OpenCode announced on October 15, 2026 that DeepSeek V4.1 Flash is now free to use inside its platform, with calls no longer counting against plan quota. For people already working with coding agents, this is a concrete change: a native multimodal MoE model released on September 10 can now sit in the terminal for everyday code work at zero marginal cost, provided you accept the free tier's trimmed context limit.
The boundary of free is in the fine print: 200K versus 1M
The free tier's context window measures at 200K, against one million for the model's full version. The path of this promotion explains the shape: in September, OpenCode's related campaign first raised the model's allowance in the Go plan from $15 to a temporary $60, then to $80 on October 12, and has now switched it to free with no quota draw at all. On the other side, DeepSeek's own platform has no free API tier and remains prepaid pay-as-you-go. So this free access is a platform subsidy, not a price cut to the model itself; if the platform changes policy, the cost returns to the user's bill unchanged. Treating it as a window with an expiry feel, rather than a long-term price promise, is the safer planning assumption. Whether the tool itself deserves a long-term place was covered separately in the hands-on look at OpenCode's setup cost and configuration burden; today's change moves only the price line.
The model's base and community feedback: speed swings are the main complaint
DeepSeek V4.1 Flash uses a causal encoder-decoder design with a backbone of roughly 552B parameters plus about 196B of conditional memory parameters; of its 40 backbone layers, the first 20 form a causal encoder and the last 20 a decoder, and the official description puts its prefill compute at close to half. In real use, community discussion concentrates on unstable generation speed: some users saw peaks near 490 tokens per second on another cloud platform, others report drops to fifty or sixty per second from midday into the afternoon; inside OpenCode's plan, some report speeds falling below 100 per second, thinning further as context grows, to roughly 50 per second near the 200K mark. These are user measurements, not vendor promises, but they are enough to show that the free tier's experience discounts itself at peak hours and long contexts.
Worth switching now: everyday changes with clear boundaries that never need more than 200K of context at once, such as bug fixes, added tests and reading small-to-medium repositories. That work is insensitive to speed swings, and the free tier saves real money. Not worth a full migration: long refactors that need a large repository loaded in one context, and pair-programming setups with hard latency requirements. Keep those on a paid tier or the full 1M version, run the same task batch on the free tier as a comparison first, and only then decide the migration share.
