DeepSeek is one of the best value models you can put behind an AI assistant today: a 1M token context window, thinking mode, tool calling, JSON output and vision on the Flash model, at off-peak prices from $0.15 per million input tokens. Very little else at this price comes close.
The catch is that the model name on the API does not tell you who is actually running it or how. DeepSeek publishes its weights, which is good for the ecosystem — and it is also why the same model can be served in very different ways.
Why we recommend the official DeepSeek API
Five reasons, and they all come back to the same thing: on the official API, the model runs the way its authors built it.
Its sharper. To serve DeepSeek more cheaply, most providers run a compressed version of the weights. DeepSeek releases DeepSeek-V4 in FP8 mixed precision, with FP4 used for the MoE expert parameters and FP8 for most of the rest. Many providers go further and convert the model down to 4-bit formats such as W4A16, INT4, NVFP4 and MXFP4. That cuts memory and compute across the whole model, and it costs accuracy — which is why the same model can feel noticeably sharper in one place than another. On the official API you get the model as DeepSeek built and released it, not a further-compressed conversion.
It is faster. The official API runs on the infrastructure DeepSeek designed and tuned for this exact model, instead of a general-purpose stack that someone else adapted afterwards. Less adaptation means less overhead between your question and the answer.
Tools run correctly. DeepSeek's thinking mode follows a strict protocol, and not every layer in between handles it properly. Elsewhere, the result is an endpoint that answers a simple question fine but slows down or breaks partway through a multi-turn tool call. On the official API, tools do what you asked, in the order you asked for.
The caching is better. This is the part that decides your bill. The official price list, per 1M tokens at off-peak rates:
- Flash — input $0.003 on a cache hit, $0.15 on a cache miss; output $0.60.
- V4 Pro — input $0.022 on a cache hit, $0.66 on a cache miss; output $1.98.
A cached input token costs about 2% of an uncached one, so the real cost of a long conversation is governed by how much of it can be reused rather than by the headline rate. Caching is best-effort — DeepSeek states it does not guarantee a 100% cache hit rate — so an endpoint that reuses less of your conversation charges you more for tokens that never appear on a price page.
The real cost is lower. Off-peak input starts at $0.15 per million tokens, and peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday) are exactly double — every other hour is off-peak. Combined with the cache rate above, the effective cost of everyday work on the official API is lower than it looks, even next to a provider advertising a cheaper number per token.
That is why, in Ceres, we recommend pointing your API key at the official DeepSeek API rather than a third-party provider. The list price may look slightly higher, but the cost of real work is lower, and the behaviour is the one every benchmark and every guide is written against.
Using DeepSeek with Ceres
Ceres does not bundle a model. You bring your own API key and choose the provider yourself — and with DeepSeek on the official API you get capable, fast, genuinely cheap assistance for day-to-day coding, writing and document work.
Try DeepSeek in Ceres wherever you already work:
- In your browser — Ceres Browser on the Chrome Web Store.
- In VS Code — Ceres: AI Coding Assistant with Voice on the Visual Studio Marketplace.
- In JetBrains IDEs — Ceres on the JetBrains Marketplace.
Try Ceres Assistant with your official DeepSeek API key and see what it costs.
Header image: "Energy Efficient Processors" by Intel Free Press, CC BY 2.0, via Wikimedia Commons.