MiniMax M3: what the model is
MiniMax M3 (exactly "M3", not "M3.0") came out on 1 June 2026 and became the flagship of the company's text lineup in place of M2.7 from 18 March 2026. The context window reaches 1 million tokens, with a guaranteed minimum of 512K. MiniMax describes it as the first open model that combines strong coding, agentic ability and native multimodality: besides text, it accepts images and video as input (MiniMax release blog).
The model is fully open. The weights sit on HuggingFace and GitHub, and you can download and deploy them without the license restrictions that some closed competitors carry.
Benchmarks
According to MiniMax, M3 scores 59.0% on SWE-Bench Pro (real bug-fixing and refactoring tasks in open repositories) and 83.5 on BrowseComp (agentic information search on the internet). On SWE-Bench Pro the company places the model above GPT-5.5 and Gemini 3.1 Pro (M3 release post). These are the vendor's own figures, not an independent measurement, so read them as a MiniMax claim rather than a neutral comparison.
For reference: the previous M2.7 scored 56.22% on the same SWE-Bench Pro, and MiniMax called it close in quality to Claude Sonnet 4.6 on the OpenClaw benchmark (M2.7 model page). MiniMax has not published a direct official comparison of M3 with the current Sonnet 5 or Opus 5 from Anthropic.
MiniMax Code: open source since 21 September 2026
Searches for "minimax code" and "minimax coding" point not to a separate model but to a product, MiniMax Code. It is a terminal coding agent with a text interface (TUI), a headless mode for the CLI and support for the Agent Client Protocol. Under the hood it uses models of the M2, M2.7 and M3 families (MiniMax Code documentation).
On 21 September 2026 MiniMax opened the source code of MiniMax Code, so the terminal agent can now be studied and self-hosted (TechNode, 21 September 2026). The official repository is on GitHub: github.com/MiniMax-AI/minimax-code.
The earlier MiniMax-M2 model is positioned by the company as a model for agents and code, and it is already integrated with several third-party tools: Claude Code, Cursor, Cline, Kilo Code, Droid (M2 release news). In other words, you can connect M2 through the branded MiniMax Code or through tools many people already use.
MiniMax Code vs Claude Code and Codex
MiniMax Code does the same practical job as Claude Code and Codex CLI: the agent reads code in the terminal, makes edits and runs commands. The key difference is the open source. You can study and modify MiniMax Code, while Claude Code and Codex remain closed products with their own ecosystems.
By the vendor's benchmarks, M3 is competitive with flagship closed models on coding tasks. But nobody has independently measured MiniMax Code with M3 against Claude Code with Claude Opus 5 or Sonnet 5 as of September 2026. We cover alternatives to Claude Code without invented numbers in Claude Code alternatives in 2026.
How much the MiniMax M3 API costs
Rates from the official page platform.minimax.io/docs/guides/pricing-paygo:
| Context | Input | Output | Cache |
|---|---|---|---|
| up to 512K tokens | $0.30 / 1M | $1.20 / 1M | $0.06 / 1M |
| over 512K tokens | $0.60 / 1M | $2.40 / 1M | $0.12 / 1M |
The older models of the lineup, M2.7, M2.5, M2.1 and M2, are billed at the same rates as M3 in standard mode: $0.30 for input and $1.20 for output per 1 million tokens. The priority tier with faster processing adds a ×1.5 markup to base prices.
For regular users there is a Token Plan subscription instead of per-token billing: the Plus, Max and Ultra plans cost from $20 to $120 a month and cover the text models and some others, but not MiniMax H3 and not the voice products. The video model H3 is covered in detail in MiniMax H3: the video model.
How to pay MiniMax from Russia
MiniMax has made no official statement blocking access for Russia: the site opens, sign-up works without a VPN, and the Terms of Service of the Singapore-registered company do not list Russia among prohibited jurisdictions. The trouble, according to users, starts at payment: Russian cards, Mir included, get declined, and the platform accepts only foreign ones. People pay with virtual foreign cards topped up through SBP (the Russian Faster Payments System), or through intermediary services that make the payment to MiniMax themselves and charge a conversion fee.
MiniMax M3 or Claude through codexkey
codexkey has MiniMax M3: the catalog lists minimax-m3 ($0.06 / $0.2 per 1M tokens, input / output), minimax-m3.1-flash ($0.26 / $0.96) and the free minimax-m3-free, with prices as of 29 October 2026. MiniMax Code itself, the subscription and the MiniMax app are a separate product, and codexkey does not sell them. If you eyed M3 for agentic coding with a large project context, Claude Opus 5 and Sonnet 5 (claude-opus-5 $1.08 / $5.4, claude-sonnet-5 $0.52 / $2.6) and the GPT-6 lineup are available alongside it. You can pay with a card issued in Russia or through SBP, with no foreign cards or intermediaries.
Claude Code connects to codexkey with these environment variables:
export ANTHROPIC_BASE_URL="https://api.codexkey.online"
export ANTHROPIC_AUTH_TOKEN="sk_cdx_your_key"
export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-5-5[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="claude-sonnet-5[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-sonnet-5"
The ANTHROPIC_DEFAULT_* lines are required: without them Claude Code asks for its own default models and answers "There's an issue with the selected model". You pick the MiniMax model inside Claude Code with the command /model minimax-m3.
Step-by-step setup is in the Claude Code guide, and current models and prices are on the models page.
FAQ
When was MiniMax M3 released?
On 1 June 2026. It took the place of the previous flagship M2.7, released on 18 March 2026.
What is MiniMax Code?
MiniMax's terminal coding agent, which runs on the M2, M2.7 and M3 models. It reads and edits code right in the terminal, works in a headless mode for scripts and supports the Agent Client Protocol. Since 21 September 2026 its source code has been open on GitHub.
How good is MiniMax M3 at programming?
By MiniMax's own claim, M3 scores 59.0% on SWE-Bench Pro, and the company calls that a result above GPT-5.5 and Gemini 3.1 Pro on this test. It is the vendor's official figure. No independent comparison with the current Anthropic models has been done.
How much does MiniMax M3 cost through the API?
$0.30 per 1 million input tokens and $1.20 per 1 million output tokens at a context of up to 512K tokens. Above 512K the price doubles. The data comes from the official pricing page as of September 2026.
How do I pay MiniMax from Russia?
A Russian card does not work directly. According to users, foreign cards and intermediary services that pay MiniMax for a fee do work. The company offers no official way to pay in rubles.
How is MiniMax Code different from Claude Code?
The main difference is the open source: you can study and modify MiniMax Code yourself, while Claude Code remains a closed product. There is no direct independent comparison of the two setups on coding tasks yet, so test them on your own real scenarios.