Model comparison
GPT-5.4 vs Qwen3.5 397B
Head-to-head evidence from 30 shared benchmark results across 6 categories. Overall scores shown here use the public BenchAlign v5 ranking lane.
Public leaderboard positions: GPT-5.4 #8 (Supported); Qwen3.5 397B #71 (Estimated). Intervals and evidence labels describe ranking uncertainty, not a guarantee for a specific workload.
Evidence parity. GPT-5.4 and Qwen3.5 397B share 30 comparable benchmark results. 5 of 8 categories are comparable. 22 results are unique to GPT-5.4; 25 to Qwen3.5 397B.
Updated July 20, 2026- Shared results
- 30
- GPT-5.4 only
- 22
- Qwen3.5 397B only
- 25
- Comparable categories
- 5 / 8
Pick GPT-5.4 if you want the stronger benchmark profile. Qwen3.5 397B 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 30 shared benchmark results across 6 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 57.01. 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 56.5. The single biggest benchmark swing on the page is HLE, 52.1% to 28.7%. Qwen3.5 397B 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 / $3.60 output per 1M tokens for Qwen3.5 397B. That is roughly 4.2x on output cost alone. GPT-5.4 is the reasoning model in the pair, while Qwen3.5 397B 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 128K for Qwen3.5 397B.
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 | Qwen3.5 397B |
|---|---|---|---|
| Math | GPT-5.442.5 | Margin→ 48.1 | Qwen3.5 397B90.6 |
| Agentic | GPT-5.477.2 | Margin← 20.7 | Qwen3.5 397B56.5 |
| Coding | GPT-5.457.7 | Margin→ 8.8 | Qwen3.5 397B66.5 |
| Multimodal | GPT-5.473.2 | Margin→ 6.4 | Qwen3.5 397B79.6 |
| Knowledge | GPT-5.457.6 | Margin← 1.0 | Qwen3.5 397B56.6 |
| Reasoning | GPT-5.4Not measured | MarginNo overlap | Qwen3.5 397B63.2 |
| Multilingual | GPT-5.4Not measured | MarginNo overlap | Qwen3.5 397B84.7 |
| Inst. Following | GPT-5.4Not measured | MarginNo overlap | Qwen3.5 397B92.6 |
Decisive benchmark drivers
The largest measured benchmark gaps in this matchup, with exact reported values.
More
- Source ↗
HLE
KnowledgeA 52.1%B 28.7%Winner: GPT-5.4Δ 23.4HLE: GPT-5.4 scored 52.1%; Qwen3.5 397B scored 28.7%. GPT-5.4 wins this benchmark. - Source ↗
Terminal-Bench 2.0
AgenticA 75.1%B 52.5%Winner: GPT-5.4Δ 22.6Terminal-Bench 2.0: GPT-5.4 scored 75.1%; Qwen3.5 397B scored 52.5%. GPT-5.4 wins this benchmark. - Source ↗
BrowseComp
AgenticA 82.7%B 62%Winner: GPT-5.4Δ 20.7BrowseComp: GPT-5.4 scored 82.7%; Qwen3.5 397B scored 62%. GPT-5.4 wins this benchmark. - Source ↗
SWE-bench Pro
CodingA 57.7%B 50.9%Winner: GPT-5.4Δ 6.8SWE-bench Pro: GPT-5.4 scored 57.7%; Qwen3.5 397B scored 50.9%. GPT-5.4 wins this benchmark. - Source ↗
GPQA
KnowledgeA 92.8%B 88.4%Winner: GPT-5.4Δ 4.4GPQA: GPT-5.4 scored 92.8%; Qwen3.5 397B scored 88.4%. 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 | Qwen3.5 397B | Comparison |
|---|---|---|---|
| Input / output priceUSD per 1M tokens | GPT-5.4$2.5 input / $15 output | Qwen3.5 397B$0.6 input / $3.6 output | Qwen3.5 397B has the lower combined listed price. |
| Generation speedtokens per second | GPT-5.474 tok/s | Qwen3.5 397B96 tok/s | Qwen3.5 397B has the higher measured throughput. |
| First-answer latencyseconds to first token | GPT-5.4151.79 s | Qwen3.5 397B2.44 s | Qwen3.5 397B reaches the first token sooner. |
| Context windowmaximum listed tokens | GPT-5.41.05M | Qwen3.5 397B128K | GPT-5.4 lists the larger context window. |
Benchmark Deep Dive
AgenticGPT-5.4 wins23 benchmarks
| Benchmark | GPT-5.4 | Qwen3.5 397B | Result |
|---|---|---|---|
| Terminal-Bench 2.0Source | 75.1% | 52.5% | GPT-5.4 leads |
| CyberGymSource | 79.0% | — | Not comparable |
| BrowseCompSource | 82.7% | 62% | GPT-5.4 leads |
| OSWorld-VerifiedSource | 75% | — | Not comparable |
| MCP AtlasSource | 70.6% | 46.1% | GPT-5.4 leads |
| ToolathlonSource | 54.6% | 36.3% | GPT-5.4 leads |
| τ²-bench resultsSource | 87.1% | 95.6% | Qwen3.5 397B leads |
| Claw-EvalSource | 60.3% | 56.8% | GPT-5.4 leads |
| DeepSearchQASource | 73.6% | — | Not comparable |
| AA Agentic IndexSource | 41.1% | 19.9% | GPT-5.4 leads |
| APEX-Agents-AASource | 33.3% | 15.3% | GPT-5.4 leads |
| GDPval-AASource | 44.7% | 23.1% | GPT-5.4 leads |
| GDPval-AASource | 1395 | 962 | GPT-5.4 leads |
| Gert LabsSource | 64.89% | 46.76% | GPT-5.4 leads |
| ResearchClawBenchSource | 15.3% | 14.2% | GPT-5.4 leads |
| JobBenchSource | 38.9% | — | Not comparable |
| ExploitGymSource | 6.0% | — | Not comparable |
| QwenClawBenchSource | — | 51.8% | Not comparable |
| τ³-bench resultsSource | — | 68.4% | Not comparable |
| VITA-BenchSource | — | 43.7% | Not comparable |
| DeepPlanningSource | — | 37.6% | Not comparable |
| MCP-TasksSource | — | 74.2% | Not comparable |
| WideResearchSource | — | 74.0% | Not comparable |
CodingQwen3.5 397B wins8 benchmarks
| Benchmark | GPT-5.4 | Qwen3.5 397B | Result |
|---|---|---|---|
| LiveCodeBench ProSource | 87.5% | — | Not comparable |
| SWE-bench ProSource | 57.7% | 50.9% | GPT-5.4 leads |
| React Native EvalsSource | 85.3% | — | Not comparable |
| Vibe Code BenchSource | 67.42% | — | Not comparable |
| AA Coding IndexSource | 71.0% | 48.2% | GPT-5.4 leads |
| AA-SciCodeSource | 56.6% | 42.0% | GPT-5.4 leads |
| SWE-bench VerifiedSource | — | 76.2% | Not comparable |
| LiveCodeBench v6Source | — | 83.6% | Not comparable |
Reasoning4 benchmarks
KnowledgeGPT-5.4 wins17 benchmarks
| Benchmark | GPT-5.4 | Qwen3.5 397B | Result |
|---|---|---|---|
| GPQASource | 92.8% | 88.4% | GPT-5.4 leads |
| HLESource | 52.1% | 28.7% | GPT-5.4 leads |
| HLE w/o toolsSource | 39.8% | — | Not comparable |
| GPQA-DSource | 92.8% | — | Not comparable |
| HealthBench HardSource | 40.1% | — | Not comparable |
| MedXpertQA (Text)Source | 59.6% | — | Not comparable |
| Artificial Analysis Intelligence IndexSource | 51.4% | 33.7% | GPT-5.4 leads |
| AA-GPQA DiamondSource | 92.0% | 89.3% | GPT-5.4 leads |
| AA-HLESource | 41.6% | 27.3% | GPT-5.4 leads |
| AA-Omniscience IndexSource | 5.7% | -29.8% | GPT-5.4 leads |
| AA-Omniscience AccuracySource | 50.0% | 31.4% | GPT-5.4 leads |
| AA-Omniscience Hallucination RateSource | 88.6% | 89.1% | GPT-5.4 leads |
| HealthBench ProfessionalSource | 48.1% | — | Not comparable |
| SuperGPQASource | — | 70.4% | Not comparable |
| MMLU-ProSource | — | 87.8% | Not comparable |
| MMLU-ReduxSource | — | 94.9% | Not comparable |
| C-EvalSource | — | 93% | Not comparable |
MathQwen3.5 397B wins7 benchmarks
| Benchmark | GPT-5.4 | Qwen3.5 397B | Result |
|---|---|---|---|
| FrontierMath v2 (Tiers 1-3)Source | 47.600% | — | Not comparable |
| FrontierMath v2 (Tier 4)Source | 27.100% | — | Not comparable |
| AIME26Source | — | 93.3% | Not comparable |
| HMMT Feb 2025Source | — | 94.8% | Not comparable |
| HMMT Nov 2025Source | — | 92.7% | Not comparable |
| HMMT Feb 2026Source | — | 87.9% | Not comparable |
| MMAnswerBenchSource | — | 80.9% | Not comparable |
Multilingual2 benchmarks
MultimodalQwen3.5 397B wins14 benchmarks
| Benchmark | GPT-5.4 | Qwen3.5 397B | Result |
|---|---|---|---|
| MMMU-ProSource | 81.2% | 79% | GPT-5.4 leads |
| OfficeQA ProSource | 53.2% | — | Not comparable |
| MMMU-Pro w/ PythonSource | 82.1% | — | Not comparable |
| CharXivSource | 82.8% | 80.8% | GPT-5.4 leads |
| ERQASource | 65.4% | — | Not comparable |
| SimpleVQASource | 61.1% | — | Not comparable |
| ScreenSpot ProSource | 85.4% | 65.6% | GPT-5.4 leads |
| ZeroBenchSource | 41.0% | — | Not comparable |
| MedXpertQA (MM)Source | 77.1% | — | Not comparable |
| AA-MMMU-ProSource | 78.4% | 77.3% | GPT-5.4 leads |
| Design Arena WebsiteSource | 1252 | — | Not comparable |
| MathVisionSource | — | 88.6% | Not comparable |
| VideoMMMUSource | — | 84.7% | Not comparable |
| V*Source | — | 95.8% | Not comparable |
Frequently Asked Questions (6)
Which is better, GPT-5.4 or Qwen3.5 397B?
GPT-5.4 is ahead on BenchLM's BenchAlign leaderboard, 74.24 to 57.01. The biggest single separator in this matchup is HLE, where the scores are 52.1% and 28.7%.
Which is better for knowledge tasks, GPT-5.4 or Qwen3.5 397B?
GPT-5.4 has the edge for knowledge tasks in this comparison, averaging 57.6 versus 56.6. Inside this category, AA-Omniscience Index is the benchmark that creates the most daylight between them.
Which is better for coding, GPT-5.4 or Qwen3.5 397B?
Qwen3.5 397B has the edge for coding in this comparison, averaging 66.5 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 Qwen3.5 397B?
Qwen3.5 397B has the edge for math in this comparison, averaging 90.6 versus 42.5. GPT-5.4 stays close enough that the answer can still flip depending on your workload.
Which is better for agentic tasks, GPT-5.4 or Qwen3.5 397B?
GPT-5.4 has the edge for agentic tasks in this comparison, averaging 77.2 versus 56.5. 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 Qwen3.5 397B?
Qwen3.5 397B has the edge for multimodal and grounded tasks in this comparison, averaging 79.6 versus 73.2. Inside this category, ScreenSpot Pro is the benchmark that creates the most daylight between them.
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.