Artem Voloskovets
⬤ DeepSeek rolled out its newest models, with V3.2 Speciale leading the pack and shaking up what we expect from frontier AI. V3.2 Speciale puts up numbers that beat GPT-5 High and compete with Google's Gemini 3.0 Pro, all while being fully open-source under MIT license. The benchmarks show this model hitting top scores in tough math and coding tests.
⬤ Looking at the numbers, DeepSeek V3.2 Speciale scored 96.0 on AIME 2025, beating GPT-5 High's 94.6 and nearly matching Gemini 3.0 Pro's 95.0. It also pulled 99.2 on HMMT Feb 2025 and 94.4 on HMMT Nov 2025. For coding, it got 88.7 on LiveCodeBench, which tops GPT-5 High's 84.5. The model's competitive programming game is solid too, with a 2701 Codeforces rating that beats GPT-5 High.
"The rapid advances shown by DeepSeek V3.2 Speciale may reshape the competitive dynamics between open-source and closed-source AI ecosystems."
⬤ DeepSeek's full lineup shows how tight the race between open-source and proprietary AI has gotten. What really stands out is the pricing: V3.2 Speciale runs about five times cheaper than GPT-5 High for regular use and up to 24 times cheaper for output tokens. Mix that with these benchmark wins, and you've got a model that's seriously disrupting the top tier of AI systems.
⬤ These gains from DeepSeek V3.2 Speciale could flip the script on how open-source and closed-source AI compete. Strong performance at much lower costs might push more companies to adopt it, change how cloud inference works financially, and speed up open-source use across research, dev work, and AI apps. As open models keep closing the gap, the whole AI scene might shift pretty dramatically in the coming months.
Artem Voloskovets
