Alibaba's flagship and Moonshot's previous-generation flagship both offer about a million tokens of context. Qwen3.8-Max is cheaper on every listed rate; Kimi K3's open weights are the main reason to pay more.
Qwen3.8-Max is a third cheaper than Kimi K3 on input and 2.5× cheaper on output, and its cache hits cost about half as much. Kimi K3 is an open-weights model, which matters if you may later self-host or switch inference providers.
| Qwen3.8-Max | Kimi K3 | |
|---|---|---|
| Provider | Alibaba | Moonshot AI |
| Tier | Flagship | Flagship |
| Input / 1M tokens | $2.00 | $3.00 |
| Output / 1M tokens | $6.00 | $15.00 |
| Cache hit / 1M | $0.17 | $0.30 |
| Cache write / 1M | — | $3.00 |
| Context window | 1M | 1M |
| Notes | New Alibaba flagship (Aug 3, 2026); 2.4T MoE | Previous-generation flagship, superseded by Kimi K3.1 Pro but still sold; open weights; cache write $3 (5-min TTL) or $6 (1-h TTL) |
Qwen3.8-Max is 33% cheaper on input tokens ($2.00 vs $3.00 per 1M). Qwen3.8-Max is 2.5× cheaper on output tokens ($6.00 vs $15.00 per 1M).
| Workload | Qwen3.8-Max | Kimi K3 | Difference |
|---|---|---|---|
| Support chatbot 10,000 conversations × (2,000 input + 400 output tokens) | $64.00 | $120 | Qwen3.8-Max saves $56.00 |
| Document Q&A 1,000 questions × (30,000 input + 800 output tokens) | $64.80 | $102 | Qwen3.8-Max saves $37.20 |
| Coding agent session 4M input tokens (90% cache hits) + 150,000 output tokens | $2.31 | $4.53 | Qwen3.8-Max saves $2.22 |
Standard rates per 1M tokens from data on this site. The coding-agent row assumes 90% of input is read from the prompt cache (full input price where a model has no cache rate). Cache-write fees, batch discounts and long-context surcharges listed in the notes are not included. Updated November 2, 2026.
you want the lower API bill, especially on output-heavy or cache-heavy work, and are comfortable running on Alibaba Cloud.
you value open weights — the option to self-host, fine-tune or move to another inference provider — and want Moonshot's batch inference at 60% of standard price.
The input gap is modest ($2.00 vs $3.00), but output is $6.00 vs $15.00. Agentic and reasoning workloads that produce long chains of output feel the bigger ratio. Alibaba bills Qwen's thinking and non-thinking modes at the same per-token price.
Both apply cache pricing to repeated prefixes. Moonshot bills Kimi K3 cache writes separately by TTL (5 minutes or 1 hour); Alibaba points to its Context Cache documentation for Qwen's cache rules. For overnight jobs, Moonshot lists kimi-k3 batch inference at 60% of standard price; Alibaba's batch rate is 50% of real-time price for supported models, but its pricing page shows that discount for qwen3.8-max only in the China (Beijing) region, and batch and cache discounts can't be combined.
Moonshot lists 1,048,576 tokens for the K3 series; Alibaba prices qwen3.8-max as a single tier up to 1M input tokens. Alibaba also lists a 1M-token free quota for qwen3.8-max in its Singapore region.
Qwen3.8-Max, on input ($2.00 vs $3.00), output ($6.00 vs $15.00) and cache hits ($0.17 vs $0.30).
Yes. That is the main structural difference: you can run Kimi K3 outside Moonshot's API.
Moonshot lists kimi-k3 batch inference at 60% of standard price. Alibaba lists 50% batch pricing for qwen3.8-max only in its China (Beijing) region; the international (Singapore) table shows no batch rate.
How we choose: prices come from each vendor's official pricing page and are re-checked every week. We take no affiliate fees, so rankings aren't for sale.