Best LLM Inference Providers

ComparisonChecked Nov 8, 2026

Together AI vs Fireworks AI

Both sell serverless per-token access to open-weight models with a path to dedicated GPUs and custom models. They differ in published data-handling statements, certifications, and how custom models are deployed.

What differs

  • Deployment: Together offers serverless, provisioned throughput, dedicated model inference per GPU-hour, and dedicated containers for custom models. Fireworks offers serverless modes (Standard, Priority, Fast, Batch, Spot) and on-demand GPU deployments; custom base models and uploaded LoRAs run only on on-demand deployments.
  • API: both document OpenAI client compatibility; Together also supports the OpenAI Responses and Anthropic Messages APIs.
  • Data handling: Fireworks states zero data retention by default for open models, with Response API data kept up to 30 days when store=True. Together’s privacy policy says no training without opt-in; inference retention options were not stated on the pages we checked.
  • Published security: Together lists SOC 2 Type 2 and ISO 27001:2022. Fireworks lists SOC 2 Type 2, HIPAA with a BAA, GDPR, ISO 27001:2022, ISO 27701, and ISO/IEC 42001:2023.
  • Pricing: both publish per-token and GPU-hour pricing on their pricing pages.

Who each tends to fit

Teams that need a HIPAA BAA or written zero-retention defaults may look first at Fireworks’ published terms. Teams that want dedicated containers for custom serving code may look at Together. Test both on your own prompts.

Last checked: . No winner declared.

Side by side

FactTogether AIFireworks AI
StatusActive, privately held.Active, privately held.
OwnershipPrivately heldPrivately held
Buying modelServerless per-token API, Dedicated or provisioned capacity, Bring your own modelServerless per-token API, Dedicated or provisioned capacity, Bring your own model
CategoriesServerless per-token API; Dedicated endpoints; Bring-your-own-model containersServerless per-token API; Dedicated GPU deployments (on-demand); Bring-your-own-model uploads
PricingPublished per-token pricing on https://www.together.ai/pricing (serverless price per 1M tokens by model); dedicated inference billed per GPU-hour (for example, NVIDIA HGX H100 listed at $5.49 per GPU-hour on demand).Published per-token and GPU-hour pricing on https://fireworks.ai/pricing (for example, H100 80 GB listed at $8.00 per hour for on-demand deployments). Serverless usage draws down prepaid credits.
CertificationsSOC 2 Type 2; ISO 27001:2022SOC 2 Type II; HIPAA (with a BAA); GDPR; ISO 27001:2022; ISO 27701; ISO/IEC 42001:2023
LimitationsHIPAA not listed in the Trust Center; ask for BAA terms if needed; Data retention options for inference are not stated on the pages checked; confirm in the DPA; Region choice is not documented on the pages checkedResponse API stores conversation data by default (store=True) for up to 30 days unless you set store=False; Custom base models and uploaded LoRA adapters run only on on-demand deployments, not serverless; GPU availability varies by region
How to buytogether.aifireworks.ai
Sources[1] [2] [3] [4] [5] Checked 2026-11-08[1] [2] [3] [4] [5] Checked 2026-11-08

Facts come from each vendor’s profile sources. Run a paid pilot before choosing.