DeepSeek vs Kimi vs Qwen API Pricing for Agent Workloads
Route DeepSeek, Kimi, and Qwen by official per-million-token prices, peak windows, cache rules, and region constraints—with failure modes and a fixed-prompt load test.
The spine remains DeepSeek vs Kimi vs Qwen API pricing, context window, and agent workload routing.
This DeepSeek vs Kimi vs Qwen API pricing guide helps cost-sensitive developers choose a default model route for agent and Chinese-language workloads—not another leaderboard post.
When you ship agents or Chinese apps, social rankings are a distraction. Three questions matter: whether official per-million-token prices survive your volume; whether context and cache rules match the job; and whether region plus billing can actually land. This comparison uses official pricing pages observed on 2026-09-20 to route DeepSeek, Kimi, and Qwen (Tongyi Qianwen).
Cost-first general chat and tool calls: start with DeepSeek
DeepSeek’s official pricing page lists deepseek-flash and deepseek-v4-pro per 1M tokens, with peak/off-peak and cache-hit rates. For example, flash off-peak cache-miss input is about $0.15 and output about $0.60; peak rates are roughly double. Pro is higher off-peak and carries a lower concurrency limit. Peak hours are documented as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; all other hours are off-peak. That fits pushing volume inside a budget if you can account for UTC peak windows.
Do not read input price alone. Tool calls, retries, and longer answers push the bill toward output. If most jobs are short and your system prompt caches well, flash plus cache hits is usually the cheapest controllable tier among the three.
Ultra-long context and flagship reasoning: look at Kimi
Kimi’s docs list kimi-k3 at $3.00 input and $15.00 output per 1M tokens, with cached input at $0.30 and roughly a 1,048,576-token context, plus cache-write charges and 5-minute / 1-hour TTL tiers. Unit price is higher than DeepSeek’s flash tier, but long-document and long-horizon coding work often lands here. K2-series models are cheaper with about 262K context and can serve as a downgrade sidecar—keep flagship and flash rows on separate budget lines.
Write “how much source text must fit in one request” as a hard constraint. If most calls stay far below 100K tokens, paying for a million-token window is usually waste.
Alibaba Cloud ecosystem and regional quotas: look at Qwen
Qwen is billed through Alibaba Cloud Model Studio / DashScope. International docs list tiers such as qwen3.7-max around $2.5 input / $7.5 output per 1M tokens, and newer qwen3.8-max international around $2 / $6, with notes that some free quotas are region-limited. China (Beijing) list prices for equivalent classes are often lower, but endpoints, invoicing, and compliance differ. If you already run on Alibaba Cloud and need invoices plus same-cloud networking, integration cost can beat “cheapest absolute model.”
Region mismatch is the common failure: treating a Singapore free quota as global, or calling an enterprise model on the wrong regional endpoint, will falsify your load-test numbers.
When not to choose each option
Do not default to Kimi only because social posts call it the strongest reasoner—if tasks are short, cacheable, and budget-tight, output price will erase the advantage. Do not assume DeepSeek flash stays cheap if you ignore peak multipliers and concurrency caps. Do not pick Qwen only for domestic invoicing while forcing international-compliance traffic onto the wrong region. If contracts lock you to one cloud or your team can maintain only one key store, write those constraints before debating model names.
Load-test checklist
Fix the same prompt, the same tool schema, and the same success criteria. Run at least three loads: short chat, a job with two or three tool calls, and a long-document summary. Record input/output tokens, latency, error rate, and whether peak pricing applied. Convert cost per 1,000 real jobs, not per single completion. Confirm regional endpoints, content-policy fit, and invoice path. Freeze the model that stays stable inside budget; keep others as downgrade/upgrade sidecars in config, not in hallway chat.
Three ledgers before you declare a winner
Keep three ledgers. Ledger one is unit price: input, output, cache hit, cache write if any, and peak multipliers. Ledger two is task shape: average input tokens, average output tokens, tool-call count, and retry rate. Ledger three is region and invoice: endpoint, currency, tax invoice path, and content-policy posture. If the ledgers do not align, you are not comparing prices—you are comparing incompatible constraints.
Agent workloads amplify the gap. A real job may include a planning turn, tool-result refill, and a final answer. Pricing only a single completion systematically undercounts output and retries. In the load-test script, define success as an executable result, not the model saying “done.”
Teams that run both Chinese support and English engineering chat can keep dual defaults: high-volume Chinese on DeepSeek flash, long-document knowledge work on Kimi, and Alibaba-internal batch jobs on Qwen. Dual defaults only work if a config center can switch by route key—not if engineers remember to rename the model in chat.
Current takeaway and defaults
Budget-sensitive general chat and tool calls: default DeepSeek flash and account for peak hours. Ultra-long context or flagship reasoning with budget headroom: default Kimi k3, with K2 as a cheaper sidecar for shorter work. Strong Alibaba Cloud binding with regional billing needs: default Qwen and price the actual region. Agent workloads also burn tokens on tool schemas, retries, and reasoning traces—price a full real task and set truncation/summarization so API pricing wins are not eaten by useless context.
Decision tree you can paste into an architecture doc
- Do jobs regularly exceed ~100K input tokens of source text? Yes → evaluate Kimi k3 first; No → continue.
- Must you stay on Alibaba Cloud networking or a domestic invoice loop? Yes → prefer Qwen and price the region; No → continue.
- Is the load mostly high-concurrency short jobs under a tight budget? Yes → default DeepSeek flash with peak alerts; No → load-test flash/pro versus Kimi against a quality gate.
- Does compliance ban a vendor? Yes → remove it from the shortlist before politics reopen it after testing.
Write these four steps as a “model routing” section. Prices change; the tree should point to a dated re-check of official pricing pages rather than freezing one screenshot forever.