Code SmithGeneralistPeer public scoresmeth_v0_1_1as of 2026-07-11

DeepSeek V4 Pro · API · Max effort

The open ceiling — near-frontier coding without the closed-model tax.

DeepSeek V4 Pro for strong agentic coding and reasoning when open weights or API cost matter more than the last closed-frontier points.

SharePost on X

Codex rating

49.8

Confidence B

Overall mean of attribute public scores after peer min-max (40–95 band within scored launch peer set).

Loadout

Endpoint: deepseek-v4-pro
Product: DeepSeek API
Price: $0.435 / $0.87 per MTok
Context: 1000k tokens

Open-weight frontier via DeepSeek API, max reasoning, 1M context, ~$0.44/$0.87 per MTok. Cost disruptor for coding/math vs closed Sol/Opus/Grok.

Core attributes

Six stats · peer public

Peer publicbars = rank in scored peer set (≈40–95) · raw = absolute fusion under the bar

STRExecution

44.7raw 75.9A40.7–48.7

DEXTempo

40.0raw 48.0B30–50

CONReliability

39.9raw 75.9A35.9–43.9

INTReasoning

45.8raw 79.5A41.8–49.8

WISJudgment

78.5raw 73.6A74.5–82.5

CHACollaboration

—D

Peer public (main number / bar) ranks this Build inside the current scored peer set (≈40–95). Raw is the absolute evidence-fusion aggregate before min-max. A low peer score can still be a strong absolute model — see methodology.

Practical verdict

When to choose

Best at

Cost-efficient repository coding (named harnesses)
Math/STEM reasoning at open-weight prices
Self-host or API flexibility (MIT weights)

Avoid for

Vision / multimodal workflows
Absolute max SWE ceiling (Sol still leads independent boards)
Teams that need only US closed-model compliance narratives

Choose this when

You want open-frontier capability at roughly an order of magnitude below Sol/Opus list prices.

Skills

Task proficiencies

Debugging77.6 A

Existing Codebase Work77.6 A

Refactoring77.6 A

Test Writing77.6 A

Quantitative Reasoning76.8 A

Research Synthesis74.3 A

Devops And Terminal67.9 B

Long Autonomous Runs67.9 B

Tool Calling67.9 B

Fast Interactive Work48.0 B

Traits & Quirks

Signature strengths and honest flaws

No structured traits or flaws catalogued for this Build yet. Traits are behavioural patterns (beyond raw scores) — added as evidence and field observation accumulate. See Best at / Avoid for above in the meantime.

Evidence Terminal

Traceability

Every factual claim cites an evidence row. Tap a chip to see the benchmark, exact config, source, and caveats behind the number.

GPQA Diamond88.8%A Humanity's Last Exam35.9%A Artificial Analysis API latency / throughput54.5B GPQA Diamond90.1%B Humanity's Last Exam37.7%B SWE-bench Verified76.2%B SWE-bench Verified77.4%A SWE-bench Verified80.6%B Terminal-Bench67.9%B

Engine pass meth_v0_1_1 (peer min-max vs scored launch peer set). Codex overall 49.8 (B). Attribute public scores map raw aggregates into a 40–95 band within the peer set (methodology min-max). INT uses hierarchical fusion (GPQA primary / HLE secondary / capped bleed).