Comparisons · both · is kimi better than claude

Is Kimi better than Claude? A dated scorecard, not a winner

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

“Better” depends on the job and on which Claude you mean. This site’s comparator is whatever the build resolved as the latest Opus and the latest Sol. The paragraphs do not freeze those names. The charts do, because they print the snapshot.

Price first, because it is not close

Coding-agent session: 3 million input tokens, 90% served from cache, 60 thousand output tokens. Cache-write fees are not included. All rows use the OpenRouter list price. Comparators are OpenRouter list prices for Anthropic: Claude Opus 5.5 ($8.00/M blended) and OpenAI: GPT-6.1 Sol. Snapshot 2026-10-16. Cache-write fees are excluded. A missing cache price is billed as a full input token.
Model	Session	vs Opus	vs Sol	Cache price
kimi-k3	$1.35	54% cheaper	8% cheaper	published
qwen3.8-flash	$0.116	96% cheaper	92% cheaper	published
Anthropic: Claude Opus 5.5	$2.94	—	—	published
OpenAI: GPT-6.1 Sol	$1.47	—	—	published

The percentages in this table compare session cost. The price table's percentages compare blended token price. They answer different questions.

The task chart uses OpenRouter list prices, so the Kimi bar is the cheap-host fiction. Read the note above it before you believe the bar. The official output rate is $15 per million. A coding agent that emits a lot of tokens will not feel like a discount product. Qwen’s flash id is the discount product.

Preference data

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

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: overall. 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)	1508.17	rank 3, 8,300 votes, 1500.84–1515.5
Kimi K3 (Max)	1487.07	rank 18, 32,795 votes, 1482.49–1491.66
gpt-6.1-sol-max	1483.85	rank 22, 6,320 votes, 1475.71–1491.98
qwen3.8-max	1482.35	rank 24, 27,766 votes, 1477.38–1487.33
mimo-v2.6-pro	1479.19	rank 27, 4,075 votes, 1469.83–1488.56
glm-5.3-max	1478.54	rank 28, 22,543 votes, 1473.31–1483.77
deepseek-v4.1-flash-max	1474.63	rank 38, 12,793 votes, 1468.41–1480.86
glm-5.3-flash	1473.84	rank 41, 32,925 votes, 1468.97–1478.71
kimi-k2.6	1460.83	rank 59, 40,013 votes, 1456.4–1465.26
deepseek-v4-pro	1457.79	rank 64, 57,594 votes, 1453.89–1461.69
mimo-v2.6-flash	1450.93	rank 77, 10,917 votes, 1444.44–1457.42
minimax-m3	1441.32	rank 96, 63,973 votes, 1437.48–1445.16
qwen3.8-27b	1437.75	rank 100, 23,294 votes, 1432.64–1442.86
deepseek-v4-flash	1435.65	rank 104, 51,987 votes, 1431.61–1439.68
hunyuan-hy3-preview	1409.13	rank 152, 6,879 votes, 1401.72–1416.54

Kimi K3 (Max) is in the top cluster on both cuts of the 15 Oct 2026 publish. The current Opus row is ahead. The intervals in the detail column are wide enough that a one-rank gap is not a product decision. These are blind votes, mostly short conversations people chose to send. They are not a private-repo agent eval.

Terminal-Bench is Moonshot’s table

Scores copied from the Moonshot model card. Kimi K3 is evaluated with the Kimi Code harness. Other models use the best score across harnesses, as the card states: GLM-5.2 with Claude Code, Claude Opus 4.8 and Claude Fable 5 with Terminus 2, GPT-5.5 and GPT-5.6 Sol with Codex. These are not the current Opus alias, and the harnesses are not the same, so the gaps are not a clean comparison. Fetched 2026-10-16. Source: https://huggingface.co/moonshotai/Kimi-K3.
Entry	score	Detail
Kimi K3 (max)	88.3	Kimi Code
Claude Fable 5 (max, w/ fallback)	88	Terminus 2
GPT-5.6 Sol (max)	88.8	Codex
Claude Opus 4.8 (max)	84.6	Terminus 2
GPT-5.5 (xhigh)	83.4	Codex
GLM-5.2 (max)	82.7	Claude Code

Use it as a vendor table with a harness footnote. Do not import it into a sentence that says Kimi beat “Claude” without naming the harness and the fact that the card’s Claude rows are not the current alias.

For what tasks

Job	Lean
Cheap coding loop, boilerplate, logs	A flash model, not K3
Long context you are willing to pay output for	K3, after a short paid test
Blind chat preference close to the frontier	K3 is in the cluster; it is not the top row
Data that cannot leave a regulated laptop	Neither, until you self-host weights you are allowed to run

SWE-bench Verified, in the file we fetched, has Kimi K2.5 and Kimi K2 rows and no K3 row. Absence is the finding.

Questions

Is Kimi better than Claude at coding?

On the LMArena coding snapshot from 15 Oct 2026, the Kimi K3 row sits behind the current Opus row. On Moonshot's own Terminal-Bench table the harnesses differ. Neither result is a universal win.

Is Kimi cheaper than Claude?

On Moonshot's published rate, output is $15 per million tokens. The blended official price is not a flash-tier discount. See the price note in the chart.

Sources

LMArena leaderboard dataset, accessed 2026-10-16
Kimi K3 model card, accessed 2026-10-16
Kimi context caching, accessed 2026-10-16
OpenRouter models API, accessed 2026-10-16