DeepSeek V4-Pro

Independent coverage: "DeepSeek-V4-Pro-Max outperforms all prior open models across reasoning benchmarks and matches state-of-the-art closed models on many metrics"; MMLU-Pro ~87.5% (V4-Pro-Max config). These figures are for the max-effort configuration and are from secondary coverage, not a verified reading of the tech report's tables.

Independent coverage (V4-Pro-Max): LiveCodeBench Pass@1 ~93.5, Codeforces rating ~3206, statistically tied with Claude Opus 4.7 on SWE-bench Verified (~80.6 vs ~80.8). Strong open-model coding performance; the registry's `supports_coding` / `agentic_tool_use` flags are set.

Drives DeepSeek's in-house agentic coding infrastructure; integrates with Claude Code / OpenCode / OpenClaw per the release notes. The agent-oriented Responses API support for this tier was rolling out as of the registry's verification date.

1,000,000-token default context.

Strong Chinese and English; not especially broad beyond those.

~49B active parameters; the compressed-attention design keeps long-context cost manageable but this is the slower, dearer tier.

$1.32 / 1M cache-miss input, $3.96 / 1M output at peak rates, half that off-peak (DeepSeek's pricing page, checked 2026-10-10). About 4.4x V4.1-Flash on input and 3.3x on output, but still cheap versus closed frontier models of comparable reasoning/coding quality, and MIT open weights.