DeepSeek V4.1 Flash hands-on tests and comparisons

The author crunches numbers: off-peak output at $0.60 per million tokens vs Opus 5.5's $20, near-tie on DeepSWE v1.1, a 4.5-point Automation-Bench lead, but nearly 20 points behind on HLE.

A user relays an unverified claim that ds-4.1-flash-1017's fingerprint test points to GPT-6 Astra, joking about getting Astra at DeepSeek prices.

V4.1-Flash-1017 reaches 42.3 on Terminal-Bench 4.0, up from 31.2, while Opus 5 leads at 51.8 as open weights narrow the agentic coding gap.

In a same-task test DeepSeek took about 1.5 hours versus Gemini's 34 minutes; Gemini built a more complete track, but DeepSeek's 3D Roadster modeling looked notably better.

A user's hands-on take: very cheap, quite fast, decent quality, suited to real-time customer service, but data going to China is a dealbreaker.

A user reports DeepSeek V4.1 Flash 1017 spending about 20 million tokens to redraw an image as SVG, costing roughly $0.4.

In a same-task 3D voxel reconstruction, V4.1 Flash 1017 completed the simulation for $0.89 while GPT-6 Luna cost $1.07 per attempt with a clear quality gap.

A hands-on test has DeepSeek V4.1 Flash 1017 build an interactive Three.js site with 3D sculpture elements, using about 1.7M tokens at roughly $0.9.

A hands-on test had V4.1-Flash-1017 recreate Jet Moto from a single prompt, producing three modes, 20 riders, 4 teams and 10 tracks, all code-generated with no external assets.

Ten hands-on notes: GPT-6-Astra-like style, code-to-video and motion stronger than Haiku 5.5, long and effective Max-effort thinking, in-house search, ~15% fewer tokens and 15–20% faster per task, MIT weights, API name deepseek-flash.

TP4 run on four DGX Sparks: 6-stream throughput 131.9→165.7 tok/s, a 93K prompt drops from 78s to 42s, and the KV pool grows to 1.52M tokens.

A link roundup of the official X post, Hugging Face model card, API announcement, pricing page, a third-party hands-on writeup and Harness v0.2.1-rc.2, noting all benchmarks are vendor self-reported and unreproduced.