DeepSeek

DeepSeek V4 Flash 0731 ($0.14/$0.28/M, MIT, Terminal-Bench 82.7%, FLI F — failing, Chinese lab, US Treasury deprecation deadline Oct 24) vs Mistral (European, various models $0.10-$8/M, Le Chat, FLI F — failing, dead last of 9 labs, GDPR-native). Both fail FLI. Different strengths and risk profiles.

Qwen3.8-Max-Preview (2.4T, no benchmark, no price, preview only) vs DeepSeek V4 Flash 0731 ($0.14/$0.28/M, Terminal-Bench 82.7%, MIT open weights, Codex-compatible). V4 Flash wins on every verifiable dimension today. Different use cases: V4 Flash for agentic coding, Qwen3.8 for general multimodal frontier work.

DeepSeek V4 Flash 0731 ($0.14/M), Laguna S 2.1 ($0.10/M), Kimi K3 ($3/M, sanctions risk), Grok 4.5 ($2/M) — ranked by price and Terminal-Bench. V4 Flash 0731 now offers the best cost-performance ratio among models with verified agent benchmarks. Updated August 1 after V4 Flash official release.

DeepSeek V4 Flash 0731 ($0.14/$0.28/M, 82.7% Terminal-Bench) vs Grok 4.5 ($2/$6/M, 83.3% Terminal-Bench). V4 Flash 0731 is 14× cheaper on input and 21× cheaper on output, and now within 0.6 percentage points on Terminal-Bench. The practical difference: Grok 4.5 has live X data and Cursor integration. V4 Flash is open-weight and costs almost nothing per run.

Claude Haiku 4.5 (Anthropic's fastest model, classification/routing specialist, US data residency, ~$0.80/M) vs DeepSeek V4 Flash 0731 ($0.14/M, Terminal-Bench 82.7%, MIT open weights, Codex-compatible). Very different positioning: Haiku is for latency-sensitive high-volume tasks. V4 Flash is for agentic coding at lowest cost.