights model. An open weights model is a neural network whose parameter files you can download and run on your own hardware. loses to cloud ones on the overall board, but it works without the internet and without sending data out. For each model, Arena shows a range of places, and the ranges of neighboring versions overlap.
Tasks: texts, code, documents, search
Texts. For emails and routine answers, Luna or Sol is enough. ChatGPT has a Russian interface according to OpenAI’s help pages; check the text quality on your own examples.
Code. OpenAI recommends Astra for complex reasoning and code, and on the WebDev board it ranks above Sol. The Codex agent is in all ChatGPT plans; its comparison with Claude Code is on the page Claude Code or Codex.
Documents. All three flagship models have a context windowcontext window. The context window is the amount of text, counted in tokens, that a language model sees at one time: instructions, documents, the question and its own answer. of 1.05 million tokens and an answer of up to 128 thousand tokens. A long contract fits into any of them; the difference is in the depth of analysis and the price.
Search and data. Arena has its own board for answers with links to sources; it is covered on the page about AI search. For analyzing closed documents inside a company, gpt-oss on your own server fits: the larger version runs on one graphics card with 80 GB of memory.
Price and limits
API, price per 1 million input and output tokens: GPT-6 Astra $10 and $50, GPT-6.1 Sol $2 and $10, GPT-6 Luna $0.1 and $0.5.
Subscription. Free: no charge, GPT-6 Luna in the Work and Codex modes. Go: $8 a month, also Luna. Plus: $20, GPT-6.1 Sol and GPT-6 Luna. Pro: from $100 a month, with tiers of $100, $200 and $500. Business: $20 per user a month when paid yearly.
gpt-oss has two versions: gpt-oss-120b for a server with one 80 GB graphics card and the smaller gpt-oss-20b for more modest hardware. OpenAI released both on 5 August 2025. The model itself is free; you pay for the hardware. The memory calculator helps estimate how much memory a model needs.
Access in Russia and data
OpenAI does not include Russia and Belarus in the list of countries where ChatGPT officially works; Kazakhstan is on it. Anthropic does the same with Claude, and Google with the Gemini API. We give no instructions on getting around these restrictions.
gpt-oss works differently: it is an open model, and the weights download with no country restrictions. The Apache 2.0 license with a short usage policy allows commercial use.
Under 152-FZ, clients’ personal data is stored in databases in Russia, so the cloud ChatGPT does not fit it. How to run an open model on a company server is described in the section “Sovereign AI”. Other assistants by the same criteria are collected in our review of AI assistantsAI assistant. An AI assistant is a helper built on a neural network that, at a person’s request, answers questions, prepares texts and does simple tasks in work programs..
How we built this top
- Place of ChatGPT versions and gpt-oss in Arena Text and Arena WebDev on the check date
- Tasks for which OpenAI recommends each model
- Price per 1 million tokens and plans from official pages on the check date
- Access from Russia, open weights and license
Order: Arena Text. Places go only to tools whose models are on the board.
Checked:
How we build ratingsQuestions and answers
Which version of ChatGPT is best?
Sources
- Arena, March 2026 update, March 2026arena.ai
- OpenAI Help Center, 16 October 2026help.openai.com
- Anthropic, 16 October 2026anthropic.com
- Google AI for Developers, 16 October 2026ai.google.dev
- OpenAI, API models page, 16 October 2026developers.openai.com
- OpenAI, ChatGPT Work and Codex pricing, 16 October 2026learn.chatgpt.com
- Arena Text, ranking snapshot, 16 October 2026arena.ai
- Arena WebDev, ranking snapshot, 16 October 2026arena.ai