MiniMax M2.7 Review: Full‑Modal Token Plan Beats Opus at 1/50 the Cost

The author, a practitioner in AI programming and cloud‑native tools, tested MiniMax M2.7 and found its software‑engineering ability (SWE‑Pro score 56.22%) essentially on par with Claude Opus 4.6, while ranking first among open‑source models. In the GDPval‑AA office‑productivity benchmark the model achieved an ELO of 1495, comparable to Opus’s ~1500+, and in the MLE Bench Lite competition it secured a 66.6% win rate, placing just behind Opus‑4.6 (75.7%) and GPT‑5.4 (71.2%).

Self‑evolution is the model’s standout feature: it can run more than 100 automatic refinement cycles, analyzing its own failures, planning fixes, re‑testing, comparing results, and deciding whether to keep or roll back changes, which boosted its internal test‑set performance by 30%.

Token Plan pricing offers three standard tiers: Starter (¥29/month, 600 requests per 5 h, text‑only), Plus (¥49/month, 1,500 requests per 5 h, voice 4,000 chars/day and 50 images/day), and Max (¥119/month, 4,500 requests per 5 h, voice 11,000 chars/day, 120 images/day, 4 videos/day, 4 music pieces/day). High‑speed variants add Plus‑highspeed (¥98/month), Max‑highspeed (¥199/month) and Ultra‑highspeed (¥899/month) with up to 30,000 requests per 5 h for teams.

From 23 Mar 2026 a weekly quota was introduced: the 5‑hour rolling window remains, but a weekly limit equal to ten times the 5‑hour quota now applies. For example, the Plus tier gets 1,500 requests per 5 h and a weekly cap of 15,000 requests; exceeding this early can leave the rest of the week unusable.

Price comparison shows MiniMax M2.7 costs $0.30 per million tokens versus Claude Opus’s $15+, and the monthly subscription starts at ¥29 versus Claude’s ¥140+. Domestic users can connect directly, while Claude and GPT require a proxy.

Practical experience includes programming with over ten AI coding assistants (Claude Code, Cursor, Cline, etc.), production‑environment debugging that reduced mean‑time‑to‑recovery to under three minutes, and a Kaggle‑style data‑analysis task on a 446 MB dataset (3000+ rows) that covered cleaning, feature engineering, visualization, and a Streamlit dashboard. Writing tasks covered financial analysis, office documents, and multimodal generation (text, speech, image, video, music) using a single API key.

Compared with competitors, MiniMax is the only model offering self‑evolution, matches Opus in SWE‑Pro and GDPval‑AA scores, and provides direct domestic access without a proxy.

Pitfalls include peak‑hour rate limiting (especially 15:00‑17:30), the new weekly quota that can be exhausted in a single day, and limited suitability for large batch jobs. The author advises planning usage cadence, considering pay‑as‑you‑go for high‑concurrency workloads, and using the high‑speed plans to avoid throttling.

Community feedback notes the model’s high benchmark scores but occasional shallow code‑thinking, reinforcing the recommendation to start with the Starter tier, verify fit, and only upgrade if needed.

Who should use it : domestic developers who need a no‑proxy solution, multimodal creators wanting an all‑in‑one subscription, and budget‑constrained individuals (¥49/month entry). Less suitable for deep‑thinking research, highly stable production environments, or heavy batch processing.

Getting started :

Select a Token Plan (Starter is recommended for initial testing).

Configure your tool (example for Claude Code):

{
  "api_key": "YOUR_API_KEY",
  "base_url": "https://api.minimaxi.com/v1",
  "model": "MiniMax-M2.7"
}

Test common scenarios such as code completion, refactoring, bug localisation, and document generation.

Final rating (out of 5): Functionality 4, Ease of use 4, Cost‑performance 5, Stability 3, Recommendation 4. One‑sentence summary: for domestic developers seeking a cost‑effective, full‑modal AI assistant, MiniMax M2.7 is definitely worth a try.