MiniMax Launches M2.5 AI Model That Matches Google and Anthropic Performance at 1/20th the Cost
Summary
MiniMax launches M2.5 AI model that matches Google and Anthropic's performance while costing just 1/20th the price at $0.15 per million tokens, with the company now using AI to complete 30% of internal tasks and generate 80% of new code.
Key Points
- MiniMax releases M2.5 language model in two variants that deliver performance rivaling top-tier models from Google and Anthropic while costing 1/20th of Claude Opus 4.6, with pricing as low as $0.15 per 1M input tokens
- The model uses Mixture of Experts architecture activating only 10 billion of its 230 billion parameters per generation, achieving 80.2% on SWE-Bench Verified and matching Claude Opus 4.6 speeds while completing tasks for $0.15 versus $3.00
- MiniMax deploys M2.5 internally with 30% of company tasks completed by the AI and 80% of newly committed code generated by the model, signaling a shift from AI chatbots to autonomous AI workers