DeepSeek V4.1 Flash hands-on tests and comparisons

The author corrects earlier figures: per million tokens it is $0.15 in / $0.60 out off-peak and $0.30 / $1.20 at peak, per DeepSeek's pricing page and OpenRouter.

A blogger relays a friend's claim that V4.1 Flash-1017 may be distilled from GPT-6 Astra yet outscores it on coding benchmarks, and asks users to test it.

The author notes the semiconductor index's eight-day slide began Oct 6 while V4.1-Flash-1017 shipped Oct 17, but argues cheap open models still erode the compute-hoarding pricing narrative, citing MiniMax and Z.ai dropping over 8% after the September V4.1 Flash launch.

A user cites community testing that DeepSeek-V4.1-Flash-1017's fingerprint attribution overlaps GPT-6 Astra by over 80%, criticizing the low-cost copying of reasoning paths and warning of a possible ban.

The poster lists GPT-6.1, Gemini 4 Argon and DeepSeek V4.1 Flash alongside Opus 5.5 and asks why Claude users still stay on Opus 5.5, with no test data.

Terminal-Bench 4.0 rises 11 points at the same API price, alongside in-house Search and Fetch APIs at $0.75 per 1,000 calls.

A Turkish finance account says independent comparisons show DeepSeek V4.1 Flash offers far lower API cost than GPT-6.1 on similar tasks, without specific figures or methods.

A user wants to run DeepSeek V4.1 Flash locally but finds every option too expensive, joking that an M5 Ultra Mac Studio may be the best route.

Open weights at an unchanged API price, but the author warns benchmarks aren't your workload and suggests validating three real tasks before switching.

A user suspects the V4.1 Flash in OpenCode is not yet the 1017 build, notes no official notice, and guesses it waits on the overseas team.

Comparison data: V4.1 Flash 1017 scores 77.0 on DeepSWE v1.1, edging Claude Opus 5 at 74.0, with MIT-licensed weights and API pricing at $0.30/$1.20 per million tokens.

The post says DeepSeek-V4.1-Flash-1017 broke Chinese labs' silence, and its aggressive price-performance positioning could affect the release timing of Kimi K3.1, MiniMax 3.1 Pro and Qwen M4.