DeepSeek vs Kimi: V4.1 Flash and V4 Pro against K3.1 Pro

DeepSeek vs Kimi in November 2026 comes down to price against measured intelligence. DeepSeek's API costs a fraction of Moonshot's, while Moonshot's Kimi K3.1 Pro scores higher on Artificial Analysis. This comparison is based on the products' docs and published results as of November 4, 2026, not on our own hands-on testing. DeepSeek's prices are from its pricing page, Kimi's from Moonshot's pricing page.

The short version: pick DeepSeek when the bill matters most, and Kimi K3.1 Pro when you want the highest-scoring open-weight model and can pay for it.

DeepSeek vs Kimi at a glance

DeepSeek's API lists two current models, deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). Moonshot lists Kimi K3.1 Pro and Kimi K3, plus the older K2.7 Code and K2.6. Prices are per million tokens, in US dollars, as of November 4, 2026. DeepSeek charges half as much off-peak as at peak, so both are shown.

DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Kimi has one rate all day, with no tiering by context length according to Moonshot's quickstart.

Which is cheaper

DeepSeek, by a wide margin. On output tokens, which dominate the bill for reasoning models, Kimi K3.1 Pro at $10.00 costs about 5 times DeepSeek V4 Pro's off-peak $1.98 and about 2.5 times its peak $3.96. Against DeepSeek Flash at $0.60, it is about 17 times off-peak.

Cached input shows the biggest gap. A cache hit costs $0.20 on Kimi K3.1 Pro and $0.022 on DeepSeek V4 Pro off-peak. Agents that resend a long prompt on every turn lean on caching, so this matters more than the headline input price.

Kimi has one saving DeepSeek's page doesn't mention: Moonshot bills batch inference on K3 and K3.1 Pro at 60% of the standard price, for work that doesn't need a real-time answer. That brings K3.1 Pro's input to $1.20 and output to $6.00, still well above DeepSeek's list prices.

For a longer look at what DeepSeek's API costs and whether it has a free tier, see is the DeepSeek API free. Moonshot's rates are broken down in our Kimi K3.1 Pro pricing explainer.

Which scores higher

Kimi, on the one independent index we could read. Artificial Analysis lists Kimi K3.1 Pro at 46 on its Intelligence Index, the highest open-weights model on the page we opened on November 4, 2026. It lists DeepSeek V4.1 Flash 1017 (Max) at 33, second among open-weights models. The page showed no score for DeepSeek V4 Pro, so we can't compare the two flagships directly, and it carried no update date.

Moonshot's own numbers are less clear-cut, and they are the company's claims. Its K3.1 Pro blog post reports 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. On its internal xdev-bench-hard, Moonshot lists K3.1 Pro at 81.0 and DeepSeek-V4.1-Flash at 85.7, so DeepSeek's cheaper model comes out ahead on that test. Treat vendor-run benchmarks with care, and check whether a task like yours is in them.

Our view: a 13-point gap on one index is large, but Flash and K3.1 Pro are not the same class of model. DeepSeek's closer match to Kimi is V4 Pro, and it has no independent score on the page we read yet.

Features and openness

Both families offer a 1M-token context window and tool calls. DeepSeek supports JSON output, the Responses API, the Anthropic API format and, in beta, chat prefix completion and fill-in-the-middle (non-thinking mode only). Only Flash accepts images; V4 Pro does not.

DeepSeek-V4.1-Flash-1017 has MIT-licensed weights on Hugging Face with 763B parameters. Kimi K3.1 Pro has a Hugging Face card listing 2.8T total parameters, 104B active per token, a vision encoder and its own license rather than MIT. Read that license before you plan commercial self-hosting. We covered the weights plan in Kimi K3.1 Pro open weights, and DeepSeek's Flash model in DeepSeek V4.1 Flash explained.

DeepSeek's models offer non-thinking and thinking modes, with thinking as the default. Moonshot's benchmark notes say its K3.1 Pro results used a reasoning effort of "max".

Availability and what to watch

Moonshot's K2 models closed to new users on October 31, 2026 and are planned for retirement by the end of December 2026, so new projects should start on K3. DeepSeek's old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but are served by Flash at Flash prices. To start with DeepSeek, see how to get a DeepSeek API key.

Neither pricing page gives a change date, and both say prices may change. Moonshot has also published no independent benchmark for K3, and the Kimi pricing page does not state a maximum output length. Run your own task on both before committing.

How this comparison was made

This comparison uses the two companies' docs, pricing pages, model cards and the Artificial Analysis index, all read on November 4, 2026. We did not run the models ourselves.