RedRouterContinuous inference provider benchmarks
Updated Trailing 72 hours
Provider comparison
DeepSeek V4.1 Flash 1017- DeepSeek V4.1 Flash (undated derivative)
GLM 5.3 Flash- GLM 5.3
 Experiential Labsvia dealignai |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.394 | $0.225 | $0.900 | 243.4 | 0.6 | 2.7 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:05 UTC |
 NanoGPT → Heabsy |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.610 | $0.350 | $1.390 | 227.4 | 1.3 | 3.7 | 62.5%20 / 32 requests | failing | 07 Nov 2026, 18:14 UTC |
| NanoGPT → OrcaRouter |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $1.050 | $0.840 | $1.680 | 158.0 | 1.4 | 4.9 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Sable |
DeepSeek V4.1 Flash 1017 | heretic | NATIVE | 512,000 | $0.303 | $0.173 | $0.693 | 135.8 | 1.5 | 5.3 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:14 UTC |
 OpenRoutervia morph/fp8 |
DeepSeek V4.1 Flash 1017 | none | FP8 | 1,048,576 | $0.117 | $0.022 | $0.400 | 119.4 | 0.8 | 5.5 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:11 UTC |
 Venice |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.578 | $0.330 | $1.320 | 190.7 | 0.6 | 3.2 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:10 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- Experiential Labs · dealignai
- NanoGPT → Heabsy
- NanoGPT → OrcaRouter
- NanoGPT → Sable
- OpenRouter · morph/fp8
- Venice
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| Experiential Labsvia dealignai |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.394 | $0.225 | $0.900 | 250.6 | 0.7 | 2.8 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:05 UTC |
| NanoGPT → Heabsy |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.610 | $0.350 | $1.390 | 182.6 | 1.5 | 3.9 | 62.5%10 / 16 requests | failing | 07 Nov 2026, 18:14 UTC |
| NanoGPT → OrcaRouter |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $1.050 | $0.840 | $1.680 | 148.8 | 1.3 | 4.6 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Sable |
DeepSeek V4.1 Flash 1017 | heretic | NATIVE | 512,000 | $0.303 | $0.173 | $0.693 | 135.8 | 1.5 | 5.2 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| OpenRoutervia morph/fp8 |
DeepSeek V4.1 Flash 1017 | none | FP8 | 1,048,576 | $0.117 | $0.022 | $0.400 | 107.4 | 1.0 | 5.8 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
| Venice |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.578 | $0.330 | $1.320 | 213.7 | 0.5 | 2.9 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- Experiential Labs · dealignai
- NanoGPT → Heabsy
- NanoGPT → OrcaRouter
- NanoGPT → Sable
- OpenRouter · morph/fp8
- Venice
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| Experiential Labsvia dealignai |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.394 | $0.225 | $0.900 | 230.8 | 0.5 | 2.7 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:05 UTC |
| NanoGPT → Heabsy |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.610 | $0.350 | $1.390 | 289.9 | 1.2 | 2.5 | 62.5%10 / 16 requests | failing | 07 Nov 2026, 18:14 UTC |
| NanoGPT → OrcaRouter |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $1.050 | $0.840 | $1.680 | 163.0 | 1.6 | 5.1 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Sable |
DeepSeek V4.1 Flash 1017 | heretic | NATIVE | 512,000 | $0.303 | $0.173 | $0.693 | 134.5 | 1.5 | 5.4 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| OpenRoutervia morph/fp8 |
DeepSeek V4.1 Flash 1017 | none | FP8 | 1,048,576 | $0.117 | $0.022 | $0.400 | 122.6 | 0.7 | 5.3 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
| Venice |
DeepSeek V4.1 Flash 1017 | uncensored | Unknown | 1,000,000 | $0.578 | $0.330 | $1.320 | 182.2 | 0.6 | 3.4 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- Experiential Labs · dealignai
- NanoGPT → Heabsy
- NanoGPT → OrcaRouter
- NanoGPT → Sable
- OpenRouter · morph/fp8
- Venice
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
 abliteration.aivia Abliterated Large Flash |
DeepSeek V4.1 Flash (undated derivative) | abliterated | Unknown | 1,048,576 | $1.250 | $1.000 | $2.000 | 350.1 | 1.1 | 2.7 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:10 UTC |
| NanoGPT → abliteration.ai |
DeepSeek V4.1 Flash (undated derivative) | abliterated | FP8 | 1,000,000 | $1.312 | $1.050 | $2.100 | 300.8 | 2.3 | 4.0 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Venice |
DeepSeek V4.1 Flash (undated derivative) | uncensored | FP8 | 524,288 | $0.608 | $0.347 | $1.390 | 181.4 | 1.8 | 4.5 | 75.0%24 / 32 requests | failing | 07 Nov 2026, 18:13 UTC |
 Refuseless |
DeepSeek V4.1 Flash (undated derivative) | abliterated | FP8 | 1,048,576 | $0.203 | $0.120 | $0.450 | 140.3 | 1.7 | 5.3 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:11 UTC |
 Yeta Labs |
DeepSeek V4.1 Flash (undated derivative) | uncensored | FP8 | 1,000,000 | $0.147 | $0.080 | $0.350 | 140.8 | 1.7 | 5.1 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:04 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- abliteration.ai · Abliterated Large Flash
- NanoGPT → abliteration.ai
- NanoGPT → Venice
- Refuseless
- Yeta Labs
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| abliteration.aivia Abliterated Large Flash |
DeepSeek V4.1 Flash (undated derivative) | abliterated | Unknown | 1,048,576 | $1.250 | $1.000 | $2.000 | 306.5 | 1.0 | 2.9 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
| NanoGPT → abliteration.ai |
DeepSeek V4.1 Flash (undated derivative) | abliterated | FP8 | 1,000,000 | $1.312 | $1.050 | $2.100 | 283.9 | 2.2 | 3.8 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Venice |
DeepSeek V4.1 Flash (undated derivative) | uncensored | FP8 | 524,288 | $0.608 | $0.347 | $1.390 | 164.8 | 1.7 | 4.7 | 75.0%12 / 16 requests | failing | 07 Nov 2026, 18:13 UTC |
| Refuseless |
DeepSeek V4.1 Flash (undated derivative) | abliterated | FP8 | 1,048,576 | $0.203 | $0.120 | $0.450 | 128.4 | 1.7 | 5.9 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
| Yeta Labs |
DeepSeek V4.1 Flash (undated derivative) | uncensored | FP8 | 1,000,000 | $0.147 | $0.080 | $0.350 | 141.8 | 1.7 | 5.2 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:04 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- abliteration.ai · Abliterated Large Flash
- NanoGPT → abliteration.ai
- NanoGPT → Venice
- Refuseless
- Yeta Labs
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| abliteration.aivia Abliterated Large Flash |
DeepSeek V4.1 Flash (undated derivative) | abliterated | Unknown | 1,048,576 | $1.250 | $1.000 | $2.000 | 390.3 | 1.2 | 2.5 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
| NanoGPT → abliteration.ai |
DeepSeek V4.1 Flash (undated derivative) | abliterated | FP8 | 1,000,000 | $1.312 | $1.050 | $2.100 | 314.8 | 2.5 | 4.0 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Venice |
DeepSeek V4.1 Flash (undated derivative) | uncensored | FP8 | 524,288 | $0.608 | $0.347 | $1.390 | 216.3 | 1.9 | 4.1 | 75.0%12 / 16 requests | failing | 07 Nov 2026, 18:13 UTC |
| Refuseless |
DeepSeek V4.1 Flash (undated derivative) | abliterated | FP8 | 1,048,576 | $0.203 | $0.120 | $0.450 | 148.3 | 2.2 | 5.1 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
| Yeta Labs |
DeepSeek V4.1 Flash (undated derivative) | uncensored | FP8 | 1,000,000 | $0.147 | $0.080 | $0.350 | 140.8 | 1.5 | 4.8 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:04 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- abliteration.ai · Abliterated Large Flash
- NanoGPT → abliteration.ai
- NanoGPT → Venice
- Refuseless
- Yeta Labs
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| Experiential Labsvia dealignai |
GLM 5.3 Flash | uncensored | Unknown | 1,000,000 | $0.237 | $0.150 | $0.500 | 76.8 | 0.9 | 7.5 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:09 UTC |
| NanoGPT → Arnict |
GLM 5.3 Flash | uncensored | NVFP4 | 1,000,000 | $0.230 | $0.131 | $0.525 | 195.6 | 1.2 | 3.6 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:13 UTC |
| NanoGPT → Heabsy |
GLM 5.3 Flash | uncensored | NVFP4 | 1,000,000 | $0.350 | $0.200 | $0.800 | 219.3 | 1.0 | 3.4 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:15 UTC |
| NanoGPT → OrcaRouter |
GLM 5.3 Flash | uncensored | Unknown | 1,000,000 | $1.050 | $0.840 | $1.680 | 50.2 | 2.2 | 12.2 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:15 UTC |
| OpenRoutervia z-ai/fp8 |
GLM 5.3 Flash | none | FP8 | 1,048,576 | $0.237 | $0.150 | $0.500 | 51.8 | 1.6 | 12.1 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:10 UTC |
| Refuseless |
GLM 5.3 Flash | abliterated | NVFP4 | 1,048,576 | $0.215 | $0.120 | $0.500 | 53.0 | 1.6 | 12.9 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:11 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- Experiential Labs · dealignai
- NanoGPT → Arnict
- NanoGPT → Heabsy
- NanoGPT → OrcaRouter
- OpenRouter · z-ai/fp8
- Refuseless
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| Experiential Labsvia dealignai |
GLM 5.3 Flash | uncensored | Unknown | 1,000,000 | $0.237 | $0.150 | $0.500 | 77.1 | 0.9 | 7.4 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:09 UTC |
| NanoGPT → Arnict |
GLM 5.3 Flash | uncensored | NVFP4 | 1,000,000 | $0.230 | $0.131 | $0.525 | 213.3 | 1.0 | 3.4 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:13 UTC |
| NanoGPT → Heabsy |
GLM 5.3 Flash | uncensored | NVFP4 | 1,000,000 | $0.350 | $0.200 | $0.800 | 212.7 | 0.9 | 3.3 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:15 UTC |
| NanoGPT → OrcaRouter |
GLM 5.3 Flash | uncensored | Unknown | 1,000,000 | $1.050 | $0.840 | $1.680 | 50.6 | 2.0 | 11.8 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:15 UTC |
| OpenRoutervia z-ai/fp8 |
GLM 5.3 Flash | none | FP8 | 1,048,576 | $0.237 | $0.150 | $0.500 | 51.8 | 1.6 | 11.7 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
| Refuseless |
GLM 5.3 Flash | abliterated | NVFP4 | 1,048,576 | $0.215 | $0.120 | $0.500 | 51.6 | 2.2 | 16.2 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- Experiential Labs · dealignai
- NanoGPT → Arnict
- NanoGPT → Heabsy
- NanoGPT → OrcaRouter
- OpenRouter · z-ai/fp8
- Refuseless
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| Experiential Labsvia dealignai |
GLM 5.3 Flash | uncensored | Unknown | 1,000,000 | $0.237 | $0.150 | $0.500 | 73.4 | 0.9 | 7.7 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:09 UTC |
| NanoGPT → Arnict |
GLM 5.3 Flash | uncensored | NVFP4 | 1,000,000 | $0.230 | $0.131 | $0.525 | 194.4 | 1.3 | 4.1 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:13 UTC |
| NanoGPT → Heabsy |
GLM 5.3 Flash | uncensored | NVFP4 | 1,000,000 | $0.350 | $0.200 | $0.800 | 226.9 | 1.2 | 3.7 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:15 UTC |
| NanoGPT → OrcaRouter |
GLM 5.3 Flash | uncensored | Unknown | 1,000,000 | $1.050 | $0.840 | $1.680 | 49.7 | 2.3 | 12.5 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:15 UTC |
| OpenRoutervia z-ai/fp8 |
GLM 5.3 Flash | none | FP8 | 1,048,576 | $0.237 | $0.150 | $0.500 | 52.4 | 1.6 | 12.7 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
| Refuseless |
GLM 5.3 Flash | abliterated | NVFP4 | 1,048,576 | $0.215 | $0.120 | $0.500 | 53.0 | 1.6 | 11.0 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- Experiential Labs · dealignai
- NanoGPT → Arnict
- NanoGPT → Heabsy
- NanoGPT → OrcaRouter
- OpenRouter · z-ai/fp8
- Refuseless
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| abliteration.aivia Abliterated Large v2 |
GLM 5.3 | abliterated | Unknown | 1,000,000 | $3.500 | $3.000 | $5.000 | 145.0 | 1.3 | 5.0 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:08 UTC |
| NanoGPT → abliteration.ai |
GLM 5.3 | abliterated | FP8 | 1,000,000 | $3.675 | $3.150 | $5.250 | 132.8 | 2.4 | 6.9 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Arnict |
GLM 5.3 | uncensored | NVFP4 | 1,000,000 | $1.302 | $1.040 | $2.090 | 69.5 | 2.3 | 9.7 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:12 UTC |
| OpenRoutervia z-ai/fp8 |
GLM 5.3 | none | FP8 | 1,048,576 | $2.150 | $1.400 | $4.400 | 68.0 | 1.2 | 8.8 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:11 UTC |
| Refuseless |
GLM 5.3 | abliterated | NVFP4 | 1,048,576 | $1.250 | $1.000 | $2.000 | 67.4 | 1.3 | 8.3 | 96.9%31 / 32 requests | ok | 07 Nov 2026, 18:10 UTC |
| Venicevia Abliterated Large v2 via abliteration.ai |
GLM 5.3 | abliterated | Unknown | 1,000,000 | $3.500 | $3.000 | $5.000 | 214.2 | 1.4 | 4.0 | 100.0%32 / 32 requests | ok | 07 Nov 2026, 18:08 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- abliteration.ai · Abliterated Large v2
- NanoGPT → abliteration.ai
- NanoGPT → Arnict
- OpenRouter · z-ai/fp8
- Refuseless
- Venice · Abliterated Large v2 via abliteration.ai
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| abliteration.aivia Abliterated Large v2 |
GLM 5.3 | abliterated | Unknown | 1,000,000 | $3.500 | $3.000 | $5.000 | 145.0 | 1.3 | 4.9 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:08 UTC |
| NanoGPT → abliteration.ai |
GLM 5.3 | abliterated | FP8 | 1,000,000 | $3.675 | $3.150 | $5.250 | 132.8 | 2.9 | 7.3 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Arnict |
GLM 5.3 | uncensored | NVFP4 | 1,000,000 | $1.302 | $1.040 | $2.090 | 70.4 | 2.6 | 9.4 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:12 UTC |
| OpenRoutervia z-ai/fp8 |
GLM 5.3 | none | FP8 | 1,048,576 | $2.150 | $1.400 | $4.400 | 65.4 | 1.2 | 9.0 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
| Refuseless |
GLM 5.3 | abliterated | NVFP4 | 1,048,576 | $1.250 | $1.000 | $2.000 | 63.1 | 1.3 | 8.7 | 93.8%15 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
| Venicevia Abliterated Large v2 via abliteration.ai |
GLM 5.3 | abliterated | Unknown | 1,000,000 | $3.500 | $3.000 | $5.000 | 208.3 | 1.4 | 3.9 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:08 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- abliteration.ai · Abliterated Large v2
- NanoGPT → abliteration.ai
- NanoGPT → Arnict
- OpenRouter · z-ai/fp8
- Refuseless
- Venice · Abliterated Large v2 via abliteration.ai
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗
| abliteration.aivia Abliterated Large v2 |
GLM 5.3 | abliterated | Unknown | 1,000,000 | $3.500 | $3.000 | $5.000 | 180.6 | 1.2 | 5.0 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:08 UTC |
| NanoGPT → abliteration.ai |
GLM 5.3 | abliterated | FP8 | 1,000,000 | $3.675 | $3.150 | $5.250 | 136.5 | 2.1 | 6.1 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:14 UTC |
| NanoGPT → Arnict |
GLM 5.3 | uncensored | NVFP4 | 1,000,000 | $1.302 | $1.040 | $2.090 | 69.5 | 2.1 | 10.1 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:12 UTC |
| OpenRoutervia z-ai/fp8 |
GLM 5.3 | none | FP8 | 1,048,576 | $2.150 | $1.400 | $4.400 | 69.5 | 1.2 | 8.4 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:11 UTC |
| Refuseless |
GLM 5.3 | abliterated | NVFP4 | 1,048,576 | $1.250 | $1.000 | $2.000 | 77.5 | 1.2 | 7.6 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:10 UTC |
| Venicevia Abliterated Large v2 via abliteration.ai |
GLM 5.3 | abliterated | Unknown | 1,000,000 | $3.500 | $3.000 | $5.000 | 217.1 | 1.5 | 4.0 | 100.0%16 / 16 requests | ok | 07 Nov 2026, 18:08 UTC |
p50 = median. Missing speed or latency values mean fewer than 8 usable successes. Success rate includes failed requests. Status and last success describe the route across all regions. Precision and context are provider declarations; different modifications are different products. Measurement details →
Output speed
tok/s · Higher is better · p50
Reported completion tokens after the first chunk, divided by streaming time. Samples without reported usage are excluded. Methodology ↗
Time to first token
s · Lower is better · p50
Request sent to first non-empty content, reasoning or tool-argument delta. Includes network time from the selected region. Methodology ↗
Time to 500 tokens
s · Lower is better · p50
Per request: TTFT + 500 ÷ output tok/s. Percentiles are computed after this calculation; this is a derived response time. Methodology ↗
Blended price
$ / 1M tokens · Lower is better
Declared price: (3 × input + output) ÷ 4. Excludes cache discounts. Prices are not measured billing receipts. Methodology ↗
Output speed vs. price
Median output speed and declared blended price
- abliteration.ai · Abliterated Large v2
- NanoGPT → abliteration.ai
- NanoGPT → Arnict
- OpenRouter · z-ai/fp8
- Refuseless
- Venice · Abliterated Large v2 via abliteration.ai
Each point is one route. Shading divides the displayed axes at their midpoints; it is not a quality rating. Blended price uses a 3:1 input/output mix. Routes without usable speed samples are omitted. Methodology ↗