DeepSeek V3.2 achieves GPT-5 level performance at one-tenth the cost. 671B parameter MoE model, MIT licensed, 96% on AIME 2025. Complete analysis.
What “GPT-5 level at one-tenth the cost” actually changes
DeepSeek V3.2 is positioned as an open-source model that reaches GPT-5-level performance while costing roughly one-tenth as much to run. That combination matters less as a headline and more as a budgeting constraint: if quality is comparable on the tasks you care about, the cheaper option lets you raise volume, add retries, or serve more users without expanding the same spend.
Treat the claim as a hypothesis to validate on your workload, not as a guarantee. Closed models and open weights differ in latency profiles, rate limits, tool support, and operational overhead. Cost-per-token or cost-per-request only tells part of the story; total cost also includes evaluation time, hosting (if self-hosted), monitoring, and the engineering needed to keep outputs reliable.
671B parameters and mixture-of-experts in practice
DeepSeek V3.2 is a 671B-parameter mixture-of-experts (MoE) model. In MoE designs, only a subset of experts activates per token, so effective compute per request can be far lower than a dense model of similar total size. That is how very large parameter counts can still be practical for inference when the routing and expert layout are efficient.
For teams evaluating the model, focus on routing behavior under your traffic shape: long contexts, multi-turn chat, code generation, and tool-using agents stress MoE systems differently. Measure tokens per second, tail latency, and quality when many concurrent sessions share the same deployment. A model that looks strong in isolation can degrade if expert capacity or memory bandwidth becomes the bottleneck under load.
MIT license and what you can actually do with it
Because DeepSeek V3.2 is MIT licensed, you can typically use, modify, and redistribute the weights with minimal friction compared with restrictive open-weight licenses. That opens paths that pure API products do not: fine-tuning on private data, air-gapped deployment, custom safety layers, and shipping the model inside products without per-token vendor lock-in for inference itself.
License freedom does not remove operational work. You still need infrastructure, model updates, evaluation harnesses, and content policies. Decide early whether you will call a hosted endpoint or self-host. Hosted use preserves the cost advantage with less ops load; self-hosting maximizes control and data locality but shifts failure modes onto your stack—GPU capacity, quantization choices, batching, and version rollouts.
How to read the AIME 2025 result and test for yourself
A reported 96% on AIME 2025 signals strong mathematical reasoning on a hard contest-style benchmark. That is useful signal for STEM tutoring, formal problem solving, and multi-step symbolic work. It does not automatically transfer to product writing, domain-specific APIs, or messy enterprise documents. Benchmarks measure slices of capability; your users measure end-to-end task success.
- Build a fixed eval set of 50–100 real tasks from production (or a close proxy), with clear pass/fail criteria.
- Score DeepSeek V3.2 against your current GPT-5 baseline on the same prompts, temperature settings, and tool access.
- Track cost and latency per successful task, not only raw quality—include retries and human review when those are part of the workflow.
- Re-run after any system-prompt, tool, or quantization change so regressions are visible before rollout.
If DeepSeek V3.2 matches your baseline on that set at substantially lower cost, expand traffic gradually with logging and sampling. If it falls short on a subset of tasks, route those cases to a stronger or more specialized model and keep the open model for the bulk of volume. The practical win is not replacing every call overnight—it is matching capability where it holds and spending less where the fit is real.