{"id":268,"date":"2026-04-20T15:59:05","date_gmt":"2026-04-20T15:59:05","guid":{"rendered":"https:\/\/santiagomarquezsolis.com\/?p=268"},"modified":"2026-04-20T15:59:05","modified_gmt":"2026-04-20T15:59:05","slug":"open-source-llms-that-rival-paid-models-in-2026-a-practical-comparison","status":"publish","type":"post","link":"https:\/\/santiagomarquezsolis.com\/index.php\/en\/2026\/04\/20\/open-source-llms-that-rival-paid-models-in-2026-a-practical-comparison\/","title":{"rendered":"Open-Source LLMs that rival paid models in 2026: a practical comparison"},"content":{"rendered":"

Open-Source LLMs that rival paid models in 2026: a practical comparison<\/h2>\n

For years, the conversation about state-of-the-art large language models was dominated by closed, paid APIs: GPT, Claude and Gemini. That gap has narrowed dramatically. In the most recent Artificial Analysis Intelligence Index (v4.0)<\/strong>, which aggregates ten independent evaluations (GDPval-AA, \u03c4\u00b2-Bench Telecom, Terminal-Bench Hard, SciCode, AA-LCR, AA-Omniscience, IFBench, Humanity’s Last Exam, GPQA Diamond and CritPt), several open-weights models now sit only a handful of points below the best proprietary systems, and often at a fraction of the cost.<\/p>\n

This article reviews the open-source alternatives that are closest in quality to the paid flagships today, where each one shines, and when it genuinely makes sense to replace a commercial API with a self-hosted or open-weights option.<\/p>\n

What \u00abopen-source\u00bb really means here<\/h3>\n

In the LLM world, \u00abopen-source\u00bb is often used loosely. Most of the models discussed below are open-weights<\/strong>: the weights are downloadable and can be self-hosted, and many (but not all) allow commercial use. A minority are fully open (weights + training data + recipe). In this article, \u00abopen\u00bb refers to open-weights models usable at zero or near-zero cost when self-hosted, or very cheaply through third-party inference providers.<\/p>\n

The leading open-weights contenders<\/h3>\n

According to the Intelligence Index v4.0, the top open-weights models are led by GLM-5.1<\/strong> (Zhipu\/Z AI) with a score of 51, followed by MiniMax-M2.7<\/strong> (50) and Qwen3.5 397B A17B<\/strong> (Alibaba, 46). Other serious contenders include DeepSeek V3.2<\/strong> (42), gpt-oss-120B<\/strong> from OpenAI (36), Gemma 4 31B<\/strong> from Google (39), Mistral Small 4<\/strong> (26), NVIDIA Nemotron 3 Super<\/strong> (33) and Llama 4 Maverick<\/strong> (18) from Meta. For reference, the proprietary leaders \u2014 Claude Opus 4.7, Gemini 3.1 Pro Preview and GPT-5.4 \u2014 are tied at 57.<\/p>\n

Comparative table: Intelligence Index<\/h3>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Model<\/th>\nType<\/th>\nIntelligence Index<\/th>\nOpen weights<\/th>\nNotes<\/th>\n<\/tr>\n<\/thead>\n
Claude Opus 4.7 (max)<\/td>\nProprietary<\/td>\n57<\/td>\n\u274c<\/td>\nAnthropic flagship<\/td>\n<\/tr>\n
Gemini 3.1 Pro Preview<\/td>\nProprietary<\/td>\n57<\/td>\n\u274c<\/td>\nGoogle flagship<\/td>\n<\/tr>\n
GPT-5.4 (xhigh)<\/td>\nProprietary<\/td>\n57<\/td>\n\u274c<\/td>\nOpenAI flagship<\/td>\n<\/tr>\n
GLM-5.1<\/strong><\/td>\nOpen<\/td>\n51<\/td>\n\u2705<\/td>\nBest open model overall<\/td>\n<\/tr>\n
MiniMax-M2.7<\/strong><\/td>\nOpen<\/td>\n50<\/td>\n\u2705<\/td>\nStrong reasoning<\/td>\n<\/tr>\n
Qwen3.5 397B A17B<\/strong><\/td>\nOpen<\/td>\n46<\/td>\n\u2705<\/td>\nMoE, 17B active params<\/td>\n<\/tr>\n
DeepSeek V3.2<\/strong><\/td>\nOpen<\/td>\n42<\/td>\n\u2705<\/td>\nGreat quality\/price ratio<\/td>\n<\/tr>\n
Gemma 4 31B<\/strong><\/td>\nOpen<\/td>\n39<\/td>\n\u2705<\/td>\nLight and efficient<\/td>\n<\/tr>\n
gpt-oss-120B (high)<\/strong><\/td>\nOpen<\/td>\n36<\/td>\n\u2705<\/td>\nOpenAI’s open release<\/td>\n<\/tr>\n
NVIDIA Nemotron 3 Super<\/strong><\/td>\nOpen<\/td>\n33<\/td>\n\u2705<\/td>\nNVIDIA ecosystem<\/td>\n<\/tr>\n
Mistral Small 4<\/strong><\/td>\nOpen<\/td>\n26<\/td>\n\u2705<\/td>\nSmall footprint<\/td>\n<\/tr>\n
Llama 4 Maverick<\/strong><\/td>\nOpen<\/td>\n18<\/td>\n\u2705<\/td>\nMeta’s MoE model<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Source: Artificial Analysis Intelligence Index v4.0.<\/em><\/p>\n

Comparative table: Price (blended, USD per 1M tokens)<\/h3>\n\n\n\n\n\n\n\n\n\n\n\n\n
Model<\/th>\nBlended price (USD \/ 1M tok)<\/th>\n<\/tr>\n<\/thead>\n
Claude Opus 4.7 (max)<\/td>\n$10.00<\/td>\n<\/tr>\n
GPT-5.4 (xhigh)<\/td>\n$6.00<\/td>\n<\/tr>\n
Gemini 3.1 Pro Preview<\/td>\n$5.60<\/td>\n<\/tr>\n
Claude Sonnet 4.6 (max)<\/td>\n$4.50<\/td>\n<\/tr>\n
GLM-5.1<\/strong><\/td>\n$2.10<\/strong><\/td>\n<\/tr>\n
DeepSeek V3.2<\/strong><\/td>\n$0.30<\/strong><\/td>\n<\/tr>\n
gpt-oss-120B (high)<\/strong><\/td>\n$0.30<\/strong><\/td>\n<\/tr>\n
NVIDIA Nemotron 3 Super<\/strong><\/td>\n$0.40<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

DeepSeek V3.2 costs roughly 33\u00d7 less<\/strong> than Claude Opus 4.7 while scoring 42 vs 57 on intelligence \u2014 very close to the paid flagship for many everyday tasks.<\/p>\n

Specific strengths by use case<\/h3>\n