
Model comparison
DeepSeek V4 Pro vs Kimi K3 API price, context and cost.
DeepSeek V4 Pro (DeepSeek) and Kimi K3 (Moonshot) side by side: per-1M prices, cached-input rates, context, output limits and capability flags from the live RouterPlex catalog, with three worked costs. A coding-agent session costs $0.90 on DeepSeek V4 Pro and $3.82 on Kimi K3. Both run on one prepaid key with the same OpenAI-compatible request.
Side by side
Catalog fields, one row each.
| Catalog field | DeepSeek V4 Pro | Kimi K3 |
|---|---|---|
| Provider | DeepSeek | Moonshot |
| Model ID | deepseek-v4-pro | kimi-k3 |
| Input / 1M | $0.80 | $3.00 |
| Output / 1M | $2.39 | $15.00 |
| Cached input / 1M | $0.21 | $0.70 |
| Context window | 1M (1,000,000 tokens) | 1M (1,048,576 tokens) |
| Max output | 384K tokens | 131K tokens |
| Image input | No | Yes |
| Tool calling | Yes | Yes |
| Reasoning | Yes | Yes |
| Long-prompt tier | None | None |
| Price basis | Vendor list price, 0% markup | Vendor list price, 0% markup |
- DeepSeek V4 Pro: DeepSeek prices this model by time of day: $0.66 input / $1.98 output per 1M tokens off-peak, and double that during peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday). RouterPlex bills a single flat rate of $0.80 / $2.39 in between -- the time-weighted average of the two -- so you are never billed by the clock. Cached prompt tokens are billed at $0.21 per 1M, set from what our upstream charges us rather than DeepSeek's cache-hit rate.
- Kimi K3: RouterPlex bills $3 input / $15 output per 1M tokens, Moonshot's list price. Cached prompt tokens are billed at $0.70 per 1M, not Moonshot's published $0.30 cache-hit rate: our cache price is set from what our upstream charges us for a cached token, so it does not follow the lab's rate.
What the catalog says
DeepSeek V4 Pro is 73% cheaper on input ($0.80 against $3.00 per 1M tokens). DeepSeek V4 Pro is 84% cheaper on output ($2.39 against $15.00 per 1M tokens). DeepSeek V4 Pro is 70% cheaper on cached input ($0.21 against $0.70 per 1M tokens). On the coding-agent session below, DeepSeek V4 Pro costs $0.90 against $3.82 for Kimi K3. Both accept a 1M context window. DeepSeek V4 Pro can return up to 384K tokens per response, against 131K for Kimi K3. Only Kimi K3 lists image input in the catalog; DeepSeek V4 Pro does not.
Generated from catalog prices, limits and capability flags. Price does not measure output quality.
Published benchmarks
Where real data exists.
Raw score as the source publishes it, percentile in brackets. Only rows with a cited source are shown; "Not listed" means the source has no entry for that model.
| Source / category | DeepSeek V4 Pro | Kimi K3 |
|---|---|---|
| overallLMArena Text Leaderboard | 1457.8 (87) | Not listed |
| codingLMArena Text Leaderboard | 1501.8 (84) | Not listed |
| reasoningLMArena Text Leaderboard | 1445.1 (82) | Not listed |
| writingLMArena Text Leaderboard | 1444.5 (89) | Not listed |
| generalLMArena Text Leaderboard | 1453.2 (88) | Not listed |
LMArena Text Leaderboard, version 2026-09-02, retrieved 2026-09-10. A leaderboard measures its own prompt set, not your workload.
Worked cost examples
Same workload, both bills.
Each line multiplies the catalog rate per 1M tokens by the tokens shown. Cached tokens use the route's cached-input rate, or the full input rate when it has no cache discount. Reasoning tokens bill as output, so a model that thinks at length raises the output line beyond these figures.
40 requests of 50K prompt tokens, 80% of each prompt a cache hit, 2.5K output each.
0.4M input x $0.80 + 1.6M cached x $0.21 + 0.1M output x $2.39 = $0.90
0.4M input x $3.00 + 1.6M cached x $0.70 + 0.1M output x $15.00 = $3.82
10,000 conversations of 2K prompt tokens and 500 output tokens, no caching.
20M input x $0.80 + 5M output x $2.39 = $27.95
20M input x $3.00 + 5M output x $15.00 = $135
200 requests of 150K prompt tokens and 2K output tokens, no caching.
30M input x $0.80 + 0.4M output x $2.39 = $24.96
30M input x $3.00 + 0.4M output x $15.00 = $96.00
Switching
One key, one line.
Both models sit behind the same RouterPlex key and the same OpenAI-compatible endpoint. Keep the request shape and change the model string. Claude Code and the Anthropic SDK use the same ids on https://api.routerplex.com.
Read the quickstartimport osfrom openai import OpenAIclient = OpenAI(base_url="https://api.routerplex.com/v1",api_key=os.environ["ROUTERPLEX_API_KEY"],)reply = client.chat.completions.create(model="deepseek-v4-pro", # switch: "kimi-k3"messages=[{"role": "user", "content": "Hello!"}],)print(reply.choices[0].message.content)
# DeepSeek V4 Procurl https://api.routerplex.com/v1/chat/completions \-H "Authorization: Bearer $ROUTERPLEX_API_KEY" \-H "Content-Type: application/json" \-d '{"model": "deepseek-v4-pro", "messages": [{"role": "user", "content": "Hello!"}]}'# Kimi K3: same request, one field changedcurl https://api.routerplex.com/v1/chat/completions \-H "Authorization: Bearer $ROUTERPLEX_API_KEY" \-H "Content-Type: application/json" \-d '{"model": "kimi-k3", "messages": [{"role": "user", "content": "Hello!"}]}'
Questions
Before you pick one of the two.
01Is DeepSeek V4 Pro cheaper than Kimi K3?+
DeepSeek V4 Pro costs $0.80 input and $2.39 output per 1M tokens on RouterPlex; Kimi K3 costs $3.00 and $15.00. On a coding-agent session (40 requests of 50K prompt tokens, 80% of each prompt a cache hit, 2.5K output each) DeepSeek V4 Pro comes to $0.90 and Kimi K3 to $3.82, so DeepSeek V4 Pro is cheaper for that workload. A chat app month (10,000 conversations of 2K prompt tokens and 500 output tokens, no caching) costs $27.95 on DeepSeek V4 Pro and $135 on Kimi K3.
02Which has the larger context window, DeepSeek V4 Pro or Kimi K3?+
They match: both accept 1M tokens of context on RouterPlex. Maximum output is 384K for DeepSeek V4 Pro and 131K for Kimi K3.
03Do DeepSeek V4 Pro and Kimi K3 support image input and tool calling?+
Per the RouterPlex catalog, DeepSeek V4 Pro: tool calling, reasoning. Kimi K3: image input, tool calling, reasoning. These flags come from the labs' published model metadata, not from a quality test.
04How do I switch from DeepSeek V4 Pro to Kimi K3?+
Change one string. Both run on the same RouterPlex key and the same OpenAI-compatible endpoint (https://api.routerplex.com/v1), so the request stays identical and only model changes from "deepseek-v4-pro" to "kimi-k3". The same ids work on the Anthropic-compatible /v1/messages endpoint for Claude Code.
05What do benchmarks say about DeepSeek V4 Pro vs Kimi K3?+
LMArena Text Leaderboard (2026-09-02, retrieved 2026-09-10) lists an overall score of 1457.8 for DeepSeek V4 Pro and no entry for Kimi K3. A public leaderboard measures its own prompt set; test both on your prompts before moving traffic.
06When is Kimi K3 worth the higher price?+
When the task needs image input, which Kimi K3 offers and DeepSeek V4 Pro does not per the catalog. Otherwise the reason to pay more is output you prefer on your own prompts; this page does not rank quality, so run both in the RouterPlex Arena before moving traffic.
Provider pricing
Try both
Both models behind one key.
Top up from $5 with no subscription, then send the same prompt to each.