DeepSeek V4.1 Flash hands-on tests and comparisons

Same one-prompt 3D test: DeepSeek V4.1 Flash looked good for under ¥1, Gemini 4 Argon web was weak but feature-rich, GPT-6 Astra was fastest yet rough, and Opus 5.5 had overexposed lighting.

V4.1 Flash 1017 generated a fully interactive browser moon base as a single three.js HTML file with zero manual edits, at nearly 200 tokens/s for a few cents.

DeepSeek's pricing page says Search and Fetch calls in Harness, Claude Code and Codex are free now but will cost ¥5 per 1,000 successful requests each from November 1; the model name stays deepseek-flash and old names silently route to 1017.

A Japanese user says V4.1 Flash is rumored to be cheap and Astra-level, possibly Astra inside, but opts out because free or ultra-low pricing may mean data collection or dumping.

Lists three head-to-head scores against Claude Opus 5: real bug fixes 77.0 vs 74.0, task automation 60.5 vs 50.3, long agentic workflows 35.4 vs 28.6 — and notes this is only the Flash tier.

Testing V4.1 Flash 1017 with an SVG pelican-on-a-bike HTML prompt, a user finds the output style closely resembles the GPT-6 series, suspecting style-specific training or distillation.

Specs roundup: 552B parameters, 8B active on input, 1M context, Terminal-Bench 4.0 at 42.3 and DeepSWE v1.1 at 77.0, beating the previous version on key coding benchmarks.

A French blog post claims the small model beats Claude Opus 5 and GPT-5.6 Sol on agent tests at roughly 10x lower cost per task, with MIT weights and API access, linking a full analysis.

Using DeepSeek-V4.1-Flash-1017 MAX, the user built a Chongqing-track racing game in an hour with rocket propulsion and launch jumps, noting the model added the top banner itself and linking a playable build.

A roundup of official info and HN feedback: coding improves notably, thinking and output are shorter, DeepSWE cost drops from $0.5 to $0.2 per task, pricing is $0.15/$0.6 per million tokens with 1M context, MIT open weights, and Claude Code compatibility.

The post cites 1M context, image input and MIT licensing, plus numbers showing DeepSWE 77.0 vs 74.0 and Automation-Bench 60.5 vs 50.3 ahead of Opus 5, while Opus 5 wins Terminal-Bench 4.0.

The poster notes the new DeepSeek v4.1 flash release and says it is time to optimize this local model, without giving concrete numbers.