Open source is catching up to GPT-5.5 in just months — and costing pennies

Aggregator OpenRouter reports that the performance gap between open-source and closed-source frontier models has stabilized at just 3 to 6 months. Over the past 18 months, frontier labs haven't pulled ahead as expected, while open-source players — from both the US and China — are accelerating a global replacement wave with extreme cost advantages.

DeepSeek V4 Flash, launched just two months ago, has become the top replacement choice. With 284 billion parameters, it scored 79.0% on SWE-bench Verified, approaching GPT-5.5-level performance. Its official first-party pricing is $0.14 per million input tokens and $0.28 per million output tokens — roughly 150x cheaper than GPT-5.5 for output. Even with the premium for Western cloud hosting that doesn't retain data for training, actual costs hover around just 1.3% of leading closed-source models.

Beyond price, Zhipu AI's GLM 5.2 (released June 2026) ranks first on the Artificial Analysis open-weight intelligence index and matches GPT-5.5 on real-world agent benchmarks, making it a viable alternative for long-horizon coding and planning. However, GLM 5.2 can be token-hungry during deep reasoning, so enterprises will need to balance output costs. Meanwhile, multimodal open-source model MiniMax M3 leverages an innovative MSA sparse attention architecture to deliver native image and video long-context processing at low token prices, positioning it as a strong open-source rival to Gemini Flash.

On the US side, Nvidia's Nemotron 3 Ultra — built on the Mamba-2 hybrid architecture — emerges as the strongest domestic open-source contender, aiming to drive demand for Nvidia hardware and microservices through an open ecosystem.

OpenRouter emphasizes that while frontier closed-source models will keep advancing, the token cost for fixed capability levels will continue to plummet, offering enterprises significant optimization opportunities.