Cheap Models Now Match Frontier Performance

Three new model releases show a clear cost-performance shift. DeepSeek V4.1-Flash uses a new encoder-decoder design. It uses 8B active parameters for input processing and 16B for output. That cuts KV cache size by 8× and brings per-task cost to $0.27 versus $0.67 for V4 Pro. It still benchmarks above V4 Pro and supports 1M-token context under an MIT license. SWE-2 from Cognition is post-trained on Moonshot AI's 2.8T-parameter Kimi K3 using reinforcement learning. It matches Fable 5.1 on FrontierCode 1.1 at 64% lower cost.

Qwen3.8-Flash-Next and zAI GLM-5.3-Flash also benchmark above Claude Sonnet 5 and Opus 4.7 and can run locally. Together, these releases mean that for most production agent workloads, paying full frontier prices is no longer justified.

MarkTechPost · MarkTechPostCognition's SWE-2 matches Fable 5.1 on FrontierCode at 64% lower cost, post-trained via RL from Kimi K3's 2.8T parameters in a single run covering all effort levels.Read this signal

swyx · Cognition · Latent SpaceDeepSeek V4.1-Flash (763B, 8B prefill/16B decode) introduces a causal encoder-decoder split that cuts KV cache 8x and costs $0.27/task vs V4 Pro's $0.67.Read this signal

CPChristopher Penn · Trust InsightsQwen3.8-Flash-Next and zAI GLM-5.3-Flash outperform Claude Sonnet 5 and Opus 4.7 on Artificial Analysis benchmarks at ~95% lower cost.Read this signal