Comparisons · both · deepseek vs qwen

Qwen, DeepSeek, GLM, Kimi: which Chinese model for which job

Published 16 Oct 2026. Updated 16 Oct 2026. Researched 16 Oct 2026.

There is no best Chinese model. There is a cheap default, a direct-API default, a preference-vote contender, and a couple of prices that are easy to quote wrong.

The shortlist. OpenRouter list price, blended 75% input / 25% output. Snapshot 2026-10-16 05:20 UTC. Comparators resolve to Anthropic: Claude Opus 5.5 and OpenAI: GPT-6.1 Sol.
Model	Blended	vs Opus	vs Sol
qwen3.8-flash	$0.230	97% cheaper	94% cheaper
glm-5.3-flash	$0.291	96% cheaper	93% cheaper
deepseek-v4.1-flash	$0.198	98% cheaper	95% cheaper
minimax-m3	$0.525	93% cheaper	87% cheaper
glm-5.3	$1.26	84% cheaper	69% cheaper
kimi-k3	$3.13	61% cheaper	22% cheaper
qwen3.8-max-prime	$6.00	25% cheaper	50% dearer

Job	Start here	Why
Bulk drafts, classification, cheap coding help	Qwen3.8 Flash or GLM-5.3 Flash	Lowest serious blend. GLM Flash matches the vendor page
You can schedule off-peak and you want DeepSeek’s own bill	deepseek-flash on DeepSeek, not the old OpenRouter id	Off-peak is half of peak
You want a coding subscription with a written cap	GLM Coding Plan Lite, then measure the week	Credits are published. Token conversion is not
The cheap model failed and output tokens are acceptable	Kimi K3 at Moonshot’s $3 / $15, or the max Qwen id	You are paying for the row on the preference board
Video	Kling or Hailuo, official sites	Not a token chart
Data cannot leave the machine	A small Qwen weight	Not K3’s full size

Style-controlled Bradley-Terry ratings from the LMArena leaderboard dataset. Not Elo. A higher rating means the model was preferred more often in blind votes. Intervals are the dataset's rating_lower and rating_upper. Ratings are rounded to 1 decimal for display in the chart; full values are stored. Category: coding. Published 2026-10-15, fetched 2026-10-16. Source: https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset (text_style_control, latest).
Entry	rating	Detail
Claude Opus 5.5 (High)	1547.52	rank 6, 1,894 votes, 1533.61–1561.44
gpt-6.1-sol-max	1539.79	rank 10, 1,257 votes, 1522.63–1556.95
Kimi K3 (Max)	1537.83	rank 12, 8,306 votes, 1530.58–1545.09
mimo-v2.6-pro	1537.58	rank 13, 1,105 votes, 1519.93–1555.23
glm-5.3-flash	1524.53	rank 27, 8,457 votes, 1517.07–1531.99
deepseek-v4.1-flash-max	1523.37	rank 29, 3,527 votes, 1513.1–1533.65
qwen3.8-max	1523.36	rank 30, 7,677 votes, 1515.79–1530.93
glm-5.3-max	1518.51	rank 38, 5,400 votes, 1509.88–1527.13
kimi-k2.6	1516.03	rank 44, 11,091 votes, 1509.15–1522.92
mimo-v2.6-flash	1512.93	rank 51, 2,625 votes, 1500.97–1524.89
deepseek-v4-pro	1500.58	rank 78, 17,136 votes, 1494.66–1506.51
qwen3.8-27b	1499.13	rank 81, 6,417 votes, 1491.03–1507.23
minimax-m3	1496.69	rank 87, 18,242 votes, 1490.99–1502.39
deepseek-v4-flash	1484.16	rank 105, 15,288 votes, 1477.95–1490.37
hunyuan-hy3-preview	1452.74	rank 153, 2,055 votes, 1439.33–1466.16

Resolved percentage on SWE-bench Verified, as published. Harnesses differ, so this is not a pure model-vs-model ranking. The current Opus and Sol aliases do not appear as rows on this board. Newer Chinese flagships (K3, GLM-5.3, DeepSeek V4) are also absent. Fetched 2026-10-16. Source: https://raw.githubusercontent.com/SWE-bench/swe-bench.github.io/master/data/leaderboards.json.
Entry	% resolved	Detail
Claude 4.5 Opus (high)	76.8	2026-02-17
MiniMax M2.5 (high)	75.8	2026-02-17
Claude 4.6 Opus	75.6	2026-02-17
GLM 5 (high)	72.8	2026-02-17
Lingxi v1.5 x Kimi K2	71.2	2025-10-14
Kimi K2.5 (high)	70.8	2026-02-17
DeepSeek V3.2 (high)	70	2026-02-17
OpenHands + Qwen3-Coder-480B-A35B-Instruct	69.6	2025-08-05
GLM-4.6	68.2	2025-09-30
OpenHands + Kimi K2	65.4	2025-07-16

Read those as incompatible instruments. LMArena is votes. SWE-bench is resolved issues under somebody’s harness, and it is behind the marketing names. K3, GLM-5.3, and DeepSeek V4 are not in the Verified extract. Saying “Kimi wins coding” because of Terminal-Bench on Moonshot’s card, and “DeepSeek wins coding” because of an old V3.2 SWE-bench row, is how both sentences end up in the same thread. They are different tests from different months.

The scatter is the other half of the matrix. Models on the right spend money when they talk.

Input and output list prices on OpenRouter, snapshot 2026-10-16 05:20 UTC. Axes on the chart are logarithmic. Grey points are Anthropic: Claude Opus 5.5 and OpenAI: GPT-6.1 Sol.
Model	Input	Output
qwen3.8-flash	$0.150	$0.470
glm-5.3-flash	$0.013	$1.13
deepseek-v4.1-flash	$0.036	$0.684
minimax-m3	$0.300	$1.20
kimi-k3	$0.304	$11.59
qwen3.8-max-prime	$4.00	$12.00
glm-5.3	$0.079	$4.80
claude-opus-5.5	$4.00	$20.00
gpt-6.1-sol	$2.00	$10.00

Questions

What is the cheapest serious option?

On the OpenRouter blend, the flash cluster. GLM-5.3 Flash also matches Z.AI's page. Confirm Qwen and Hunyuan on their own invoices before you call them official prices.

What should I avoid as a default?

Kimi K3 if you wanted cheap output, and the OpenRouter headline for GLM-5.3 if you will actually call Z.AI.

Sources

OpenRouter models API, accessed 2026-10-16
Z.AI pricing, accessed 2026-10-16
DeepSeek API pricing, accessed 2026-10-16
Kimi context caching, accessed 2026-10-16
LMArena leaderboard dataset, accessed 2026-10-16
SWE-bench leaderboards.json, accessed 2026-10-16