LLM Opex

Models

OpenAI gpt-oss 20B: price on every cloud

Open weights · 21 billion parameters · Mixture of experts · 131,072-token context · licence apache-2.0

Per token, by cloud

US dollars per million tokens, each cloud's cheapest region.
CloudInput, per 1M tokensOutput, per 1M tokensCheapest region
DeepInfra$0.03$0.14global
AWS$0.07$0.30eu-north-1 · 15 regions
Azure$0.07$0.30australiaeast · 36 regions
Databricks$0.07$0.30us-east-1
OCI$0.07$0.30global · 3 regions
Groq$0.075$0.30global
Core42$0.10$0.30uae
Cloudflare$0.20$0.30global
Alibaba Cloud——not sold per token
Anthropic——not sold per token
Crusoe——not sold per token
Fireworks AI——not sold per token
Gemini API——not sold per token
Google Cloud——not sold per token
HF Endpoints——not sold per token
Huawei Cloud——not sold per token
IBM watsonx——not sold per token
Mistral——not sold per token
Nebius——not sold per token
OpenAI——not sold per token
Snowflake——not sold per token
Telnyx——not sold per token
Together AI——not sold per token
xAI——not sold per token

Run it yourself

It needs 13.1 GB of GPU memory at int4, the precision it ships in.

CloudCheapest instance that fitsGPUsAn hourA month, around the clock
HF Endpointsaws.nvidia-t4.x11 × T4 14 GB$0.50$365
AWSg4dn.xlarge1 × T4 16 GB$0.526$384
AzureStandard_NC4as_T4_v31 × T4 16 GB$0.526$384
Google Cloudg2-standard-41 × L4 24 GB$0.7068$516
Alibaba Cloudecs.gn6i-c4g1.xlarge1 × T4 16 GB$1.12$816
DeepInfraa100.1x1 × A100 80 GB$1.49$1,088
Crusoel40s.1x1 × L40S 48 GB$1.50$1,095
Nebiusgpu-l40s-a.1gpu-8vcpu-32gb1 × L40S 48 GB$1.55$1,130
OCIVM.GPU.A10.11 × A10 24 GB$2.00$1,460
Core42h100.1x1 × H100 80 GB$2.50$1,825
IBM watsonxdeploy-on-demand.1xL40S1 × L40S 48 GB$4.43$3,234
Fireworks AIh100.1x1 × H100 80 GB$8.00$5,840
Together AIhgx-h100.8gpu8 × H100 80 GB$31.92$23,302

At the cheapest per-token price, running it on HF Endpoints aws.nvidia-t4.x1 pays off above roughly 6.3 billion tokens a month.

More from OpenAI

Similar size, open weights

Inside the UAE: sold in-country by Azure, Core42.

List prices collected 21 November 2026, refreshed every three hours. No negotiated rates or taxes.