Inference cloud
How Together AI compares
Serverless open-model prices and per-GPU-hour rates for dedicated inference and GPU clusters.
Pricing page · Site · Open in compare
The serverless table was printed under a 'Batch API price' label as well as 'Price per 1M tokens'. The toggle state did not survive extraction. Token rows are the numbers in that table, and the tier is marked ambiguous in the row note.
Regions and compliance
Regions: not published / unverified
Compliance: not published / unverified
Token offerings
DeepSeek V4.1 Flash
serverless_api · deepseek-v4.1-flash · open-weight not published / unverified
- Input / 1M
- 0.3together.aichecked 2026-10-20“DeepSeek V4.1 Flash | $0.30 $0.006 (cached) | $1.20”
- Cached input / 1M
- 0.006together.aichecked 2026-10-20“DeepSeek V4.1 Flash | $0.006 (cached)”
- Output / 1M
- 1.2together.aichecked 2026-10-20“DeepSeek V4.1 Flash | $1.20”
Context not published · tokens/s not published · TTFT not published
Taken from the serverless table. A Batch API price label sat above that table, so the tier is ambiguous.
gpt-oss-120B
serverless_api · gpt-oss-120b · open-weight true
- Input / 1M
- 0.15together.aichecked 2026-10-20“gpt-oss-120B | $0.15 | $0.60”
- Cached input / 1M
- not published / unverifiedtogether.aichecked 2026-10-20“Cached input price not published.”
- Output / 1M
- 0.6together.aichecked 2026-10-20“gpt-oss-120B | $0.60”
Context not published · tokens/s not published · TTFT not published
No cached price was printed in this row. Tier label on the table is ambiguous. Open-weight flag is cited from GroqDocs, which describes the model.
Llama 3.3 70B
serverless_api · llama-3.3-70b · open-weight not published / unverified
- Input / 1M
- 1.04together.aichecked 2026-10-20“Llama 3.3 70B | $1.04 | $1.04”
- Cached input / 1M
- not published / unverifiedtogether.aichecked 2026-10-20“Cached input price not published.”
- Output / 1M
- 1.04together.aichecked 2026-10-20“Llama 3.3 70B | $1.04”
Context not published · tokens/s not published · TTFT not published
Tier label on the table is ambiguous.
Compute SKUs
HGX H100 cluster 1× H100 · gpu hour · in the run-cost model
Reserved figure is the 91–180 day column, not the shortest commit.
HGX H100 dedicated inference 1× H100 · gpu hour · in the run-cost model
Docs say a ready replica can scale to zero. This row still uses the provisioned method because the list price is per GPU-hour while the replica runs.