Pointing OpenCode at DeepSeek is an endpoint change. The API answers in both OpenAI ChatCompletions and Anthropic formats, so there is no adapter to maintain [11]. The same writeup routes some workflows through OpenRouter instead [12], and that is the more useful detail: if the provider is swappable at the URL, price becomes a term of business rather than a property of your stack. Switching between Pro and Flash mid-session is one command, no restart [13].
Now the arithmetic, because it does not close. At the listed $0.87 per million output tokens for V4-Pro [6], a dollar buys about 1.15 million output tokens. The post says 500K off-peak [9], which is 2.3x away from its own rate card [2]. Caching has the same problem: cache-hit input at $0.0036 against $0.435 standard is a 99.2 percent reduction [3], not the roughly 88 percent quoted [10]. Either the peak and off-peak tiers introduced in mid-August [20] are doing work the post does not show, or the figures come from different snapshots. The post also prices a full day of Pro-driven development at less than a coffee [17]; the coffee is unspecified, which is about the precision on offer here. Take the direction. Against Claude Opus output at $25 per million [8], list-to-list on output is roughly 29x [1], and that survives a 2x error in either direction.
The benchmark needs the same handling. V4-Flash at 69.1 on the Artificial Analysis coding index, 29th of nearly 200 models tracked [16], also means 28 models score higher [6], and the ones you are thinking of replacing are plausibly among them. The author's own scope is narrower than the headline score: boilerplate, CRUD logic, component generation and API routes, judged indistinguishable in blind use [14]. He also generated a complete Astro template unattended with DeepSeek alone [15]. That number transfers to the extent your week looks like that list. For long refactors he rates the million-token default context as mattering more than raw intelligence scores [2].
Where it stops is named in the post, and named honestly. Agentic capability trails coding capability; complex autonomous multi-step work needs tighter task boundaries and more validation than Claude Opus or GPT-5.5, and long, loosely defined runs drift [18]. Vague prompts draw more assumptions than Claude would make [21]. Individual tool calls and shell commands held up [22]. So the failure mode is not a bad edit. It is a confident wrong direction sustained across twenty steps, which is the expensive kind.
Serving cost tracks active parameters, not total. Pro fires 49B of 1.6T per token, about 3.1 percent; Flash 13B of 284B, about 4.6 percent [3][4][4]. That is why a 1.6T-parameter model lists at 3.1x its smaller sibling's output rate rather than a hundred times [5].
The MIT license [1] is what makes this a budget decision rather than a bet. DeepSeek has signalled a broader increase without publishing rates [20]; when it lands, the weights you validated against are still downloadable and someone else can serve them. Which is why I would keep the router abstraction in place even after one provider wins on price.
