OpenAI

gpt-oss-20b

Launched 2025-08-05 · Knowledge up to 2024-06 · Listed on OpenRouter 2025-08-05

Prices updated . Benchmarks updated .

gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license. It uses a Mixture-of-Experts (MoE) architecture with 3.6B active parameters per forward pass, optimized for...

You send / 1M

$0.018

tokens in

It replies / 1M

$0.090

tokens out

Context

131K

tokens it can read at once

Typical request

$0.0001

1,000 words in, 400 out

Benchmarks

BenchmarkScoreKindDateSource
Epoch Capabilities Index137.8Independent2025-08-05Epoch AI
GPQA DiamondMedium effort60.8%Independent2026-08-27Epoch AI
OTIS Mock AIMEMedium effort65.3%Independent2026-08-27Epoch AI
Terminal-Benchagent Mini-SWE-Agent3.4%Independent2025-11-03Terminal-Bench, republished by Epoch AI
CritPtHigh effort1.4%Independent2025-08-05CritPt, republished by Epoch AI
Arena Elo (text)1,287Independent2026-10-28LMArena

Rankings on this site use the Epoch Capabilities Index, an independent composite. Lab-reported rows are the lab's own figures and are kept separate.

Data privacy

No training or retention: Yes. At least one provider serves it with zero data retention.

ProviderYou send / 1MIt replies / 1MUptimePrivacy policy
AkashML · fp4$0.018$0.090100%Read
DeepInfra · bf16$0.024$0.11293%Read
CoreWeave · fp4$0.030$0.130100%Read
DekaLLM · bf16$0.029$0.140100%Read
DeepInfra · bf16$0.030$0.14098%Read
Amazon Bedrock$0.070$0.15099%Read
Amazon Bedrock$0.070$0.150100%Read
Parasail · fp4$0.030$0.150100%Read
PrimeIntellect · fp4$0.117$0.175100%Read
SiliconFlow · fp8$0.040$0.18089%Read
Groq$0.075$0.300100%Read

Based on OpenRouter and Vercel AI Gateway data. Confirm terms with the provider before sending sensitive data.

What it can do

  • Reads: text
  • Writes: text
  • Weights: Open weights
  • Reasoning: Yes
  • Tool calling: Yes
  • Longest reply: 131K tokens
  • Repeated text is billed at $0.0090 per 1M (prompt caching)

Compare with

Appears in

Estimate your monthly cost with gpt-oss-20b

Sources and licenses

  • OpenRouter: Prices, context windows, modalities and the model list come from the OpenRouter public API. Data: OpenRouter.
  • Vercel AI Gateway: The no-training flag comes from the Vercel AI Gateway model list.
  • models.dev: Lab launch dates, open-weights flags and knowledge cutoffs come from models.dev. Copyright (c) 2025 models.dev, MIT License.
  • Epoch AI: Independent benchmark scores and the Epoch Capabilities Index (ECI) come from Epoch AI's Benchmarking Hub, used under CC BY 4.0. Epoch AI, Capabilities & benchmarking, epoch.ai/benchmarks. Scores Epoch marks as external keep their original license, and the source is named on each score.
  • LMArena: Arena Elo (text, webdev and vision) comes from the LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0, leaderboard date 2026-10-28.
  • Lab-reported scores, when an editor has added them, are the lab's own numbers. Each one links to the source and is labeled lab-reported. They are not used to rank the lists.
  • What a score measures: the benchmark guide.