MiniMax M3.1 Flash Now Live on Novita AI at $0.15 per 1M Input Tokens

Access, pricing and specs

Developers can reach the model through Novita AI's OpenAI-compatible base URL, https://api.novita.ai/openai, using the model ID minimax/minimax-m3.1-flash. According to Novita AI, the listing also exposes Anthropic-compatible and Responses endpoint families, so an Anthropic-style client can target the model without a new integration.

• Input price: $0.15 per 1M tokens
• Cache-read price: $0.015 per 1M tokens, one-tenth of the input rate
• Output price: $0.60 per 1M tokens
• Context window: 1,000,000 tokens
• Maximum output: 524,288 tokens
• Input: text, image and video; output: text only

Novita AI says the rate is flat, so budgets scale linearly with token volume rather than jumping at a context threshold. The platform also lists function calling, structured output and reasoning support.

What MiniMax says about the model

MiniMax describes M3.1 Flash as a next-generation Flash model for coding and agent scenarios, built around a 1-million-token context window. According to its launch post, the model delivers nearly three times faster online inference and about 70% lower usage cost than its predecessor. MiniMax also reports output speeds of 200+ tokens per second, with a median around 170+ TPS.

MiniMax reports the following benchmark results, which Novita AI stresses are not its own measurements:

• BankerToolBench: 76.2, covering multi-step planning, document consistency, financial compliance and chained tool calls
• DeepSWE v1.1: 73.7%, covering repository-scale feature work, issue fixing and test validation
• OSWorld V2: 61.7, covering cross-application computer operation
• Agents' Last Exam: 29.6, covering complex professional tasks with economic value
• CyberGym: 82%, covering vulnerability discovery and validation

MiniMax also says the model ranks first among Chinese models on OSWorld V2 and Agents' Last Exam. Novita AI advises treating all scores as figures to validate on a team's own tasks.

Where it fits, and where it does not

Novita AI recommends the model for coding agents working over large repositories that need long context but not the largest model, tool-heavy automation across systems such as CRM, document and billing services, and screenshot- or recording-driven computer-use agents. It also points to high-volume agent loops with a stable prefix, where cache reads lower the bill.

The provider is equally direct about the limits. The model cannot generate images or video. Short, high-volume classification or chat is usually cheaper on a smaller model. MiniMax says the weights will be released later, so the hosted API is the practical route today and self-hosting is not yet possible. Because the scores are self-reported and the model is only days old, Novita AI cautions against relying on them where independent benchmark evidence is required.

What it means for developers

For teams already using Novita AI, switching requires one change: the new model ID on the existing base URL and key. Text-only request bodies should work unchanged, while image and video content needs the formatting described in Novita AI's API reference.

Novita AI suggests replaying a handful of real agent traces rather than testing a generic prompt, and recording solved-task rate, retries, latency and cost per completed task. It also recommends keeping the model ID in configuration so a workload that does not improve can move back to MiniMax M3, which keeps the same 1M-token context. There is no Novita-run head-to-head benchmark between the two, so the provider says the choice should rest on a team's own measurements.