inference/research
verifiedmedium · 0.80updated 2026-10-16

Azure OpenAI Service / Azure AI Foundry inference

azure

Fields

id
services/svc-azure-ai-foundry
Provider
azure
Endpoint
https://{resource}.openai.azure.com/openai/v1
Input $/Mtok
2
Output $/Mtok
8
Price basis
USD per 1M TOKENS, representative model GPT-4.1-2025-04-14 Global standard deployment, observed 2026-10-14 ($2.00 in / $0.50 cached / $8.00 out; Data Zone $2.20/$0.55/$8.80; Regional $2.20/$0.55/$8.80). GPT-4.1-mini Global $0.40/$1.60, nano $0.10/$0.40. Three deployment SKUs change the price on the same model: Global, Data Zone (EU or US), Regional (up to 27 regions). Batch API = 50% of Global Standard. Provisioned throughput (PTU) is reserved-capacity pricing, not token-priced. Azure AI Foundry also hosts third-party models (Meta, DeepSeek, Mistral, xAI) each with its own catalog price.
Context window
1,048,576
TTFT (ms)
2,280
Output tok/s
132.2
Measured by
artificial-analysis
Regions
east-useast-us-2west-ussweden-centralwest-europeuk-southjapan-eastaustralia-east
notes
2026-10-16: the latency gap closed. Artificial Analysis measures Azure on Llama 3.3 70B at 132.2 output tok/s median and 2.28s median first chunk, context 128k — the fields were null on 2026-10-14 because no row extracted. The representative PRICE is still GPT-4.1 Global at $2.00/$8.00 from the 2026-10-14 Azure OpenAI pricing capture; that page was not re-fetched in this pass, so the price and price_basis are unchanged and the new measurement is for a different model than the price anchor. price_basis names the priced model; the measurement names Llama 3.3 70B. The OpenAI-on-Azure path: GPT-4.1's 1,048,576-token context is the biggest published context in this record set. Three deployment types (Global / Data Zone / Regional) are a data-residency choice that shows up directly in the per-1M price - Global is cheapest, Regional pins inference inside one of 27 regions. Artificial Analysis benchmarks Azure on Llama 3.3 70B; no measured row was extracted for this pass so the latency fields stay null. Foundry's model catalog (Meta, DeepSeek, xAI, Mistral) prices per model on the same page and is the 'other models' clause in the basis.

Notes

2026-10-16: the latency gap closed. Artificial Analysis measures Azure on Llama 3.3 70B at 132.2 output tok/s median and 2.28s median first chunk, context 128k — the fields were null on 2026-10-14 because no row extracted. The representative PRICE is still GPT-4.1 Global at $2.00/$8.00 from the 2026-10-14 Azure OpenAI pricing capture; that page was not re-fetched in this pass, so the price and price_basis are unchanged and the new measurement is for a different model than the price anchor. price_basis names the priced model; the measurement names Llama 3.3 70B. The OpenAI-on-Azure path: GPT-4.1's 1,048,576-token context is the biggest published context in this record set. Three deployment types (Global / Data Zone / Regional) are a data-residency choice that shows up directly in the per-1M price - Global is cheapest, Regional pins inference inside one of 27 regions. Artificial Analysis benchmarks Azure on Llama 3.3 70B; no measured row was extracted for this pass so the latency fields stay null. Foundry's model catalog (Meta, DeepSeek, xAI, Mistral) prices per model on the same page and is the 'other models' clause in the basis.

Cross-references

Models (3)