DeepSeek V4.1 Flash hands-on tests and comparisons

The post reports open weights, self-reported coding scores near the frontier at about 20 cents per task, and a search engine turned API at $0.75 per 1,000 calls, free in Claude Code and Codex until October 31.

Higgsfield's same-prompt 3D comparison: DeepSeek's floating island, lighthouse and moon are more fantastical, while Claude's cloud sea has better volume and hull detail, though Opus 5.5 may cost 10x more.

A 100-HTML-file creativity and frontend test of V4.1 Flash: roughly on par with GLM 5.3, clearly behind Qwen3.8-Max, with some broken files.

The poster argues the new DeepSeek V4.1 Flash 1017 feels better than Gemini 4 Argon despite lower benchmark scores, and asks whether the benchmarks are misleading again.

The post frames DeepSeek-V4.1-Flash-1017 as an open-source version of Opus 5 and predicts an Opus 5.5-level open model runnable locally on a Mac Studio within months.

On the author's custom benchmark, Opus 5.5 tops at 88.9 for $8.89 a run, while DeepSeek V4.1 Flash gets 80.3 for just $0.83, a standout price/perf result.

DeepInfra says weekend-long agent demand for DeepSeek V4.1 Flash has been huge, so it is cutting the price by 51% to keep workloads running.

A fact-check rates DeepSeek's 11.1-point benchmark gain as supported while finding no independent evidence that per-task costs fell, noting unchanged token prices don't guarantee cheaper tasks.

Hands-on test: V4.1 Flash 1017 was asked to produce a 5-second DeepSeek ad clip, and the author finds the result strong given the time and length constraints.

A 3.5bpw EXL3 quant (240 GiB) of DeepSeek V4.1 Flash 1017 scores only 0.8% / 0.1% worse than base on code and prose sets, with 2.9bpw numbers and Kit KMAP_HOST setup included.

Same pricing, Terminal-Bench 4.0 up from 31.2 to 42.3, DeepSWE 77.0, shorter faster outputs and MIT weights, though tool-free HLE is 42.6 vs Opus 5's 56.3; also notes $0.75 per 1,000 search/fetch calls and free Claude Code/Codex access until Nov 1.

An opinion post arguing DeepSeek V4.1 Flash-1017 is roughly 80% as capable as Opus 5.5, with no benchmarks or side-by-side results given.