MiniMax-M3 Pricing & Context Window: Full 2026 Cost Guide

When MiniMax released the M3 model on May 31, 2026, the headline number wasn't just another "best-in-class" benchmark — it was a price-to-intelligence ratio that reset the frontier. At $0.60 per million input tokens under the 512K threshold, MiniMax-M3 is priced closer to last year's mid-tier models than to the GPT-5.5 and Claude Opus 4.7 tier it actually competes with on coding and agent benchmarks. If you are evaluating whether to migrate a coding agent, a retrieval pipeline, or a long-context RAG workload to the new M3, the cost line is the first thing you need to lock in — and the cost line is also the easiest to get wrong, because the official Pay-as-You-Go table is structured around context length bands, cached reads, and a seven-day launch discount that all interact.

This guide pulls every number from the official Pay-as-You-Go pricing document, the rate-limits page, and the pricing overview on platform.minimax.io, cross-references them against the MiniMax Group Wikipedia entry and the OpenRouter M3 channel, and walks through what each number means for real coding-agent workloads. If you have not yet wired M3 into your editor, the companion piece MiniMax-M3 Setup: OpenCode, Cursor & Claude Code Guide covers the install step in full.

MiniMax-M3 Pricing Tiers: The Official 2026 Cost Breakdown

The Pay-as-You-Go table on platform.minimax.io splits MiniMax-M3 into two bands based on the total context length of the request — not the output length. If your prompt plus any tool-result history you are sending in stays at or below 512,000 tokens, you are billed at the standard rate. The moment a single request crosses 512K, the entire call (input, output, and cache reads) is billed at the long-context rate, which is exactly double.

Source: platform.minimax.io/docs/guides/pricing-paygo — captured May 31, 2026.

A few things that catch developers off guard the first time they read this table:

• Cache writes are free. Only the read side is metered. If you build a prompt-cache layer (which the M3 server does automatically for prompts longer than a few thousand tokens), you are not paying to populate the cache — only to hit it. The cache read at $0.12 per million tokens is 5x cheaper than fresh input.
• The 7-day 50% discount applies automatically to new accounts during the launch window. The discounted rates ($0.30 / $1.20 / $0.06) are what most early benchmarks were measured against — make sure you re-quote with the standard tier if you are running a workload that will continue past day seven.
• Subscription tier tokens are priced separately. If you are on the Plus ($20/month), Max ($50/month), or Ultra ($120/month) plan, your effective per-million-token rate is governed by the included token allowance, not the Pay-as-You-Go table above. The Max tier works out to roughly 4.5 billion tokens per month at the standard input rate, which is more than enough to drive a long-running coding agent.

For an OpenAI-compatible pay-as-you-go integration, the base URL is https://api.minimax.io/v1 and the model identifier is MiniMax-M3. If you want Anthropic-style tool-use headers, the Anthropic-compatible endpoint lives at https://api.minimax.io/anthropic. Pricing is identical on both.

Context Window: What 1 Million Tokens Actually Means

The MiniMax-M3 context window is officially 1,048,576 tokens — the binary-round 1M figure, not the decimal 1,000,000. The full window is the maximum the model can ingest in a single request, including system prompt, conversation history, and any tool-call results you append.

What the marketing page doesn't always make clear is the guaranteed minimum. For workloads that are sensitive to context length — multi-file refactors, repository-level code reviews, and document-grounded Q&A — MiniMax guarantees a usable minimum of 512,000 tokens on the standard Pay-as-You-Go rate. Requests that exceed 512K are not throttled; they are simply billed at the long-context rate shown above. There is no hard cap before you fall into that second tier, but the 1,048,576 token ceiling is the absolute wall.

The architecture that makes this 1M window practical is MiniMax's MSA — MiniMax Sparse Attention. The vendor's own benchmarks show the M3 hitting 9.7x faster prefill and 15.6x faster decode than the M2 generation when both are processing 1M-token prompts, which is the speedup that makes the long-context tier billable in the first place. Sparse attention is the same broad technique used by Gemini 1.5 and Claude 3.x to extend their windows without a quadratic compute penalty.

To put 1 million tokens in developer terms:

If you are running a coding agent that drops the full repository into context, 1M tokens gives you room for a mid-size monorepo plus tool history. The companion setup guide shows the exact config flags for OpenCode and Claude Code that take advantage of the long-context band.

Multimodal Pricing: Image and Video Token Costs

MiniMax-M3 is a natively multimodal model — text, image, and video inputs share the same endpoint and the same generation meter. Text is always the cheapest modality, by a wide margin. Image and video inputs are tokenized at a different (higher) rate and billed at a separate price per million tokens that is not displayed on the headline Pay-as-You-Go card.

Per the official LLMReference spec sheet for MiniMax-M3 API pricing, the multimodal rates are:

• Image input: $1.00 per million tokens. A typical 1024×1024 PNG converts to roughly 1,200-1,500 image tokens, so a single image costs under two-tenths of a cent at list price.
• Video input: $1.00 per million tokens. Frames are sampled and encoded at roughly one frame per second for short clips. A 30-second clip at 24 fps becomes about 720 video tokens after MiniMax's frame sampling pass — well under a cent per clip.
• Output is text-only. M3 does not generate images or video in this release. All multimodal input flows through the same reasoning path and the same output meter, with the output rate staying at $2.40 per million tokens (standard) or $4.80 (long-context).

The practical implication: a coding agent that occasionally screenshots a UI bug and feeds it back to the model will barely register the multimodal cost. A document-QA pipeline that scans 200 page screenshots per query will start to feel it — at which point switching to native PDF ingestion (text tokenization) is the cheaper path.

MiniMax-M3 vs Claude Opus 4.7: A Real Cost Comparison

Comparing frontier models on benchmarks is a noisy exercise; comparing them on cost-per-million-tokens is not. The Pay-as-You-Go rate for MiniMax-M3 standard input ($0.60) is 8.3x cheaper than the Claude Opus 4.7 list rate of $5 per million input tokens. On output, M3 is $2.40 vs Opus 4.7's $25 — a 10.4x advantage. These are not promo rates, not bundled credits, and not subscription tier shenanigans; they are the numbers on the public pricing pages of both vendors as of June 1, 2026.

For a representative coding-agent workload — say, 1.5M input tokens and 400K output tokens per day, which is the median footprint of an active Claude Code session — the daily bill at list price is roughly $0.90 on MiniMax-M3 vs $12.50 on Opus 4.7. Over a 30-day month, that is $27 vs $375.

Where Claude still wins is on a small set of agent benchmarks that are not yet replicated by M3, and on the maturity of its tool-use framework. If your pipeline relies on Claude-specific features (computer use, native MCP support, or vision-on-PDF), the cost gap may be worth paying. For most general coding, retrieval, and chat workloads, the M3 is the cheaper default in 2026 — see our MiniMax-M3 vs Claude, GPT-5.5 & DeepSeek benchmark analysis for the full head-to-head with code-accuracy scores, or the Agentic AI Explained companion article for tool-use patterns.

MiniMax-M3 vs GPT-5.5: Where the M3 Wins on Cost

OpenAI's GPT-5.5 launched in late 2025 at $5 per million input tokens and $30 per million output tokens for the standard tier, with GPT-5.4 still available at the older $2.50 / $15 rate. Gemini 2.5 Pro sits in between at $1.25-$2.50 / $10-15 depending on context length. M3 undercuts the entire field on input price and comes in well below all three on output price.

For a developer who is paying out of pocket or running a small-team SaaS, the M3 price floor changes the unit economics of AI features that used to be unprofitable. A "summarize this document" feature that costs 1.2 cents per call on Opus 4.7 costs roughly 0.09 cents on M3. A code-review tool that bills 80 cents per PR on GPT-5.5 can charge a tenth of that and still keep a healthy margin.

Rate Limits, Timeouts, and the Count Tokens Endpoint

Price gets you in the door; rate limits determine whether your workload actually runs in production. MiniMax's default per-key rate-limit envelope for M3, taken directly from the rate-limits page:

M3 trades some throughput (200 RPM, 10M TPM) for the larger context window. If you are running single-turn chat or single-file completions, this is irrelevant. If you are running a parallel batch job across hundreds of tool calls per minute, the older M2.7 still has the higher ceiling — at the cost of a much smaller 128K context window.

429 backoff and timeouts. Hitting the RPM or TPM cap returns a standard 429 Too Many Requests. The M3 server enforces exponential backoff via the Retry-After header. For very long-context requests (anything over 200K tokens), the request timeout is extended to 10 minutes — long enough to cover the prefill phase on a 1M-token prompt. Standard requests time out at 60 seconds.

The count_tokens endpoint. Both the OpenAI-compatible and Anthropic-compatible surfaces expose a count_tokens helper that returns the exact token count for a given payload — useful for pre-checking whether a request will land in the standard or long-context billing band before you commit to the call. The endpoint is free of charge; it does not generate, does not count against your rate limit, and is the right tool for any budget-aware RAG pipeline.

Free Tier and Trial Credits: How to Test MiniMax-M3 at Zero Cost

MiniMax is aggressive on developer acquisition. Three zero-cost paths exist right now:

1. Trial credits on platform.minimax.io. New accounts receive an allocation of trial credits at sign-up. These credits cover the standard Pay-as-You-Go rate and are sufficient to run thousands of M3 calls before you need to enter a card.
2. OpenRouter promo pricing. The OpenRouter M3 channel currently lists M3 at $0.30 / $1.20 per million tokens — the launch-discount rate, available to all OpenRouter users regardless of MiniMax direct-account status. Useful for one-off benchmarks and side-by-side testing.
3. Puter.js wrapper. For browser-based experimentation, Puter.js offers a free, rate-limited path to M2.x models with no signup. It is a different SKU (M2.x, not M3) but a good way to feel the model's reasoning style before committing to a billing relationship.

The fastest zero-cost path is the official trial credits. Sign up at platform.minimax.io, generate a Subscription Key, and point your OpenCode / Claude Code config at the base URL shown in the setup guide. The credits appear in your dashboard within minutes of email verification and are valid for thirty days from issuance.

Final Word: Is MiniMax-M3 the Cheapest Frontier Model of 2026?

On raw per-token price, yes. The $0.60 / $2.40 standard rate and the $0.30 / $1.20 launch-discount rate put M3 at the bottom of the frontier-model price ladder in 2026, and the 1M-token context window, multimodal inputs, and 200 RPM rate limit make it a credible primary model for coding agents, RAG pipelines, and long-context Q&A — not just a cheap fallback.

The two things to watch between now and the next pricing revision are: (1) whether the 7-day launch discount is extended or rolled into a permanent intro tier, and (2) how the long-context band's 2x multiplier holds up as more workloads push past 512K. The official Pay-as-You-Go page is the only source of truth; bookmark it and re-quote before any production rollout.

Last Updated: June 01, 2026 | Source: platform.minimax.io (Official MiniMax Developer Documentation)