The State of Open-Source LLMs in 2026

A year ago I would have told you open source LLMs were a research toy and a Reddit obsession — interesting for hobbyists, unserious for anyone shipping a product. That answer is now wrong, and it has been wrong for about eight months. DeepSeek V3.2 costs $0.28 per million input tokens. Claude Opus 5 costs $15. That is a 53x delta on the same category of task, and the quality gap is nowhere near 53x. For a solo founder burning API credits, "nowhere near 53x" is the entire business case. The industry narrative is still stuck in 2024: frontier labs are unreachable, open weights are catching up, someday maybe. In 2026 that framing is upside down. Open weights are the default for anyone whose margin actually matters, and closed frontier models are the specialty tier you pay a premium for when the task genuinely demands it. This post is where each of the current open weight contenders — DeepSeek V3.2, Llama 4, Qwen 3, Mistral Large 3, Falcon 3 — actually sits, with numbers, and where the gap with GPT 5 and Claude Opus 5 still bites. The pricing gap is the entire story Let me get the ugly math out of the way first, because everything else in this post is a footnote to it. Claude Opus 5 sits at $15/M input and $75/M output [verify pricing]. GPT 5 is roughly $10/$40 [verify pricing]. Those are the reference prices for "frontier quality." Now the open weight column, all inference prices on hosted APIs so it's apples to apples: DeepSeek V3.2: $0.28/M input, $0.42/M output on the official API. Cache hits drop input to $0.028/M — one order of magnitude below the sticker. Llama 4 Maverick (400B MoE, 17B active): roughly $0.19/$0.85 on Together AI, cheaper on DeepInfra. Qwen3 235B…