openai Open weights Reasoning

gpt-oss-20b pricing & benchmarks

Released Aug 5, 2025 · served by 11 providers

Input / 1M tokens$0.030
Output / 1M tokens$0.130
Context window131K
Cached input
Intelligence index14.9
Coding index20.7
Agentic index3.1
1M in + 300K out$0.07

What gpt-oss-20b is

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...

Providers and prices for gpt-oss-20b

The same model costs different amounts depending on who serves it. Prices are per 1M tokens, cheapest first. Uptime and throughput are OpenRouter's rolling measurements, not vendor claims.

ProviderInputOutputContext QuantUptime 24hThroughput
CoreWeavecheapest $0.030 $0.130 131K fp4 99.4%
DeepInfra $0.030 $0.140 131K bf16 99.8%
Parasail $0.030 $0.150 131K fp4 80.9%
Novita $0.040 $0.150 131K fp4 94.9%
Phala $0.040 $0.150 131K 93.6%
Amazon Bedrock $0.070 $0.150 131K 99.2%
SiliconFlow $0.040 $0.180 131K fp8 96.7%
Together $0.050 $0.200 131K 93.5%
Google $0.070 $0.250 131K 97.4%
Fireworks $0.070 $0.300 131K 99.3%
Groq $0.075 $0.300 131K 99.4%

gpt-oss-20b benchmark results

Design Arena head-to-head results, as reported through the OpenRouter model API.

CategoryArenaRankEloWin rate
dataviz models #100 963 39.7%
website models #115 877 27.9%

Cheaper models in the same class

Models scoring within 4 points of gpt-oss-20b on the Intelligence Index, but with a lower output price.

FAQ

How much does the gpt-oss-20b API cost?

$0.030 per 1M input tokens and $0.130 per 1M output tokens. A typical workload of 1M input + 300K output tokens costs about $0.07.

Is gpt-oss-20b free?

No. It is a paid model starting at $0.030 per 1M input tokens, though some providers offer trial credits.

Which provider is cheapest for gpt-oss-20b?

CoreWeave at $0.030 input / $0.130 output per 1M tokens (fp4 quantization).

Can I self-host gpt-oss-20b?

Yes — weights are published as openai/gpt-oss-20b on Hugging Face, so you can run it on your own hardware.