Zhipu AIGLM-4.6
Best configuration ranks #254 of 610 on the BenchLeader Index at 52.3 ±9.7 (thinking reasoning effort). Released 30 Sept 2025.
- Blended price
- $0.963/M
- $0.550 in · $2.20 out
- Output speed
- 44 tok/s
- First answer
- 48 s
- first token 2.74 s
- Context
- 200k
- Overall index52
- Agents & tools72
- Knowledge44
- Instruction following45
- Long context53
- Composite52
Reasoning-effort configurations
The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Default” means the publisher did not say which setting was used.
| Effort | Index | Rank | Speed | First answer | Chat reply cost | Categories |
|---|---|---|---|---|---|---|
| thinkingbest | 52.3 | #254 | 44 tok/s | 48 s | $0.0009 | Agents & tools 72 · Composite 52 · Instruction following 45 · Knowledge 44 · Long context 53 |
| default | 47.6 | #349 | 65 tok/s | 2.74 s | $0.0009 | Agents & tools 40 · Coding 44 · Composite 48 · Human preference 61 · Instruction following 40 · Knowledge 51 · Long context 38 · Maths 51 · Reasoning 53 |
Benchmark results
One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.
Reasoning
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| LMArena Hard Prompts | – | 1442#104 | LMArena | |
| GPQA Diamond (AA)not in index | 78.0%#209 | 63.2%#339 | Artificial Analysis | |
| Humanity's Last Exam (AA)not in index | 14.5%#198 | 5.5%#349 | Artificial Analysis | |
| GPQA Diamond (Vals)not in index | – | 74.5%#80 | Vals AI | |
| Kagi LLM Benchmark | – | 47.4%#90 | Kagi LLM Benchmark |
Coding
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| SciCode | – | 38.4%#120 | SciCode | |
| LMArena Coding | – | 1458#118 | LMArena | |
| LMArena WebDev | – | 1341#93 | LMArena | |
| LiveCodeBench | – | 81.0%#62 | Vals AI | |
| IOI | – | 4.3%#46 | Vals AI | |
| SWE-Bench Pro | – | 9.7%#22 | Scale AI SEAL | |
| SWE-bench Verified (bash only) | – | 55.4%#28 | SWE-bench | |
| SWE-bench Verified (any scaffold)not in index | – | 55.4%#36 | SWE-bench |
Agents & tools
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| Terminal-Bench | – | 24.5%#50 | Terminal-Bench | |
| APEX-Agents | – | 4.0%#59 | Mercor | |
| Terminal-Bench Hard | 25.0%#133 | 28.8%#115 | Artificial Analysis | |
| τ²-Bench Telecom (AA)not in index | 70.5%#150 | 76.9%#129 | Artificial Analysis | |
| BFCL Overall | 72.4%#4 | – | Berkeley Function Calling Leaderboard |
Maths
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| AIME (Vals) | – | 92.7%#17 | Vals AI | |
| MGSM | – | 89.8%#42 | Vals AI | |
| MathArena Apex | – | 0.5%#41 | MathArena |
Knowledge
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| AA-Omniscience | -41.9#265 | -31.7#222 | Artificial Analysis | |
| MMLU-Pro | – | 82.2%#76 | Vals AI | |
| LegalBench | – | 79.6%#87 | Vals AI | |
| CorpFin | – | 56.8%#86 | Vals AI | |
| TaxEval | – | 66.2%#110 | Vals AI | |
| MedQA | – | 92.2%#33 | Vals AI |
Instruction following
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| IFBench | 43.4%#219 | 36.7%#282 | Artificial Analysis |
Human preference
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| LMArena Text | – | 1425#102 | LMArena |
Long context
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| AA-LCR | 54.0%#239 | 26.3%#335 | Artificial Analysis |
Composite
| Benchmark | thinking | default | Source | Trend |
|---|---|---|---|---|
| Epoch Capabilities Indexnot in index | – | 140.8#95 | Epoch AI Benchmarking Hub | |
| AA Intelligence Index | 18.5#190 | 14.9#223 | Artificial Analysis |
Where it wins
Benchmarks where this configuration ranks in the top five of every configuration measured.
Where to run it
Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.
| Provider | Speed | First token | Input $/M | Output $/M | Context | Quantisation |
|---|---|---|---|---|---|---|
| Venice | 60 tok/s | 1.07 s | $0.430 | $1.75 | 198k | fp4 |
| AtlasCloud | 44 tok/s | 1.15 s | $0.600 | $2.20 | 203k | fp8 |
| NovitaAI | 28 tok/s | 2.35 s | $0.550 | $2.20 | 205k | bf16 |
| Z.ai | 27 tok/s | 12 s | $0.600 | $2.20 | 203k | fp4 |
| DeepInfra | 21 tok/s | 1.51 s | $0.500 | $2.00 | 203k | fp4 |
Price history
Listed price per 1M tokens over time, as recorded by OpenRouter for the provider with the longest history.
What a task costs
Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.110 per 1M. Reasoning tokens are not modelled.
| Workload | Tokens in / out | Cost | With caching | Time |
|---|---|---|---|---|
| Chat reply | 400 / 300 | $0.0009 | $0.0007 | 54.8 s |
| Summarise a 30-page report | 12,000 / 600 | $0.0079 | $0.0040 | 1.0 min |
| Code edit | 6,000 / 1,500 | $0.0066 | $0.0046 | 1.4 min |
| Agentic coding session | 60,000 / 4,000 | $0.042 | $0.022 | 2.3 min |
| Structured extraction | 2,000 / 200 | $0.0015 | $0.0009 | 52.6 s |
See also
Data as of 9 Sept 2026. Compare these configurations.