Hands-on: DeepSeek V4.1 Flash 1017 in OpenCode, five observations
Testing 1017 in OpenCode: clearly ahead of GLM 5.3 Flash on normal coding and closer to Sonnet-class, $10/month Go plan yields ~$80 of inference, free long-context loophole gone with compression near 300K, and token burn still fast in long sessions.
i’ve been testing DeepSeek V4.1 Flash 1017 in OpenCode, and 5 things stood out: 1. it now feels clearly ahead of GLM 5.3 Flash for me on normal coding work. the gap is bigger than it was with the previous V4.1 checkpoint. 2. the value on OpenCode Go is kind of absurd. $10/month for roughly $80 of inference makes it hard to justify using direct API for everything. 3. the free context loophole looks gone. i’m seeing compression kick in around 300K context instead of being able to stretch it close to 1M. 4. the model itself feels closer to Sonnet-class now than the other cheap flash models, especially once tasks get a bit more complex. 5. token usage still disappears faster than i’d like on longer sessions. the model is cheap, but cache efficiency and context management still matter a lot. 1017 feels like the point where DeepSeek Flash stopped being just “good for the price.” it’s now genuinely competitive.
This post quotes another post: x.com/opencode/status/2112220734768779441