DeepSeek: DeepSeek V3.2

DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fi…

deepseek/deepseek-v3.2-20251201

DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that reduces training and inference cost while preserving quality in long-context scenarios. A scalable reinforcement learning post-training framework further improves reasoning, with reported performance in the GPT-5 class, and the model has demonstrated gold-medal results on the 2025 IMO and IOI. V3.2 also uses a large-scale agentic task synthesis pipeline to better integrate reasoning into tool-use settings, boosting compliance and generalization in interactive environments.

Users can control the reasoning behaviour with the `reasoning` `enabled` boolean. [Learn more in our docs](https://openrouter.ai/docs/use-cases/reasoning-tokens#enable-reasoning-with-default-config)

Specifications & capabilities

Input
text
Output
text
Context window · tokens
163840
Knowledge cutoff
2024-07
Published model ID
DeepSeek-V3.2
Capabilities
promptCaching, reasoning, structuredOutputs, toolCalling
Description source ↗

Evaluations

1–20 / 36

OpenRouter Benchmarks

72.8%

source id
openRouterBenchmarks
leaderboard name
tau2-bench Verified Airline · Accuracy · Friendli
dimension
agentic
metric
tau2-bench Verified Airline · Accuracy
rank
4
sample size
250
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
5
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
tau_bench_verified_airline
name
tau2-bench Verified Airline
record_kind
aggregate
provider_name
Friendli
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-17T19:16:53.251Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
tau2-bench Verified Airline · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
250
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/tau2-bench-airline
measurements
accuracy
0.728
accuracy_std_dev
0.018086336157442175
cost_per_task_usd
0.10671685199999999
output_tokens_per_task
14424.78
latency_seconds_per_task
361.548212
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
7
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenCompass LLM

85.8%

source id
openCompass
leaderboard name
MMLU-Pro
dimension
general
metric
MMLU-Pro
rank
5
aggregation scope
status
published
direction
higherIsBetter
conditions
suite
OverallTable
catalog
open
snapshot
month
REALTIME
fileName
assets/research-rank/research-data.REALTIME.20260309.json
updateTime
2026-03-09
columnFileName
assets/research-rank/research-column.REALTIME.20260309.json
metric_kind
MMLU-Pro
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
0-data.json
source_revision
117c5effa3066abf39f3e8f61109707b09f0cd95223246c07acd6289f4f7fcd2
metric_population
48
benchmark_evaluation
true
source model id
DeepSeek-V3.2
Original result ↗

OpenRouter Benchmarks

60.595%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · Mara
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
5
sample size
269
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
3
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
Mara
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T16:08:33.699Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
269
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.6059481003717473
accuracy_std_dev
0.024067533945591992
cost_per_task_usd
0.03276318587360595
output_tokens_per_task
7032.501858736059
latency_seconds_per_task
68.52027137546469
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
5
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

81.532%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · Friendli
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
6
sample size
990
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
5
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
Friendli
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-17T16:09:25.569Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
990
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.8153198
accuracy_std_dev
0.009214558967199644
cost_per_task_usd
0.030977021212121213
output_tokens_per_task
20456.146464646463
latency_seconds_per_task
329.3202121212121
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
7
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

58.736%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · SambaNova
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
7
sample size
662
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
7
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
SambaNova
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T16:41:30.138Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
662
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.5873616193353474
accuracy_std_dev
0.048784153240032144
cost_per_task_usd
0.03112247280966767
output_tokens_per_task
6675.285498489426
latency_seconds_per_task
62.86141993957704
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
8
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenCompass LLM

23.2%

source id
openCompass
leaderboard name
HLE
dimension
general
metric
HLE
rank
8
aggregation scope
status
published
direction
higherIsBetter
conditions
suite
OverallTable
catalog
open
snapshot
month
REALTIME
fileName
assets/research-rank/research-data.REALTIME.20260309.json
updateTime
2026-03-09
columnFileName
assets/research-rank/research-column.REALTIME.20260309.json
metric_kind
HLE
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
0-data.json
source_revision
117c5effa3066abf39f3e8f61109707b09f0cd95223246c07acd6289f4f7fcd2
metric_population
48
benchmark_evaluation
true
source model id
DeepSeek-V3.2
Original result ↗

OpenCompass LLM

93%

source id
openCompass
leaderboard name
AIME2025
dimension
general
metric
AIME2025
rank
9
aggregation scope
status
published
direction
higherIsBetter
conditions
suite
OverallTable
catalog
open
snapshot
month
REALTIME
fileName
assets/research-rank/research-data.REALTIME.20260309.json
updateTime
2026-03-09
columnFileName
assets/research-rank/research-column.REALTIME.20260309.json
metric_kind
AIME2025
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
0-data.json
source_revision
117c5effa3066abf39f3e8f61109707b09f0cd95223246c07acd6289f4f7fcd2
metric_population
48
benchmark_evaluation
true
source model id
DeepSeek-V3.2
Original result ↗

OpenCompass LLM

84.6%

source id
openCompass
leaderboard name
GPQA-Diamond
dimension
general
metric
GPQA-Diamond
rank
9
aggregation scope
status
published
direction
higherIsBetter
conditions
suite
OverallTable
catalog
open
snapshot
month
REALTIME
fileName
assets/research-rank/research-data.REALTIME.20260309.json
updateTime
2026-03-09
columnFileName
assets/research-rank/research-column.REALTIME.20260309.json
metric_kind
GPQA-Diamond
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
0-data.json
source_revision
117c5effa3066abf39f3e8f61109707b09f0cd95223246c07acd6289f4f7fcd2
metric_population
48
benchmark_evaluation
true
source model id
DeepSeek-V3.2
Original result ↗

OpenCompass LLM

75.3%

source id
openCompass
leaderboard name
Average
dimension
general
metric
Average
rank
9
aggregation scope
status
published
direction
higherIsBetter
conditions
suite
OverallTable
catalog
open
snapshot
month
REALTIME
fileName
assets/research-rank/research-data.REALTIME.20260309.json
updateTime
2026-03-09
columnFileName
assets/research-rank/research-column.REALTIME.20260309.json
metric_kind
Average
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
0-data.json
source_revision
117c5effa3066abf39f3e8f61109707b09f0cd95223246c07acd6289f4f7fcd2
metric_population
48
benchmark_evaluation
true
source model id
DeepSeek-V3.2
Original result ↗

OpenRouter Benchmarks

82.448%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · Baidu
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
10
sample size
1563
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
8
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
Baidu
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-11-10T18:18:26.718Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
1563
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.8244826858605246
accuracy_std_dev
0.009259257906785777
cost_per_task_usd
0.008255877799104288
output_tokens_per_task
19223.356365962893
latency_seconds_per_task
352.54179846449136
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
12
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

73.667%

source id
openRouterBenchmarks
leaderboard name
tau2-bench Verified Airline · Accuracy · StreamLake
dimension
agentic
metric
tau2-bench Verified Airline · Accuracy
rank
11
sample size
100
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
2
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
tau_bench_verified_airline
name
tau2-bench Verified Airline
record_kind
aggregate
provider_name
StreamLake
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-08-29T16:31:42.782Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
tau2-bench Verified Airline · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
100
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/tau2-bench-airline
measurements
accuracy
0.736667
accuracy_std_dev
0.010000000000000009
cost_per_task_usd
0.02084315
output_tokens_per_task
14704.35
latency_seconds_per_task
478.62581
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
25
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenCompass LLM

75.4%

source id
openCompass
leaderboard name
LiveCodeBenchV6
dimension
general
metric
LiveCodeBenchV6
rank
13
aggregation scope
status
published
direction
higherIsBetter
conditions
suite
OverallTable
catalog
open
snapshot
month
REALTIME
fileName
assets/research-rank/research-data.REALTIME.20260309.json
updateTime
2026-03-09
columnFileName
assets/research-rank/research-column.REALTIME.20260309.json
metric_kind
LiveCodeBenchV6
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
source_file
0-data.json
source_revision
117c5effa3066abf39f3e8f61109707b09f0cd95223246c07acd6289f4f7fcd2
metric_population
48
benchmark_evaluation
true
source model id
DeepSeek-V3.2
Original result ↗

OpenRouter Benchmarks

73.969%

source id
openRouterBenchmarks
leaderboard name
tau2-bench Verified Airline · Accuracy · Venice
dimension
agentic
metric
tau2-bench Verified Airline · Accuracy
rank
14
sample size
299
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
6
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
tau_bench_verified_airline
name
tau2-bench Verified Airline
record_kind
aggregate
provider_name
Venice
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-24T19:40:41.765Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
tau2-bench Verified Airline · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
299
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/tau2-bench-airline
measurements
accuracy
0.739687695652174
accuracy_std_dev
0.02325959803799046
cost_per_task_usd
0.050316361204013386
output_tokens_per_task
13915.548494983277
latency_seconds_per_task
654.799304347826
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
38
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

73.444%

source id
openRouterBenchmarks
leaderboard name
tau2-bench Verified Airline · Accuracy · SiliconFlow
dimension
agentic
metric
tau2-bench Verified Airline · Accuracy
rank
15
sample size
300
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
6
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
tau_bench_verified_airline
name
tau2-bench Verified Airline
record_kind
aggregate
provider_name
SiliconFlow
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T20:02:53.411Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
tau2-bench Verified Airline · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
300
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/tau2-bench-airline
measurements
accuracy
0.7344443333333334
accuracy_std_dev
0.019404823126108514
cost_per_task_usd
0.056632510000000004
output_tokens_per_task
15944.676666666666
latency_seconds_per_task
908.3642133333334
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
28
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

71.905%

source id
openRouterBenchmarks
leaderboard name
tau2-bench Verified Airline · Accuracy · Phala
dimension
agentic
metric
tau2-bench Verified Airline · Accuracy
rank
17
sample size
350
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
7
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
tau_bench_verified_airline
name
tau2-bench Verified Airline
record_kind
aggregate
provider_name
Phala
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T19:47:23.162Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
tau2-bench Verified Airline · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
350
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/tau2-bench-airline
measurements
accuracy
0.7190477142857142
accuracy_std_dev
0.026047291719454555
cost_per_task_usd
0.21749334
output_tokens_per_task
14108.717142857144
latency_seconds_per_task
1099.8125457142858
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
23
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

82.941%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · Alibaba
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
18
sample size
594
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
3
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
Alibaba
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-04T00:17:30.035Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
594
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.8294050000000001
accuracy_std_dev
0.004827531943619518
cost_per_task_usd
0.02358655555555556
output_tokens_per_task
21014.917508417508
latency_seconds_per_task
376.7206851851852
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
29
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

82.256%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · DigitalOcean
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
18
sample size
1547
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
8
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
DigitalOcean
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T16:13:07.320Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
1547
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.8225597382029736
accuracy_std_dev
0.013482485406421735
cost_per_task_usd
0.018713002585649646
output_tokens_per_task
20229.648351648353
latency_seconds_per_task
601.9765061409179
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
23
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

80.079%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · GMICloud
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
18
sample size
1782
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
9
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
GMICloud
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-11-10T07:09:23.085Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
1782
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.8007854444444445
accuracy_std_dev
0.13709711112858092
cost_per_task_usd
0.009282588103254769
output_tokens_per_task
29453.966891133558
latency_seconds_per_task
461.3563843995511
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
26
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

83.237%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · AtlasCloud
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
19
sample size
1386
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
7
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
AtlasCloud
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T16:01:53.649Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
1386
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.8323711428571429
accuracy_std_dev
0.01158903863174871
cost_per_task_usd
0.008056092352092353
output_tokens_per_task
20733.657287157286
latency_seconds_per_task
525.7800266955267
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
30
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

OpenRouter Benchmarks

79.342%

source id
openRouterBenchmarks
leaderboard name
GPQA Diamond · Accuracy · Phala
dimension
knowledge
metric
GPQA Diamond · Accuracy
rank
19
sample size
1186
aggregation scope
status
published
harness
url
https://github.com/OpenRouterTeam/benchmark-harness
name
OpenRouter benchmark harness
direction
higherIsBetter
run_count
6
conditions
window
newest 90 days; successful visible nonzero runs
dataset
benchmark-leaderboard
benchmark
id
gpqa_diamond
name
GPQA Diamond
record_kind
aggregate
provider_name
Phala
attribution
Source: OpenRouter (openrouter.ai), licensed under CC BY 4.0: reuse and republish with attribution to OpenRouter.
last_run_at
2026-10-31T16:45:39.993Z
license_url
https://creativecommons.org/licenses/by/4.0/
methodology
Rows aggregate the benchmark-runs dataset exactly as the leaderboard pages do, over each benchmark's non-zero-score runs from the newest 90 days: accuracy is the task-weighted mean of per-run accuracy, accuracy_std_dev is the population standard deviation of per-run accuracy (null with a single run), and the per-task averages divide summed cost, generation time, and output tokens by summed tasks. Rows are ordered by accuracy descending, then model name.
metric_kind
GPQA Diamond · Accuracy
rank_method
competition rank among reported evaluations in identical conditions, including unlinked observations
sample_size
1186
source_file
benchmark-leaderboard.json
source_page
https://openrouter.ai/benchmarks/gpqa-diamond
measurements
accuracy
0.7934233709949409
accuracy_std_dev
0.05668542998401107
cost_per_task_usd
0.01885696121416526
output_tokens_per_task
18169.33305227656
latency_seconds_per_task
1103.0432150084318
published_at
2026-11-16T01:46:02.653Z
snapshot_date
2026-11-16
source_revision
7ccaf17664252a3053dcb7df8ad0a46e72954a3b4524ddc1c8a7ef8a41b15568
transformations
TensorTrace preserves source results, converts latency to seconds and derives ranks within matching conditions. Individual runs have no derived rank.
metric_population
24
benchmark_evaluation
true
openrouter_permaslug
deepseek/deepseek-v3.2-20251201
source model id
deepseek/deepseek-v3.2-20251201
Original result ↗

Published prices

1–13 / 13

deepseek/deepseek-v3.2 ↗

$0.13

Inference provider
openRouter
Platform
OpenRouter
Unit
cache_read_tokens_per_million
Currency
USD
Billing conditions

deepseek/deepseek-v3.2 ↗

$0.26

Inference provider
openRouter
Platform
OpenRouter
Unit
input_tokens_per_million
Currency
USD
Billing conditions

deepseek/deepseek-v3.2 ↗

$0.38

Inference provider
openRouter
Platform
OpenRouter
Unit
output_tokens_per_million
Currency
USD
Billing conditions

deepseek-ai/DeepSeek-V3.2 ↗

$0.14

Inference provider
siliconFlow
Platform
SiliconFlow
Unit
cache_read_tokens_per_million
Currency
USD
Billing conditions
Billing label
Cached input tokens (per 1M)

deepseek-ai/DeepSeek-V3.2 ↗

$0.27

Inference provider
siliconFlow
Platform
SiliconFlow
Unit
input_tokens_per_million
Currency
USD
Billing conditions
Billing label
Input tokens (per 1M)

deepseek-ai/DeepSeek-V3.2 ↗

$0.42

Inference provider
siliconFlow
Platform
SiliconFlow
Unit
output_tokens_per_million
Currency
USD
Billing conditions
Billing label
Output tokens (per 1M)

deepseek/deepseek-v3.2 ↗

$0.62

Inference provider
vercelAIGatewayCatalog
Platform
bedrock
Unit
input_tokens_per_million
Currency
USD
Billing conditions
endpoint
bedrock | deepseek/deepseek-v3.2
model_type
language
billing_scope
via Vercel AI Gateway
Billing label
Input tokens / 1M
Pricing via
Vercel AI Gateway

deepseek/deepseek-v3.2 ↗

$1.85

Inference provider
vercelAIGatewayCatalog
Platform
bedrock
Unit
output_tokens_per_million
Currency
USD
Billing conditions
endpoint
bedrock | deepseek/deepseek-v3.2
model_type
language
billing_scope
via Vercel AI Gateway
Billing label
Output tokens / 1M
Pricing via
Vercel AI Gateway

deepseek/deepseek-v3.2 ↗

$0.13

Inference provider
vercelAIGatewayCatalog
Platform
deepinfra
Unit
cache_read_tokens_per_million
Currency
USD
Billing conditions
endpoint
deepinfra | deepseek/deepseek-v3.2
model_type
language
quantization
fp4
billing_scope
via Vercel AI Gateway
Billing label
Cached input / 1M
Pricing via
Vercel AI Gateway

deepseek/deepseek-v3.2 ↗

$0.26

Inference provider
vercelAIGatewayCatalog
Platform
deepinfra
Unit
input_tokens_per_million
Currency
USD
Billing conditions
endpoint
deepinfra | deepseek/deepseek-v3.2
model_type
language
quantization
fp4
billing_scope
via Vercel AI Gateway
Billing label
Input tokens / 1M
Pricing via
Vercel AI Gateway

deepseek/deepseek-v3.2 ↗

$0.38

Inference provider
vercelAIGatewayCatalog
Platform
deepinfra
Unit
output_tokens_per_million
Currency
USD
Billing conditions
endpoint
deepinfra | deepseek/deepseek-v3.2
model_type
language
quantization
fp4
billing_scope
via Vercel AI Gateway
Billing label
Output tokens / 1M
Pricing via
Vercel AI Gateway

deepseek/deepseek-v3.2 ↗

$0.62

Inference provider
vercelAIGatewayCatalog
Platform
Vercel AI Gateway
Unit
input_tokens_per_million
Currency
USD
Billing conditions
model_type
language
quote_scope
gateway_catalog
billing_scope
Vercel public catalog; endpoint prices may differ
varies_by_provider
true
Billing label
Input tokens / 1M
Pricing via
Vercel AI Gateway

deepseek/deepseek-v3.2 ↗

$1.85

Inference provider
vercelAIGatewayCatalog
Platform
Vercel AI Gateway
Unit
output_tokens_per_million
Currency
USD
Billing conditions
model_type
language
quote_scope
gateway_catalog
billing_scope
Vercel public catalog; endpoint prices may differ
varies_by_provider
true
Billing label
Output tokens / 1M
Pricing via
Vercel AI Gateway

Provider specifications & access

1–4 / 4

DeepSeek-V3.2 ↗

source model id
deepseek-ai/DeepSeek-V3.2
hosting
siliconflow-hosted
platforms
  • SiliconFlow
max output tokens
163840
architecture
input_modalities
  • text
output_modalities
  • text
capabilities
  • promptCaching
  • reasoning
  • structuredOutputs
  • toolCalling

DeepSeek V3.2 ↗

source model id
deepseek/deepseek-v3.2
hosting
vercel-gateway
platforms
  • Vercel AI Gateway
context length
128000
max output tokens
8000
architecture
input_modalities
  • text
output_modalities
  • text
capabilities
  • promptCaching
  • toolCalling

deepseek-ai/DeepSeek-V3.2 ↗

source model id
publishers/deepseek-ai/models/deepseek-v3-2@deepseek-v3-2
hosting
vertex-self-deployed
platforms
  • Google Cloud Vertex AI
architecture
input_modalities
  • text
output_modalities
  • text
capabilities

    deepseek-ai/DeepSeek-V3.2 ↗

    source model id
    publishers/hf-deepseek-ai/models/deepseek-v3.2@001
    hosting
    vertex-self-deployed
    platforms
    • Google Cloud Vertex AI
    architecture
    input_modalities
    • text
    output_modalities
    • text
    capabilities

      Safety assessments

      1–1 / 1

      Foundation Model Transparency Index 2025 ↗

      32 / 100

      source id
      stanfordFMTI
      canonical model id
      deepseek/deepseek-v3.2-20251201
      taxonomy
      transparency
      hazard category
      transparency
      hazard name
      Organization transparency
      policy area
      DeepSeek
      assessment type
      organization-level disclosure
      metric
      disclosure_score
      result label
      32 / 100
      sample size
      100
      updated at
      2026-11-09T18:08:05.227Z

      Source specification observations

      1–7 / 7

      openRouter ↗

      field
      context_length
      value
      163840
      normalized value
      163840
      scope
      source-model
      source id
      openRouter
      confidence
      1

      siliconFlow ↗

      field
      context_length
      value
      164K
      scope
      source-model
      source id
      siliconFlow
      confidence
      1

      vercelAIGatewayCatalog ↗

      field
      context_length
      value
      128000
      normalized value
      128000
      scope
      source-model
      source id
      vercelAIGatewayCatalog
      confidence
      1

      vercelAIGatewayCatalog ↗

      field
      knowledge_cutoff
      reported field
      knowledge_cutoff
      value
      2024-07
      normalized value
      2024-07
      precision
      month
      scope
      source-model
      source id
      vercelAIGatewayCatalog
      confidence
      1

      openRouter ↗

      field
      max_output_tokens
      value
      147456
      normalized value
      147456
      scope
      source-model
      source id
      openRouter
      confidence
      1

      siliconFlow ↗

      field
      max_output_tokens
      value
      164K
      scope
      source-model
      source id
      siliconFlow
      confidence
      1

      vercelAIGatewayCatalog ↗

      field
      max_output_tokens
      value
      8000
      normalized value
      8000
      scope
      source-model
      source id
      vercelAIGatewayCatalog
      confidence
      1

      Original sources

      1–16 / 16

      SiliconFlow · Model page ↗

      Hugging Face · Model repository ↗

      Google Cloud · Model page ↗

      Google Cloud · Model page ↗

      OpenRouter · Model page ↗

      SiliconFlow · Playground ↗

      Vercel AI Gateway · Model page ↗

      SiliconFlow · API documentation ↗

      GitHub · Catalog ↗

      Google Cloud · Catalog ↗

      Hugging Face · Catalog ↗

      Hugging Face · Catalog ↗

      OpenRouter · Catalog ↗

      OpenRouter · Catalog ↗

      SiliconFlow · Catalog ↗

      Vercel AI Gateway · Catalog ↗

      Source links

      1–9 / 9

      Hugging Face model card

      1–1 / 1

      Explore more models

      Models

      Loading
      EVIDENCE / COMPARISON

      Benchmarks

      Explore every benchmark

      Loading

      These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

      Install TensorTrace