Long documents
Prompts that approach the documented context limit
Kimi K3
Kimi K3 has the larger documented context window.
Confidence: documented
Kimi K2.5 vs Kimi K3
Updated October 2, 2026. Rank says Kimi K3 is ahead. Price, access, and your workload can each overturn that. Public scores include evidence status and uncertainty.
Both of these models will change. Get the price, version and retirement notices for the pair, each with its source. Follow model changes
Share or export
Decision reading
Kimi K3 has the higher public score, 72.14 versus 53.67, and the 90% score intervals do not overlap. 8 results are shared. Category rows resting on Estimated evidence or different benchmark sets are marked directional and do not name a winner.
- Shared results
- 8
- Kimi K2.5 only
- 37
- Kimi K3 only
- 40
- Like-for-like categories
- 0 / 8
Supported: Kimi K2.5 and Kimi K3How the comparison works
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-ratesCache-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 K2.5 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-ratesCoding work
Code generation, repair, and software-engineering tasks
Not enough matched evidence
Kimi K2.5 is scored on Estimated evidence for coding, so the reading is directional rather than like-for-like.
Confidence: limitedAgentic work
Tool use, computer use, and multi-step task completion
Not enough matched evidence
Kimi K2.5 is scored on Estimated evidence for agentic, so the reading is directional rather than like-for-like.
Confidence: limited
Which one for a specific job
Choose a job from the LLM Selector's task catalog to see the category row it rests on, under the same basis rules as the table below. A directional row stays directional; choosing a job never creates a winner.
The same task catalog as the LLM Selector. Each job names the evidence surface it rests on; nothing here adds a new score.
40.1Kimi K2.561.4Kimi K3
Directional only · BenchAlign v5.8
Kimi K3 has the higher coding point estimate. Conditional score ranges do not establish rank confidence.
Coding scores combine specific tasks and setups. Match the editor, harness, and effort to your workflow.
Same basis rules as the category table below
What is actually comparable
Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.
37 Kimi K2.5 only
8 Shared
40 Kimi K3 only
4 categories rest on Estimated evidence or different evidence sets. Those rows remain visible for coverage context but do not name a winner.
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.
BrowseCompAgentic
Normalized gap 30.6Kimi K2.5:60.6%Kimi K3:91.2%HLEKnowledge
Normalized gap 25.9Kimi K2.5:30.1%Kimi K3:56%GPQAKnowledge
Normalized gap 5.9Kimi K2.5:87.6%Kimi K3:93.5%MMMU-ProMultimodal
Normalized gap 3.1Kimi K2.5:78.5%Kimi K3:81.6%
Bars run 0–100 on each benchmark’s normalized display scale
Category results, on a stated basis
Each row shows the public-lane category score for both models: the BenchAlign v5.8 lane for agentic, coding, and knowledge, and the weighted public rows elsewhere, with the same rank each model page reports. A row is like-for-like only when both scores rest on Supported evidence or the same weighted set. Directional and not-comparable rows remain visible, but they never receive a winner.
Agentic
Directional only- Kimi K2.5
- 38.0
- Estimated · #63/119
- Kimi K3
- 68.1
- Supported · #8/119
- Basis
- BenchAlign v5.8 lane · 14 vs 12 public rows
- Reading
- Directional only
Coding
Directional only- Kimi K2.5
- 40.1
- Estimated · #64/144
- Kimi K3
- 61.4
- Supported · #18/144
- Basis
- BenchAlign v5.8 lane · 8 vs 14 public rows
- Reading
- Directional only
Multimodal
Directional only- Kimi K2.5
- 66.8
- #24/49
- Kimi K3
- 89.4
- #1/49
- Basis
- Provisional lane · 1 vs 3 weighted rows
- Reading
- Directional only
Knowledge
Directional only- Kimi K2.5
- 49.5
- Estimated · #66/171
- Kimi K3
- 67.9
- Supported · #18/171
- Basis
- BenchAlign v5.8 lane · 6 vs 6 public rows
- Reading
- Directional only
Reasoning
Not comparable- Kimi K2.5
- 55.2
- Unranked · 3 rankable rows
- Kimi K3
- 65.8
- #18/27
- Basis
- Provisional lane · 1 vs 1 weighted rows
- Reading
- Not comparable
Multilingual
Not comparable- Kimi K2.5
- 87.3
- #8/16
- Kimi K3
- Not ranked
- Basis
- Provisional lane · 1 vs 0 weighted rows
- Reading
- Not comparable
Instruction following
Not comparable- Kimi K2.5
- 84.4
- #41/125
- Kimi K3
- Not ranked
- Basis
- Provisional lane · 0 vs 0 weighted rows
- Reading
- Not comparable
Math
Not comparable- Kimi K2.5
- 62.1
- #4/7
- Kimi K3
- Not ranked
- Basis
- Provisional lane · 4 vs 0 weighted rows
- Reading
- Not comparable
| Category | Kimi K2.5 | Kimi K3 | Basis | Reading |
|---|---|---|---|---|
| Agentic | 38.0Estimated · #63/119 | 68.1Supported · #8/119 | Directional onlyBenchAlign v5.8 lane · 14 vs 12 public rows | Directional only |
| Coding | 40.1Estimated · #64/144 | 61.4Supported · #18/144 | Directional onlyBenchAlign v5.8 lane · 8 vs 14 public rows | Directional only |
| Multimodal | 66.8#24/49 | 89.4#1/49 | Directional onlyProvisional lane · 1 vs 3 weighted rows | Directional only |
| Knowledge | 49.5Estimated · #66/171 | 67.9Supported · #18/171 | Directional onlyBenchAlign v5.8 lane · 6 vs 6 public rows | Directional only |
| Reasoning | 55.2Unranked · 3 rankable rows | 65.8#18/27 | Not comparableProvisional lane · 1 vs 1 weighted rows | Not comparable |
| Multilingual | 87.3#8/16 | Not ranked | Not comparableProvisional lane · 1 vs 0 weighted rows | Not comparable |
| Instruction following | 84.4#41/125 | Not ranked | Not comparableProvisional lane · 0 vs 0 weighted rows | Not comparable |
| Math | 62.1#4/7 | Not ranked | Not comparableProvisional lane · 4 vs 0 weighted rows | Not comparable |
Ranks count the models scored in each category’s lane, so the agentic, coding, and knowledge denominators (BenchAlign v5.8) differ from the provisional-lane categories. Unranked scores sit on the provisional lane’s scale but never name a winner. Like-for-like rows within 0.5 points read as a practical tie.
Supported evidence per lane · bars run 0–100Methodology
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 K2.5
- $0.0021
- Fits in one request
- Kimi K3
- $0.0105
- 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 K2.5
- $0.039
- Fits in one request
- Kimi K3
- $0.195
- 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 K2.5
- $0.162
- Fits in one request
- Cached input priced at the published list-input rate
- Kimi K3
- $0.27
- Fits in one request
Kimi K2.5 has the lower modeled cost
Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Cached input falls back to the list input rate only where a cached rate is unpublished
Specification differences
Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.
SpecificationKimi K2.5Kimi K3
Context window
Maximum documented context; output-token limits may be lower.
Kimi K2.5
256K
Kimi K3
1.05M
API model ID
Kimi K2.5
Not sourced
Kimi K3
Not sourced
Cached-input rate
A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.
Kimi K2.5
Not published
Kimi K3
$0.3 per 1M cached input tokens
Documented inputs
Kimi K2.5
Not sourced
Kimi K3
Not sourced
Documented outputs
Kimi K2.5
Not sourced
Kimi K3
Not sourced
Provider availability
Kimi K2.5
Not sourced
Kimi K3
Not sourced
Reasoning profile
Kimi K2.5
Non-Reasoning
Kimi K3
Reasoning
Weight access
Kimi K2.5
Open Weight
Kimi K3
Pending
License
Kimi K2.5
Open Weight
Kimi K3
Pending
Release date
Kimi K2.5
2026-02-01
Kimi K3
2026-07-16
If you already use one of these models
- Deployment change
- Both entries list Moonshot AI as the provider. Confirm endpoint, model ID, limits, and feature support before switching.
- Quality signal
- Kimi K3 has the higher public score, 72.14 versus 53.67, and the 90% score intervals do not overlap.
- Workload cost
- Repository review: $0.039 vs $0.195. Cache-heavy agent loop: $0.162 vs $0.27.
- Context tradeoff
- Kimi K3 has the larger documented window (1.05M).
Run the same representative tasks against both endpoints before changing production traffic.
Questions
Which is better, Kimi K2.5 or Kimi K3?
Kimi K3 has the higher public score, 72.14 versus 53.67, and the 90% score intervals do not overlap. That is the clearest overall quality signal in the current public data.
Which is better for coding, Kimi K2.5 or Kimi K3?
Kimi K3 scores higher for coding on the public lane, 61.4 to 40.1. Kimi K2.5 is scored on Estimated evidence for coding, so the reading is directional rather than like-for-like. BenchLM does not name a winner for coding from a directional reading; read the shared benchmark rows directly and test the models on the same task set.
Which is better for agentic tasks, Kimi K2.5 or Kimi K3?
Kimi K3 scores higher for agentic tasks on the public lane, 68.1 to 38. Kimi K2.5 is scored on Estimated evidence for agentic tasks, so the reading is directional rather than like-for-like. BenchLM does not name a winner for agentic tasks from a directional reading; read the shared benchmark rows directly and test the models on the same task set.
Which costs less, Kimi K2.5 or Kimi K3?
For the stated presets, chat costs $0.0021 on Kimi K2.5 and $0.0105 on Kimi K3; repository review costs $0.039 and $0.195; the cache-heavy agent loop costs $0.162 and $0.27. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Which has the larger context window, Kimi K2.5 or Kimi K3?
Kimi K3 has the larger documented context window: 1.05M, compared with 256K.
Self-host vs API cost
Estimates at 50,000 req/day · 1000 tokens/req average.
Kimi K2.5
API / mo$2,700
Self-host / mo$5,221
Break-even132M/day
Kimi K3
API / mo$0
Self-host / moNot listed
Break-even—
Proprietary model — self-hosting not applicable.
Benchmark evidence
The full public result ledger is available for audit without forcing a wide desktop table onto a phone.
Browse raw public benchmark evidence85 rows
Agentic
Terminal-Bench 2.0
Kimi K2.550.8%SourceKimi K3—Not directly comparable
BrowseComp
Kimi K2.560.6%SourceKimi K391.2%SourceKimi K3 leads this result
Claw-Eval
Kimi K2.552.3%SourceKimi K3—Not directly comparable
QwenClawBench
Kimi K2.554.3%SourceKimi K3—Not directly comparable
τ³-bench results
Kimi K2.565.7%SourceKimi K3—Not directly comparable
DeepSearchQA
Kimi K2.577.1%SourceKimi K395.0%SourceKimi K3 leads this result
DeepPlanning
Kimi K2.514.4%SourceKimi K3—Not directly comparable
Toolathlon
Kimi K2.527.8%SourceKimi K3—Not directly comparable
MCP Atlas
Kimi K2.529.5%SourceKimi K384.2%SourceKimi K3 leads this result
MCP-Tasks
Kimi K2.559.1%SourceKimi K3—Not directly comparable
WideResearch
Kimi K2.572.7%SourceKimi K3—Not directly comparable
Gert Labs
Kimi K2.545.88%SourceKimi K3—Not directly comparable
ResearchClawBench
Kimi K2.514.0%SourceKimi K3—Not directly comparable
JobBench
Kimi K2.58.7%SourceKimi K352.9%SourceKimi K3 leads this result
Terminal-Bench 2.1
Kimi K2.5—Kimi K388.3%SourceNot directly comparable
Toolathlon-Verified
Kimi K2.5—Kimi K373.2%SourceNot directly comparable
AutomationBench
Kimi K2.5—Kimi K330.8%SourceNot directly comparable
APEX-Agents
Kimi K2.5—Kimi K337.6%SourceNot directly comparable
SpreadsheetBench 2
Kimi K2.5—Kimi K334.8%SourceNot directly comparable
DECK-Bench
Kimi K2.5—Kimi K373.5%SourceNot directly comparable
Terminal-Bench 2.1 (Vals)
Kimi K2.5—Kimi K380.9%SourceNot directly comparable
ApprenticeBench
Kimi K2.5—Kimi K318%SourceNot directly comparable
Coding
SWE-bench Verified
Kimi K2.576.8%SourceKimi K3—Not directly comparable
SWE-bench Verified*
Kimi K2.570.8%SourceKimi K3—Not directly comparable
LiveCodeBench v6
Kimi K2.585.0%SourceKimi K3—Not directly comparable
SWE-bench Pro
Kimi K2.550.7%SourceKimi K3—Not directly comparable
SWE Multilingual
Kimi K2.573%SourceKimi K3—Not directly comparable
SWE-Rebench
Kimi K2.558.5%SourceKimi K3—Not directly comparable
React Native Evals
Kimi K2.577.2%SourceKimi K3—Not directly comparable
SciCode
Kimi K2.548.7%SourceKimi K3—Not directly comparable
DeepSWE
Kimi K2.5—Kimi K367.5%SourceNot directly comparable
CursorBench 3.2
Kimi K2.5—Kimi K360.8%SourceNot directly comparable
FrontierSWE
Kimi K2.5—Kimi K381.2%SourceNot directly comparable
ProgramBench
Kimi K2.5—Kimi K377.8%SourceNot directly comparable
Kimi Code Bench v2
Kimi K2.5—Kimi K372.9%SourceNot directly comparable
sweMarathon
Kimi K2.5—Kimi K342%SourceNot directly comparable
PostTrain Bench
Kimi K2.5—Kimi K336.6%SourceNot directly comparable
MLS-Bench Lite
Kimi K2.5—Kimi K348.3%SourceNot directly comparable
VulcanBench v3
Kimi K2.5—Kimi K373.7%SourceNot directly comparable
OpenHarmony Bench
Kimi K2.5—Kimi K357.3%SourceNot directly comparable
FrontierSWE v2
Kimi K2.5—Kimi K325.9%SourceNot directly comparable
LiveCodeBench (Vals)
Kimi K2.5—Kimi K387.2%SourceNot directly comparable
SWE-bench (Vals)
Kimi K2.5—Kimi K393.4%SourceNot directly comparable
PostTrainBench v1.1
Kimi K2.5—Kimi K332.0%SourceNot directly comparable
Reasoning
Multimodal
MMMU-Pro
Kimi K2.578.5%SourceKimi K381.6%SourceKimi K3 leads this result
Video-MME
Kimi K2.587.4%SourceKimi K3—Not directly comparable
MMVU
Kimi K2.580.4%SourceKimi K3—Not directly comparable
VideoMMMU
Kimi K2.586.6%SourceKimi K3—Not directly comparable
OfficeQA Pro
Kimi K2.5—Kimi K363.3%SourceNot directly comparable
MMMU-Pro w/ Python
Kimi K2.5—Kimi K383.4%SourceNot directly comparable
CharXiv w/o tools
Kimi K2.5—Kimi K384.8%SourceNot directly comparable
CharXiv
Kimi K2.5—Kimi K391.3%SourceNot directly comparable
MathVision
Kimi K2.5—Kimi K394.3%SourceNot directly comparable
MathVision w/ Python
Kimi K2.5—Kimi K397.8%SourceNot directly comparable
BabyVision w/ Python
Kimi K2.5—Kimi K385.7%SourceNot directly comparable
ZeroBench
Kimi K2.5—Kimi K323.0%SourceNot directly comparable
ZeroBench w/ Python
Kimi K2.5—Kimi K341.0%SourceNot directly comparable
WorldVQA ForceAnswer
Kimi K2.5—Kimi K351.0%SourceNot directly comparable
OmniDocBench
Kimi K2.5—Kimi K391.1%SourceNot directly comparable
PerceptionBench
Kimi K2.5—Kimi K358.5%SourceNot directly comparable
Knowledge
GPQA
Kimi K2.587.6%SourceKimi K393.5%SourceKimi K3 leads this result
GPQA-D
Kimi K2.587.6%SourceKimi K393.5%SourceKimi K3 leads this result
SuperGPQA
Kimi K2.569.2%SourceKimi K3—Not directly comparable
MMLU-Pro
Kimi K2.587.1%SourceKimi K3—Not directly comparable
MMLU-Pro (Arcee)
Kimi K2.587.1%SourceKimi K3—Not directly comparable
HLE
Kimi K2.530.1%SourceKimi K356%SourceKimi K3 leads this result
HLE w/o tools
Kimi K2.5—Kimi K343.5%SourceNot directly comparable
GPQA Diamond (Vals)
Kimi K2.5—Kimi K392.9%SourceNot directly comparable
MMLU-Pro (Vals)
Kimi K2.5—Kimi K388.0%SourceNot directly comparable
Multilingual
Instruction following
Math
AIME 2025
Kimi K2.596.1%SourceKimi K3—Not directly comparable
AIME26
Kimi K2.595.8%SourceKimi K3—Not directly comparable
AIME25 (Arcee)
Kimi K2.596.3%SourceKimi K3—Not directly comparable
HMMT Feb 2025
Kimi K2.595.4%SourceKimi K3—Not directly comparable
HMMT Nov 2025
Kimi K2.591.1%SourceKimi K3—Not directly comparable
HMMT Feb 2026
Kimi K2.587.1%SourceKimi K3—Not directly comparable
MMAnswerBench
Kimi K2.581.8%SourceKimi K3—Not directly comparable
FrontierMath v2 (Tiers 1-3)
Kimi K2.527.900%SourceKimi K3—Not directly comparable
FrontierMath v2 (Tier 4)
Kimi K2.54.200%SourceKimi K3—Not directly comparable
85 public results · 8 shared
Watch Kimi K2.5 vs Kimi K3
One weekly email when material rank, price, or benchmark evidence changes make this matchup worth revisiting.
Read a sample issueJoin 2,000+ readers.
Last updated October 2, 2026