MiniMax’s M2.5 slashes AI costs and reframes models as persistent workers

MiniMax introduced M2.5, a two-variant language model designed to move high-capability AI from occasional consultancy-style use into continuous, production-grade automation. Technically, the model uses a sparse Mixture-of-Experts configuration that contains a very large parameter budget but routes computation so only a subset is active per token, a design choice the company says preserves depth while trimming compute. MiniMax credits a custom reinforcement learning training pipeline called Forge, together with a stability technique named CISPO, for getting agentic behavior and planning skills out of the architecture within a compressed training schedule. Benchmarks reported by the developer position M2.5 among top coding and tool-using models, with specific leaderboard figures the company released showing strong performance on coding and tool-calling tasks. The startup is shipping two commercial endpoints: a cost-optimized standard offering and a lightning variant tuned for throughput, each priced per million input and output tokens at levels the company frames as a fraction of current premium alternatives. MiniMax also disclosed internal adoption metrics: a substantial share of routine tasks and newly committed code is now produced or automated by the model in-house, a practical test of the approach beyond synthetic benchmark runs. That combination of performance, internal use, and aggressive price points is intended to change the calculus for engineering and product teams that have so far rationed model calls to manage bills. Still, key pieces of the openness claim — notably publicly posted weights and a clear license — are not yet available, which limits immediate self-hosting and independent audit. From an enterprise risk perspective, lower per-call cost increases the surface area for misuse, data-exfiltration risk, and unvetted automation unless organizations pair adoption with stronger governance. Strategically, the arrival of a cheap, high-context worker model pressures incumbents to justify premium pricing by delivering clearly superior latency, safety, or integration features. In short, M2.5 is pitched as a pivot: not merely a competitive model release but an operational lever that could make always-on, long-running agents economically realistic for more companies. Whether that promise holds will depend on independent benchmark verification, license clarity, and how well MiniMax’s disclosed training and safety methods scale in real deployments. For now, the release is a noteworthy signal that architectural sparsity plus targeted RL can compress the cost-performance trade-off in practical ways, forcing a rethink of where AI budgets and engineering efforts should flow.