deepseek-ai/DeepSeek-R1

DeepSeek-R1 remains a capable text reasoning model, but its public evidence places it below newer frontier anchors such as Claude Opus 4.8 and GPT-5.5 on broad capability, coding, and reasoning. Its 671B MoE architecture (37B active) and reported SWE-bench Verified result of 49.2% support solid technical and software capability. The current Arena-Hard-v2 preview result—58.0%, ninth place—supports a good but no longer leading reasoning score. It is text-focused, so it trails multimodal models such as Gemini 3.5 Flash. Value is the strongest practical dimension. Artificial Analysis reports median provider pricing of $1.675/M input and $4.70/M output tokens, materially better than premium proprietary frontier models, while tracked providers can deliver 178.8 output tokens/s and 14.24s TTFT. MIT licensing, weights, local-running material, API access, and broad third-party serving improve flexibility and developer access. These advantages place it above DeepSeek-V3.2’s published developer-experience anchor, though deployment of the full model is operationally demanding. Coding evidence is mixed: the official 49.2% SWE-bench Verified claim is meaningful, but Terminus 1 using R1 recorded only 5.7% ±1.4 on Terminal-Bench 1.0, and no DeepSWE result is listed. LiveCodeBench was checked but the supplied evidence contains no score. Vendor-reported historical ArenaHard 92.3 is not directly comparable with the current style-controlled Arena-Hard result. Official pricing detail is incomplete, so cost and pricing scores rely partly on Artificial Analysis provider observations.