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
Kimi K3 has the higher public point estimate, 72.14 versus 56.76. Their conditional score ranges do not overlap. These ranges do not establish rank confidence. 3 results are shared. Category rows resting on Estimated evidence or different benchmark sets are marked directional and do not name a winner.
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.
Code generation, repair, and software-engineering tasks
Kimi K3
Kimi K3 has the higher public coding point estimate, 61.4 to 46.5, with Supported evidence for both models. The conditional ranges do not establish rank confidence.
Prompts that approach the documented context limit
Kimi K3
Kimi K3 has the larger documented context window.
1K fresh input + 500 output tokens
GPT-5.2-Codex
GPT-5.2-Codex has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.
200K cached + 20K fresh input + 10K output tokens
Kimi K3
Kimi K3 has the lower estimated token cost for this stated workload. GPT-5.2-Codex has no published cached-input rate, so cached tokens use its listed input rate.
50K fresh input + 3K output tokens
GPT-5.2-Codex
GPT-5.2-Codex has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.
Tool use, computer use, and multi-step task completion
Not enough matched evidence
GPT-5.2-Codex is scored on Estimated evidence for agentic, so the reading is directional rather than like-for-like.
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.
Like-for-like · 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.
Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.
2 categories rest on Estimated evidence or different evidence sets. Those rows remain visible for coverage context but do not name a winner.
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.
LiveCodeBench (Vals)Coding
Normalized gap 0.8Each 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.
| Category | GPT-5.2-Codex | Kimi K3 | Basis | Reading |
|---|---|---|---|---|
| Coding | 46.5Supported · #49/144 | 61.4Supported · #18/144 | Like-for-likeBenchAlign v5.8 lane · 3 vs 14 public rows | Kimi K3 leads · intervals overlap |
| Agentic | 40.3Estimated · #57/119 | 68.1Supported · #8/119 | Directional onlyBenchAlign v5.8 lane · 2 vs 12 public rows | Directional only |
| Knowledge | 55.5Estimated · #49/171 | 67.9Supported · #18/171 | Directional onlyBenchAlign v5.8 lane · 0 vs 6 public rows | Directional only |
| Reasoning | 78.9Unranked · 2 rankable rows | 65.8#18/27 | Not comparableProvisional lane · 0 vs 1 weighted rows | Not comparable |
| Multimodal | 73.5Unranked · 1 rankable row | 89.4#1/49 | Not comparableProvisional lane · 0 vs 3 weighted rows | Not comparable |
| Multilingual | Not ranked | Not ranked | Not comparableProvisional lane · 0 vs 0 weighted rows | Not comparable |
| Instruction following | 92.4#3/125 | Not ranked | Not comparableProvisional lane · 0 vs 0 weighted rows | Not comparable |
| Math | Not ranked | Not ranked | Not comparableProvisional lane · 0 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.
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.
1K fresh input + 500 output tokens
GPT-5.2-Codex has the lower modeled cost
Costs use the listed standard API rates.
50K fresh input + 3K output tokens
GPT-5.2-Codex has the lower modeled cost
Costs use the listed standard API rates.
200K cached + 20K fresh input + 10K output tokens
Kimi K3 has the lower modeled cost
GPT-5.2-Codex has no published cached-input rate, so cached tokens use its listed input rate.
Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.
Maximum documented context; output-token limits may be lower.
GPT-5.2-Codex
400K
Kimi K3
1.05M
GPT-5.2-Codex
Not sourced
Kimi K3
Not sourced
A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.
GPT-5.2-Codex
Not published
Kimi K3
$0.3 per 1M cached input tokens
GPT-5.2-Codex
Not sourced
Kimi K3
Not sourced
GPT-5.2-Codex
Not sourced
Kimi K3
Not sourced
GPT-5.2-Codex
Not sourced
Kimi K3
Not sourced
GPT-5.2-Codex
Reasoning
Kimi K3
Reasoning
GPT-5.2-Codex
Proprietary
Kimi K3
Pending
GPT-5.2-Codex
Proprietary
Kimi K3
Pending
GPT-5.2-Codex
2025-12-18
Kimi K3
2026-07-16
Kimi K3 has the higher public point estimate, 72.14 versus 56.76. Their conditional score ranges do not overlap. These ranges do not establish rank confidence. The page therefore keeps the decision tied to the specific documented workload.
Kimi K3 has the higher public coding point estimate, 61.4 to 46.5, with Supported evidence for both models. The conditional ranges do not establish rank confidence.
Kimi K3 scores higher for agentic tasks on the public lane, 68.1 to 40.3. GPT-5.2-Codex 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.
For the stated presets, chat costs $0.00875 on GPT-5.2-Codex and $0.0105 on Kimi K3; repository review costs $0.1295 and $0.195; the cache-heavy agent loop costs $0.525 and $0.27. GPT-5.2-Codex has no published cached-input rate, so cached tokens use its listed input rate.
Kimi K3 has the larger documented context window: 1.05M, compared with 400K.
The full public result ledger is available for audit without forcing a wide desktop table onto a phone.
Gert Labs
Not directly comparable
JobBench
Kimi K3 leads this result
Terminal-Bench 2.1
Not directly comparable
BrowseComp
Not directly comparable
DeepSearchQA
Not directly comparable
Toolathlon-Verified
Not directly comparable
MCP Atlas
Not directly comparable
AutomationBench
Not directly comparable
APEX-Agents
Not directly comparable
SpreadsheetBench 2
Not directly comparable
DECK-Bench
Not directly comparable
Terminal-Bench 2.1 (Vals)
Not directly comparable
ApprenticeBench
Not directly comparable
Vibe Code Bench
Not directly comparable
LiveCodeBench (Vals)
GPT-5.2-Codex leads this result
SWE-bench (Vals)
Kimi K3 leads this result
DeepSWE
Not directly comparable
CursorBench 3.2
Not directly comparable
FrontierSWE
Not directly comparable
ProgramBench
Not directly comparable
Kimi Code Bench v2
Not directly comparable
sweMarathon
Not directly comparable
PostTrain Bench
Not directly comparable
MLS-Bench Lite
Not directly comparable
VulcanBench v3
Not directly comparable
OpenHarmony Bench
Not directly comparable
FrontierSWE v2
Not directly comparable
PostTrainBench v1.1
Not directly comparable
OfficeQA Pro
Not directly comparable
MMMU-Pro
Not directly comparable
MMMU-Pro w/ Python
Not directly comparable
CharXiv w/o tools
Not directly comparable
CharXiv
Not directly comparable
MathVision
Not directly comparable
MathVision w/ Python
Not directly comparable
BabyVision w/ Python
Not directly comparable
ZeroBench
Not directly comparable
ZeroBench w/ Python
Not directly comparable
WorldVQA ForceAnswer
Not directly comparable
OmniDocBench
Not directly comparable
PerceptionBench
Not directly comparable
GPQA
Not directly comparable
GPQA-D
Not directly comparable
HLE
Not directly comparable
HLE w/o tools
Not directly comparable
GPQA Diamond (Vals)
Not directly comparable
MMLU-Pro (Vals)
Not directly comparable
Gray Swan IPI (15 attempts)
Not directly comparable
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
