Model comparison
GPT-5.4 vs Kimi K2.5
Head-to-head evidence from 35 shared benchmark results across 7 categories. Overall scores shown here use the public BenchAlign v5 ranking lane.
Public leaderboard positions: GPT-5.4 #8 (Supported); Kimi K2.5 #54 (Supported). Intervals and evidence labels describe ranking uncertainty, not a guarantee for a specific workload.
Evidence parity. GPT-5.4 and Kimi K2.5 share 35 comparable benchmark results. 5 of 8 categories are comparable. 17 results are unique to GPT-5.4; 28 to Kimi K2.5.
Updated July 20, 2026- Shared results
- 35
- GPT-5.4 only
- 17
- Kimi K2.5 only
- 28
- Comparable categories
- 5 / 8
Pick GPT-5.4 if you want the stronger benchmark profile. Kimi K2.5 only becomes the better choice if mathematics is the priority or you want the cheaper token bill.
Confidence note. This is a partial-evidence comparison with 35 shared benchmark results across 7 evidence categories; 5 of 8 categories currently have scoreable aggregates for both models. Treat the verdict as directional until coverage is more balanced.
Why this result
GPT-5.4 is clearly ahead on the BenchAlign aggregate, 74.24 to 59.66. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
GPT-5.4's sharpest advantage is in agentic, where it averages 77.2 against 55. The single biggest benchmark swing on the page is Terminal-Bench 2.0, 75.1% to 50.8%. Kimi K2.5 does hit back in mathematics, so the answer changes if that is the part of the workload you care about most.
GPT-5.4 is also the more expensive model on tokens at $2.50 input / $15.00 output per 1M tokens, versus $0.60 input / $3.00 output per 1M tokens for Kimi K2.5. That is roughly 5.0x on output cost alone. GPT-5.4 is the reasoning model in the pair, while Kimi K2.5 is not. That usually helps on harder chain-of-thought-heavy tests, but it can also mean more latency and more token spend in real use. GPT-5.4 gives you the larger context window at 1.05M, compared with 256K for Kimi K2.5.
Category breakdown
Exact category averages are shown below. Not measured means BenchLM does not have enough sourced public coverage for that model and category.
| Category | GPT-5.4 | Δ | Kimi K2.5 |
|---|---|---|---|
| Agentic | GPT-5.477.2 | Margin← 22.2 | Kimi K2.555.0 |
| Math | GPT-5.442.5 | Margin→ 18.1 | Kimi K2.560.6 |
| Multimodal | GPT-5.473.2 | Margin→ 5.3 | Kimi K2.578.5 |
| Coding | GPT-5.457.7 | Margin→ 1.7 | Kimi K2.559.4 |
| Knowledge | GPT-5.457.6 | Margin← 0.7 | Kimi K2.556.9 |
| Reasoning | GPT-5.4Not measured | MarginNo overlap | Kimi K2.561.0 |
| Multilingual | GPT-5.4Not measured | MarginNo overlap | Kimi K2.582.3 |
| Inst. Following | GPT-5.4Not measured | MarginNo overlap | Kimi K2.593.9 |
Decisive benchmark drivers
The largest measured benchmark gaps in this matchup, with exact reported values.
More
- Source ↗
Terminal-Bench 2.0
AgenticA 75.1%B 50.8%Winner: GPT-5.4Δ 24.3Terminal-Bench 2.0: GPT-5.4 scored 75.1%; Kimi K2.5 scored 50.8%. GPT-5.4 wins this benchmark. - Source ↗
FrontierMath v2 (Tier 4)
MathA 27.100%B 4.200%Winner: GPT-5.4Δ 22.9FrontierMath v2 (Tier 4): GPT-5.4 scored 27.100%; Kimi K2.5 scored 4.200%. GPT-5.4 wins this benchmark. - Source ↗
BrowseComp
AgenticA 82.7%B 60.6%Winner: GPT-5.4Δ 22.1BrowseComp: GPT-5.4 scored 82.7%; Kimi K2.5 scored 60.6%. GPT-5.4 wins this benchmark. - Source ↗
HLE
KnowledgeA 52.1%B 30.1%Winner: GPT-5.4Δ 22HLE: GPT-5.4 scored 52.1%; Kimi K2.5 scored 30.1%. GPT-5.4 wins this benchmark. - Source ↗
FrontierMath v2 (Tiers 1-3)
MathA 47.600%B 27.900%Winner: GPT-5.4Δ 19.7FrontierMath v2 (Tiers 1-3): GPT-5.4 scored 47.600%; Kimi K2.5 scored 27.900%. GPT-5.4 wins this benchmark.
Operational comparison
Runtime and commercial metrics are compared only when both models have a complete sourced value.
| Metric | GPT-5.4 | Kimi K2.5 | Comparison |
|---|---|---|---|
| Input / output priceUSD per 1M tokens | GPT-5.4$2.5 input / $15 output | Kimi K2.5$0.6 input / $3 output | Kimi K2.5 has the lower combined listed price. |
| Generation speedtokens per second | GPT-5.474 tok/s | Kimi K2.545 tok/s | GPT-5.4 has the higher measured throughput. |
| First-answer latencyseconds to first token | GPT-5.4151.79 s | Kimi K2.52.38 s | Kimi K2.5 reaches the first token sooner. |
| Context windowmaximum listed tokens | GPT-5.41.05M | Kimi K2.5256K | GPT-5.4 lists the larger context window. |
Benchmark Deep Dive
AgenticGPT-5.4 wins22 benchmarks
| Benchmark | GPT-5.4 | Kimi K2.5 | Result |
|---|---|---|---|
| Terminal-Bench 2.0Source | 75.1% | 50.8% | GPT-5.4 leads |
| CyberGymSource | 79.0% | — | Not comparable |
| BrowseCompSource | 82.7% | 60.6% | GPT-5.4 leads |
| OSWorld-VerifiedSource | 75% | — | Not comparable |
| MCP AtlasSource | 70.6% | 29.5% | GPT-5.4 leads |
| ToolathlonSource | 54.6% | 27.8% | GPT-5.4 leads |
| τ²-bench resultsSource | 87.1% | 95.9% | Kimi K2.5 leads |
| Claw-EvalSource | 60.3% | 52.3% | GPT-5.4 leads |
| DeepSearchQASource | 73.6% | 77.1% | Kimi K2.5 leads |
| AA Agentic IndexSource | 41.1% | 21.7% | GPT-5.4 leads |
| APEX-Agents-AASource | 33.3% | 11.5% | GPT-5.4 leads |
| GDPval-AASource | 44.7% | 25.4% | GPT-5.4 leads |
| GDPval-AASource | 1395 | 1009 | GPT-5.4 leads |
| Gert LabsSource | 64.89% | 45.88% | GPT-5.4 leads |
| ResearchClawBenchSource | 15.3% | 14.0% | GPT-5.4 leads |
| JobBenchSource | 38.9% | 8.7% | GPT-5.4 leads |
| ExploitGymSource | 6.0% | — | Not comparable |
| QwenClawBenchSource | — | 54.3% | Not comparable |
| τ³-bench resultsSource | — | 65.7% | Not comparable |
| DeepPlanningSource | — | 14.4% | Not comparable |
| MCP-TasksSource | — | 59.1% | Not comparable |
| WideResearchSource | — | 72.7% | Not comparable |
CodingKimi K2.5 wins12 benchmarks
| Benchmark | GPT-5.4 | Kimi K2.5 | Result |
|---|---|---|---|
| LiveCodeBench ProSource | 87.5% | — | Not comparable |
| SWE-bench ProSource | 57.7% | 50.7% | GPT-5.4 leads |
| React Native EvalsSource | 85.3% | 77.2% | GPT-5.4 leads |
| Vibe Code BenchSource | 67.42% | — | Not comparable |
| AA Coding IndexSource | 71.0% | 46.8% | GPT-5.4 leads |
| AA-SciCodeSource | 56.6% | 49.0% | GPT-5.4 leads |
| SWE-bench VerifiedSource | — | 76.8% | Not comparable |
| SWE-bench Verified*Source | — | 70.8% | Not comparable |
| LiveCodeBench v6Source | — | 85.0% | Not comparable |
| SWE MultilingualSource | — | 73% | Not comparable |
| SWE-RebenchSource | — | 58.5% | Not comparable |
| SciCodeSource | — | 48.7% | Not comparable |
Reasoning3 benchmarks
KnowledgeGPT-5.4 wins16 benchmarks
| Benchmark | GPT-5.4 | Kimi K2.5 | Result |
|---|---|---|---|
| GPQASource | 92.8% | 87.6% | GPT-5.4 leads |
| HLESource | 52.1% | 30.1% | GPT-5.4 leads |
| HLE w/o toolsSource | 39.8% | — | Not comparable |
| GPQA-DSource | 92.8% | 87.6% | GPT-5.4 leads |
| HealthBench HardSource | 40.1% | — | Not comparable |
| MedXpertQA (Text)Source | 59.6% | — | Not comparable |
| Artificial Analysis Intelligence IndexSource | 51.4% | 35.4% | GPT-5.4 leads |
| AA-GPQA DiamondSource | 92.0% | 87.9% | GPT-5.4 leads |
| AA-HLESource | 41.6% | 29.4% | GPT-5.4 leads |
| AA-Omniscience IndexSource | 5.7% | -8.1% | GPT-5.4 leads |
| AA-Omniscience AccuracySource | 50.0% | 34.3% | GPT-5.4 leads |
| AA-Omniscience Hallucination RateSource | 88.6% | 64.6% | Kimi K2.5 leads |
| HealthBench ProfessionalSource | 48.1% | — | Not comparable |
| SuperGPQASource | — | 69.2% | Not comparable |
| MMLU-ProSource | — | 87.1% | Not comparable |
| MMLU-Pro (Arcee)Source | — | 87.1% | Not comparable |
MathKimi K2.5 wins9 benchmarks
| Benchmark | GPT-5.4 | Kimi K2.5 | Result |
|---|---|---|---|
| FrontierMath v2 (Tiers 1-3)Source | 47.600% | 27.900% | GPT-5.4 leads |
| FrontierMath v2 (Tier 4)Source | 27.100% | 4.200% | GPT-5.4 leads |
| AIME 2025Source | — | 96.1% | Not comparable |
| AIME26Source | — | 95.8% | Not comparable |
| AIME25 (Arcee)Source | — | 96.3% | Not comparable |
| HMMT Feb 2025Source | — | 95.4% | Not comparable |
| HMMT Nov 2025Source | — | 91.1% | Not comparable |
| HMMT Feb 2026Source | — | 87.1% | Not comparable |
| MMAnswerBenchSource | — | 81.8% | Not comparable |
Multilingual2 benchmarks
MultimodalKimi K2.5 wins14 benchmarks
| Benchmark | GPT-5.4 | Kimi K2.5 | Result |
|---|---|---|---|
| MMMU-ProSource | 81.2% | 78.5% | GPT-5.4 leads |
| OfficeQA ProSource | 53.2% | — | Not comparable |
| MMMU-Pro w/ PythonSource | 82.1% | — | Not comparable |
| CharXivSource | 82.8% | — | Not comparable |
| ERQASource | 65.4% | — | Not comparable |
| SimpleVQASource | 61.1% | — | Not comparable |
| ScreenSpot ProSource | 85.4% | — | Not comparable |
| ZeroBenchSource | 41.0% | — | Not comparable |
| MedXpertQA (MM)Source | 77.1% | — | Not comparable |
| AA-MMMU-ProSource | 78.4% | 75.4% | GPT-5.4 leads |
| Design Arena WebsiteSource | 1252 | 1282 | Kimi K2.5 leads |
| Video-MMESource | — | 87.4% | Not comparable |
| MMVUSource | — | 80.4% | Not comparable |
| VideoMMMUSource | — | 86.6% | Not comparable |
Frequently Asked Questions (6)
Which is better, GPT-5.4 or Kimi K2.5?
GPT-5.4 is ahead on BenchLM's BenchAlign leaderboard, 74.24 to 59.66. The biggest single separator in this matchup is Terminal-Bench 2.0, where the scores are 75.1% and 50.8%.
Which is better for knowledge tasks, GPT-5.4 or Kimi K2.5?
GPT-5.4 has the edge for knowledge tasks in this comparison, averaging 57.6 versus 56.9. Inside this category, AA-Omniscience Hallucination Rate is the benchmark that creates the most daylight between them.
Which is better for coding, GPT-5.4 or Kimi K2.5?
Kimi K2.5 has the edge for coding in this comparison, averaging 59.4 versus 57.7. Inside this category, AA Coding Index is the benchmark that creates the most daylight between them.
Which is better for math, GPT-5.4 or Kimi K2.5?
Kimi K2.5 has the edge for math in this comparison, averaging 60.6 versus 42.5. Inside this category, FrontierMath v2 (Tier 4) is the benchmark that creates the most daylight between them.
Which is better for agentic tasks, GPT-5.4 or Kimi K2.5?
GPT-5.4 has the edge for agentic tasks in this comparison, averaging 77.2 versus 55. Inside this category, GDPval-AA is the benchmark that creates the most daylight between them.
Which is better for multimodal and grounded tasks, GPT-5.4 or Kimi K2.5?
Kimi K2.5 has the edge for multimodal and grounded tasks in this comparison, averaging 78.5 versus 73.2. Inside this category, Design Arena Website is the benchmark that creates the most daylight between them.
Self-host vs API cost
Estimates at 50,000 req/day · 1000 tokens/req average.
Related Comparisons
Explore More
Choose a model with this week’s evidence
Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.
One email each week. Unsubscribe anytime.