DeepSeek vs Grok
Open-weight cost king vs alt-frontier value coding — both punch above their list price. Public stats are peer-relative ranks (40–95 band among scored Builds), not absolute “% intelligence.”
DeepSeek V4 Pro · API · Max effort
The open ceiling — near-frontier coding without the closed-model tax.
Grok 4.5 · API · High effort
The alt-frontier code smith — near-Opus agentic coding at a third the token bill.
Six attributes (peer public)
Same-harness & loadout
| Signal | DeepSeek | Grok |
|---|---|---|
| SWE Verified (agent boards) | 77.4% Vals mini-swe · 80.6% vendor | 86.6% Vals mini-swe |
| Terminal-Bench (published) | 67.9% vendor TB 2.0 | 83.3% vendor TB 2.1 |
| List price (in/out per MTok) | ~$0.44 / $0.87 | $2 / $6 |
| AA decode speed | ~54.5 tok/s | ~93 tok/s |
| Open weights | Yes · MIT · 1.6T/49B MoE | No · closed API |
| Arena Elo (openlm) | — (not pinned) | 1505 |
Large Existing Codebase
Decisive on Execution: Grok — public ranges do not overlap (DeepSeek 40.7–48.7 vs Grok 64.8–72.8).
- Best Under BudgetDeepSeek · conf A
First-party list price is roughly 4–7× cheaper on output than Grok; open weights unlock self-host.
Quest Fit → - Large Existing CodebaseGrok · conf A
Independent mini-swe gap (86.6% vs 77.4% Vals) and peer STR favor Grok.
Quest Fit → - Autonomous Coding AgentGrok · conf B
Stronger agentic boards + tempo; DeepSeek still viable when cost dominates.
Quest Fit → - Vibe Code a New AppGrok · conf B
DEX and preference signal; DeepSeek wins if the stack is open/self-host first.
Quest Fit → - Fast EverydayGrok · conf A
Peer DEX max and snappier high-effort decode story.
Quest Fit → - Private Local CodingDeepSeek · conf B
Only if you self-host the MIT weights — this Build’s scores are API; local quant is a different Build.
Quest Fit →
DeepSeek: MIT open weights + ~$0.44/$0.87 API — strong absolute coding when you optimise for cost and self-host options.
Grok: Closed alt-frontier with higher independent SWE, class-leading tempo, and Arena preference at $2/$6.