DeepSeek V4.1 Flash hands-on tests and comparisons

A user traces Open Code's shift from official DeepSeek APIs to mixed self-hosted routing, its price increase and dropped $5 credit, slower responses, and the new traffic boost from the 1017 release.

The post says Cursor turned on DeepSeek V4.1 Flash as a production default, with CursorBench runs costing $0.60–$0.98 per task and nothing scoring higher at that price, plus aggregator input pricing as low as $0.0477 per million tokens.

Daily AI news roundup noting DeepSeek V4.1-Flash-1017 hit 33.6 trillion tokens on OpenRouter for Oct 19–25, up 801% week over week, alongside GPT-6.1 Terra and other items.

A user says Cursor added DeepSeek V4.1 Flash five days after being asked, and calls the pricing very attractive.

On the Artificial Analysis Intelligence Index, Terra Max scores 48, MiMo 46, and DeepSeek V4.1 Flash 1017 Max and GLM-5.3 both 45, while DeepSeek's peak output price of $1.20 is about a quarter of Terra's.

Cursor now lists V4.1 Flash, with CursorBench putting per-task cost at $0.60–0.98; the author finds it pricier than direct API but good for cheap grunt work.

On the same three.js interstellar black hole prompt, DeepSeek V4.1 Flash used 14.9M tokens over 34 minutes for $0.17, while GPT-6.1 Terra finished in 2.5 minutes with 0.81M tokens for $0.18.

In a same-prompt SVG animation test of a pelican riding a bicycle, the author finds GPT-6.1 Terra has more natural pelican detail but clearly weaker motion logic than DeepSeek V4.1 Flash.

A user posts a screenshot saying Cursor has integrated V4.1 Flash and cites a claim that no model scores higher at the same price, rebutting doubts about Chinese models.

Running V4.1 Flash on four DGX Sparks with TensorFold v0.3.5 and 4.0bpw weights plus two patches improved the fleet, at the cost of about 15% lower 16-user throughput.

Deployed with the TensorFold v0.3.5 recipe and 4.0bpw pack on four DGX Sparks, hitting 97 tok/s for one user, 231 tok/s at 16 users, and a 93K prompt read in 37s.

The author calculates DeepSeek V4.1 Flash 1017 is 6x cheaper than GPT-6.1 Terra and 8x cheaper on cache hits, while Terra doubles its price past 272k output tokens.