Math
Like-for-like- Kimi K2.6
- 67.1
- Kimi K2.5
- 60.6
- Weighted basis
- 4 vs 4 rows
- Reading
- Kimi K2.6 leads
Kimi K2.6 vs Kimi K2.5
Updated August 27, 2026. Public scores include evidence status and uncertainty. They are not guarantees for a specific workload.
Decision reading
Kimi K2.6 has the higher public score estimate, 60.14 versus 59.01, but the 90% score intervals overlap. Treat that as a lead, not a settled winner.
23 results are shared. Category rows based on different benchmark sets are marked directional and do not name a winner.
Which one for your work
Recommendations appear only when a shared evidence basis or an explicit operating constraint supports the call. Secondary and unsupported use cases stay disclosed below the initial list.
Chat turn cost
1K fresh input + 500 output tokens
Kimi K2.5
Kimi K2.5 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.
Confidence: listed-rates
Cache-heavy agent loop cost
200K cached + 20K fresh input + 10K output tokens
Kimi K2.5
Kimi K2.5 has the lower estimated token cost for this stated workload. Kimi K2.6 has no published cached-input rate, so cached tokens use its listed input rate. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Confidence: rate-fallback
Show secondary and unsupported calls
Repository review cost
50K fresh input + 3K output tokens
Kimi K2.5
Kimi K2.5 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.
Confidence: listed-rates
Coding work
Code generation, repair, and software-engineering tasks
Not enough matched evidence
The category averages use different weighted benchmark sets, so they are directional rather than like-for-like.
Confidence: limited
Agentic work
Tool use, computer use, and multi-step task completion
Not enough matched evidence
The category averages use different weighted benchmark sets, so they are directional rather than like-for-like.
Confidence: limited
Long documents
Prompts that approach the documented context limit
No clear pick
The documented context windows are equal.
Confidence: documented
What is actually comparable
Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.
9 Kimi K2.6 only
23 Shared
22 Kimi K2.5 only
- Shared results
- 23
- Kimi K2.6 only
- 9
- Kimi K2.5 only
- 22
- Like-for-like categories
- 1 / 8
4 categories use different evidence sets. Those rows remain visible for coverage context but do not name a winner.
Category results, on a stated basis
Each row states whether both averages use the same weighted benchmark set. Directional and not-comparable rows remain visible, but they never receive a winner in this template.
Agentic
Directional only- Kimi K2.6
- 73.5
- Kimi K2.5
- 55.0
- Weighted basis
- 3 vs 2 rows
- Reading
- Directional only
Coding
Directional only- Kimi K2.6
- 64.4
- Kimi K2.5
- 59.4
- Weighted basis
- 3 vs 4 rows
- Reading
- Directional only
Knowledge
Directional only- Kimi K2.6
- 42.2
- Kimi K2.5
- 56.9
- Weighted basis
- 2 vs 4 rows
- Reading
- Directional only
Multimodal
Directional only- Kimi K2.6
- 79.8
- Kimi K2.5
- 78.5
- Weighted basis
- 2 vs 1 rows
- Reading
- Directional only
Reasoning
Not comparable- Kimi K2.6
- Not measured
- Kimi K2.5
- 61.0
- Weighted basis
- 0 vs 1 rows
- Reading
- Not comparable
Multilingual
Not comparable- Kimi K2.6
- Not measured
- Kimi K2.5
- 82.3
- Weighted basis
- 0 vs 1 rows
- Reading
- Not comparable
Instruction following
Not comparable- Kimi K2.6
- Not measured
- Kimi K2.5
- 93.9
- Weighted basis
- 0 vs 1 rows
- Reading
- Not comparable
| Category | Kimi K2.6 | Kimi K2.5 | Weighted basis | Reading |
|---|---|---|---|---|
| Math | 67.1 | 60.6 | Like-for-like4 vs 4 rows | Kimi K2.6 leads |
| Agentic | 73.5 | 55.0 | Directional only3 vs 2 rows | Directional only |
| Coding | 64.4 | 59.4 | Directional only3 vs 4 rows | Directional only |
| Knowledge | 42.2 | 56.9 | Directional only2 vs 4 rows | Directional only |
| Multimodal | 79.8 | 78.5 | Directional only2 vs 1 rows | Directional only |
| Reasoning | Not measured | 61.0 | Not comparable0 vs 1 rows | Not comparable |
| Multilingual | Not measured | 82.3 | Not comparable0 vs 1 rows | Not comparable |
| Instruction following | Not measured | 93.9 | Not comparable0 vs 1 rows | Not comparable |
Shape of the matched evidence
Only shared public evidence is shown. Sparse evidence stays a ruled list rather than being closed into a radar shape.
Too few matched category axes support a radar. The ruled list below shows only shared benchmark results; positions use each benchmark’s normalized display scale when available.
BrowseComp
Agentic
Terminal-Bench 2.0
Agentic
FrontierMath v2 (Tiers 1-3)
Math
FrontierMath v2 (Tier 4)
Math
SWE-bench Pro
Coding
What each workload costs
Three fixed token mixes turn per-token rates into comparable decisions. Each scenario states context fit and whether cached input had to fall back to the published list-input rate.
Chat turn
1K fresh input + 500 output tokens
- Kimi K2.6
- $0.00295
- Fits in one request
- Kimi K2.5
- $0.0021
- Fits in one request
Kimi K2.5 has the lower modeled cost
Costs use the listed standard API rates.
Repository review
50K fresh input + 3K output tokens
- Kimi K2.6
- $0.0595
- Fits in one request
- Kimi K2.5
- $0.039
- Fits in one request
Kimi K2.5 has the lower modeled cost
Costs use the listed standard API rates.
Cache-heavy agent loop
200K cached + 20K fresh input + 10K output tokens
- Kimi K2.6
- $0.249
- Fits in one request
- Cached input priced at the published list-input rate
- Kimi K2.5
- $0.162
- Fits in one request
- Cached input priced at the published list-input rate
Kimi K2.5 has the lower modeled cost
Kimi K2.6 has no published cached-input rate, so cached tokens use its listed input rate. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Specification differences
Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.
SpecificationKimi K2.6Kimi K2.5
Context window
Maximum documented context; output-token limits may be lower.
Kimi K2.6
256K
Kimi K2.5
256K
API model ID
Kimi K2.6
Not sourced
Kimi K2.5
Not sourced
Cached-input rate
A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.
Kimi K2.6
Not published
Kimi K2.5
Not published
Documented inputs
Kimi K2.6
Not sourced
Kimi K2.5
Not sourced
Documented outputs
Kimi K2.6
Not sourced
Kimi K2.5
Not sourced
Provider availability
Kimi K2.6
Not sourced
Kimi K2.5
Not sourced
Reasoning profile
Kimi K2.6
Reasoning
Kimi K2.5
Non-Reasoning
Weight access
Kimi K2.6
Open Weight
Kimi K2.5
Open Weight
License
Kimi K2.6
Open Weight
Kimi K2.5
Open Weight
Release date
Kimi K2.6
2026-04-20
Kimi K2.5
2026-02-01
If you are considering the documented upgrade path
- Deployment change
- Both entries list Moonshot AI as the provider. Confirm endpoint, model ID, limits, and feature support before switching.
- Quality signal
- Kimi K2.6 has the higher public score estimate, 60.14 versus 59.01, but the 90% score intervals overlap.
- Workload cost
- Repository review: $0.0595 vs $0.039. Cache-heavy agent loop: $0.249 vs $0.162.
- Context tradeoff
- Both models list 256K.
Run the same representative tasks against both endpoints before changing production traffic.
Self-host vs API cost
Estimates at 50,000 req/day · 1000 tokens/req average.
Kimi K2.6
API / mo$3,713
Self-host / mo$18,221
Break-even326M/day
Kimi K2.5
API / mo$2,700
Self-host / mo$5,221
Break-even132M/day
Benchmark evidence
The full public result ledger is available for audit without forcing a wide desktop table onto a phone.
Browse raw public benchmark evidence54 rows
Agentic
Terminal-Bench 2.0
Kimi K2.666.7%SourceKimi K2.550.8%SourceKimi K2.6 leads this result
BrowseComp
Kimi K2.683.2%SourceKimi K2.560.6%SourceKimi K2.6 leads this result
OSWorld-Verified
Kimi K2.673.1%SourceKimi K2.5—Not directly comparable
Toolathlon
Kimi K2.650%SourceKimi K2.527.8%SourceKimi K2.6 leads this result
MCP Atlas
Kimi K2.655.9%SourceKimi K2.529.5%SourceKimi K2.6 leads this result
Claw-Eval
Shared sourceKimi K2.662.3%Kimi K2.552.3%Kimi K2.6 leads this result
DeepSearchQA
Kimi K2.692.5%SourceKimi K2.577.1%SourceKimi K2.6 leads this result
WideResearch
Kimi K2.680.8%SourceKimi K2.572.7%SourceKimi K2.6 leads this result
Gert Labs
Shared sourceKimi K2.656.82%Kimi K2.545.88%Kimi K2.6 leads this result
ResearchClawBench
Shared sourceKimi K2.618.0%Kimi K2.514.0%Kimi K2.6 leads this result
OSWorld 2.0
Kimi K2.64.6%SourceKimi K2.5—Not directly comparable
QwenClawBench
Kimi K2.6—Kimi K2.554.3%SourceNot directly comparable
τ³-bench results
Kimi K2.6—Kimi K2.565.7%SourceNot directly comparable
DeepPlanning
Kimi K2.6—Kimi K2.514.4%SourceNot directly comparable
MCP-Tasks
Kimi K2.6—Kimi K2.559.1%SourceNot directly comparable
JobBench
Kimi K2.6—Kimi K2.58.7%SourceNot directly comparable
Coding
SWE-bench Verified
Kimi K2.680.2%SourceKimi K2.576.8%SourceKimi K2.6 leads this result
LiveCodeBench v6
Kimi K2.689.6%SourceKimi K2.585.0%SourceKimi K2.6 leads this result
SWE-bench Pro
Kimi K2.658.6%SourceKimi K2.550.7%SourceKimi K2.6 leads this result
SWE Multilingual
Kimi K2.676.7%SourceKimi K2.573%SourceKimi K2.6 leads this result
SciCode
Kimi K2.652.2%SourceKimi K2.548.7%SourceKimi K2.6 leads this result
Terminal-Bench 2.0
Kimi K2.666.7%SourceKimi K2.5—Not directly comparable
Vibe Code Bench
Kimi K2.637.89%SourceKimi K2.5—Not directly comparable
cursorBench31
Kimi K2.647.6%SourceKimi K2.5—Not directly comparable
SWE-bench Verified*
Kimi K2.6—Kimi K2.570.8%SourceNot directly comparable
SWE-Rebench
Kimi K2.6—Kimi K2.558.5%SourceNot directly comparable
React Native Evals
Kimi K2.6—Kimi K2.577.2%SourceNot directly comparable
Reasoning
LongBench v2
Kimi K2.6—Kimi K2.561%SourceNot directly comparable
Knowledge
GPQA
Kimi K2.690.5%SourceKimi K2.587.6%SourceKimi K2.6 leads this result
GPQA-D
Kimi K2.690.5%SourceKimi K2.587.6%SourceKimi K2.6 leads this result
HLE
Kimi K2.634.7%SourceKimi K2.530.1%SourceKimi K2.6 leads this result
SuperGPQA
Kimi K2.6—Kimi K2.569.2%SourceNot directly comparable
MMLU-Pro
Kimi K2.6—Kimi K2.587.1%SourceNot directly comparable
MMLU-Pro (Arcee)
Kimi K2.6—Kimi K2.587.1%SourceNot directly comparable
Math
AIME26
Kimi K2.696.4%SourceKimi K2.595.8%SourceKimi K2.6 leads this result
HMMT Feb 2026
Kimi K2.692.7%SourceKimi K2.587.1%SourceKimi K2.6 leads this result
MMAnswerBench
Kimi K2.686.0%SourceKimi K2.581.8%SourceKimi K2.6 leads this result
FrontierMath v2 (Tiers 1-3)
Shared sourceKimi K2.638.966%Kimi K2.527.900%Kimi K2.6 leads this result
FrontierMath v2 (Tier 4)
Shared sourceKimi K2.614.580%Kimi K2.54.200%Kimi K2.6 leads this result
AIME 2025
Kimi K2.6—Kimi K2.596.1%SourceNot directly comparable
AIME25 (Arcee)
Kimi K2.6—Kimi K2.596.3%SourceNot directly comparable
HMMT Feb 2025
Kimi K2.6—Kimi K2.595.4%SourceNot directly comparable
HMMT Nov 2025
Kimi K2.6—Kimi K2.591.1%SourceNot directly comparable
Multilingual
Multimodal
MMMU-Pro
Kimi K2.679.4%SourceKimi K2.578.5%SourceKimi K2.6 leads this result
MMMU-Pro w/ Python
Kimi K2.680.1%SourceKimi K2.5—Not directly comparable
CharXiv
Kimi K2.680.4%SourceKimi K2.5—Not directly comparable
MathVision
Kimi K2.687.4%SourceKimi K2.5—Not directly comparable
V*
Kimi K2.696.9%SourceKimi K2.5—Not directly comparable
Video-MME
Kimi K2.6—Kimi K2.587.4%SourceNot directly comparable
MMVU
Kimi K2.6—Kimi K2.580.4%SourceNot directly comparable
VideoMMMU
Kimi K2.6—Kimi K2.586.6%SourceNot directly comparable
Instruction following
IFEval
Kimi K2.6—Kimi K2.593.9%SourceNot directly comparable
Frequently asked questions
Which is better, Kimi K2.6 or Kimi K2.5?
Kimi K2.6 has the higher public score estimate, 60.14 versus 59.01, but the 90% score intervals overlap. The higher estimate is not a decisive winner because the uncertainty ranges overlap.
Which is better for coding, Kimi K2.6 or Kimi K2.5?
The current coding averages use different weighted benchmark sets, so BenchLM does not name a winner from them. Read the shared benchmark rows directly and test the models on the same task set.
Which is better for agentic tasks, Kimi K2.6 or Kimi K2.5?
The current agentic tasks averages use different weighted benchmark sets, so BenchLM does not name a winner from them. Read the shared benchmark rows directly and test the models on the same task set.
Which costs less, Kimi K2.6 or Kimi K2.5?
For the stated presets, chat costs $0.00295 on Kimi K2.6 and $0.0021 on Kimi K2.5; repository review costs $0.0595 and $0.039; the cache-heavy agent loop costs $0.249 and $0.162. Kimi K2.6 has no published cached-input rate, so cached tokens use its listed input rate. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Which has the larger context window, Kimi K2.6 or Kimi K2.5?
Both models list the same context window, 256K.
Related comparisons
Last updated August 27, 2026
Watch Kimi K2.6 vs Kimi K2.5
One weekly email when material rank, price, or benchmark evidence changes make this matchup worth revisiting.
Read a sample issueJoin 2,000+ readers.