Open-weight models now 10x cheaper than frontier APIs, closing capability gap to 7 months

Open-weight models like GLM-5.2, DeepSeek V4, MiniMax M3, and Kimi K2.6 have closed the capability gap with frontier models to approximately 7 months of training equivalence, according to Epoch AI analysis updated in mid-2026. Chinese open-weight models now account for 45% of token volume on OpenRouter and other aggregators—Xiaomi's MiMo V2 Pro alone processes 4.79 trillion tokens per week, the #1 model by a 3x margin. The narrative that frontier-only systems are necessary for production is outdated.

On cost, the gap is dramatic. GLM-5.2, which scores #1 on Artificial Analysis' Intelligence Index for open weights and rivals Claude Opus and GPT-5.5 on coding and planning tasks, costs $1.40/$4.40 per million input/output tokens—vs. GPT-5.4 at $2.50/$15.00 and Claude Opus 4.8 at around $7.50/$22.50. For high-volume agentic tasks, users report 10–50x cost reductions by routing from frontier to open-weight pipelines. MiniMax M2.5 costs $0.30/million input tokens while matching frontier performance on SWE-bench Verified. DeepSeek V4 Flash is now the defacto budget option for agentic coding.

Capability remains stratified by task. Frontier models (o3, Opus 4.8, GPT-5.5) still lead on: multi-step reasoning chains, nuanced instruction-following on complex prompts, tool-use reliability under load, vision and audio, and edge-case safety tuning. Open-weight models dominate on: coding/software engineering (especially GLM-5.2 on long-horizon planning), cost-per-task efficiency (because they avoid wasted turns), and data residency (self-hosted without training-data logging). Epoch AI documents that open-weight models lag frontier by 4–14 months on advanced reasoning, but for 80–90% of production tasks—summarization, classification, extraction, Q&A—the gap is irrelevant.

For stack builders, the decision framework has inverted. Where early 2025 meant frontier-first with open-weight fallback, mid-2026 means route-by-task. High-stakes reasoning, vision, or real-time instruction → frontier. High-volume code, document work, or agent loops → open-weight + frontier hedge. The unit economics of that split now favors open-weight for 70%+ of tokens in many pipelines, which explains why Chinese model providers and Alibaba are growing faster than OpenAI and Anthropic on inference traffic.