Bedrock: managed fine-tuning plus Custom Model Import
Bedrock's managed fine-tuning catalog is the smallest of the three hyperscalers, but Custom Model Import opens the door to any Hugging Face model — billed at $0.0785 per minute per Custom Model Unit (CMU).
Only Claude 3 Haiku has first-party fine-tuning support on Bedrock as of September 2026 (Anthropic Claude 3 Haiku SFT, GA November 2024). Every other model on Bedrock — including Claude Sonnet 4.6, Llama, Mistral, and Cohere — must be fine-tuned through Custom Model Import or external SageMaker. Custom Model Import serves your own weights behind the same unified API but requires Provisioned Throughput to serve (kensink.com, September 2026).
The CMU pricing model is not directly comparable to per-token training rates on OpenAI or Vertex. CMUs measure throughput capacity, not training compute, so the bill depends on how many tokens per second the model serves rather than how many training tokens were used. For a production deployment with consistent traffic, CMUs can be more cost-efficient than Vertex's 1.5x multiplier because the CMU is sized to actual demand (kensink.com, September 2026).
Azure AI Foundry: mirrors OpenAI with bigger hidden costs
Azure OpenAI fine-tuning mirrors OpenAI's headline pricing, but Azure's hidden costs add 15-40% on top.
SFT is $25 per million training tokens, RFT is $100 per hour, hosting is $1-$2 per hour per deployed model — same as OpenAI direct. The hidden costs come from Azure's broader stack: Azure Monitor logs (automatic and included), Azure AI Search for RAG ($0.75 per 1K documents indexed plus query costs), Azure ML compute for fine-tuning orchestration (GPU/CPU hourly), and data transfer fees for cross-region deployments (technologymatch.com, June 2026).
For most production Azure deployments, fine-tuning spend is the line item teams most often lose track of. Hosting during development alone — sandboxed fine-tuned deployments at $1-$2/hour even at zero traffic — accumulates fast when there are multiple concurrent experiments. The rule of thumb is to budget 1.5x the headline training rate for total Azure fine-tuning TCO in the first year (technologymatch.com, June 2026).
Together AI: cheapest fine-tuning for open-weight models
Together AI charges per million training tokens for supervised, LoRA, full SFT, and DPO fine-tuning on open-weight models up to 100B parameters — the cheapest fine-tuning path for teams comfortable with self-hosted deployment.
Pricing is tiered by model size: $0.48 input / $1.20 output per million tokens for 16B models, $1.50/$3.75 for 17-69B models, $2.90/$7.25 for 70-100B models. Together offers serverless multi-LoRA endpoints and dedicated endpoints, with the multi-LoRA path allowing multiple adapters on a single base model deployment — useful for A/B testing variants (kensink.com, September 2026).
For teams that need fine-tuning on Llama, Mistral, Qwen, or DeepSeek open-weight models, Together is the cheapest path at scale. The trade-off is operational: open-weight models require more evaluation and iteration because the base capability is lower than GPT-4.1 or Claude Opus. Together works best when the application requires data isolation (VPC deployment), model ownership (the weights are yours), or specific compliance regimes (kensink.com, September 2026).
What enterprise buyers should do next
Three actions for organisations evaluating LLM fine-tuning pricing in 2026.
- Try prompt engineering and RAG before fine-tuning. The 80/20 of fine-tuning decisions: prompt engineering covers most use cases, RAG covers most knowledge cases, fine-tuning is the last 5-15%. Most teams that jump straight to fine-tuning spend 3-10x more than necessary and get worse results.
- Model the lifetime TCO, not the training bill. Vertex's 1.5x permanent multiplier, OpenAI's $1-$2/hour hosting, and Bedrock's CMU costs all add up over the life of the deployment. A $250 training job on Vertex with 100M production tokens/day costs $45,000 over a year at 1.5x Gemini Flash rates.
- Plan for retraining when base models update. When GPT-6 Astra ships, fine-tuned GPT-5 deployments often need to be retrained on GPT-6 Astra. The retraining cost is the same as the initial training — budget for at least one retraining cycle per year for production fine-tuned deployments.
What to watch next
Three near-term datapoints. First, the introduction of fine-tuning support for Claude Sonnet 5 and Opus 5 on Anthropic direct or AWS Bedrock — currently only Claude 3 Haiku supports fine-tuning. Second, Vertex AI's launch of free-tier fine-tuning for new Google Cloud accounts, similar to the existing $300/90-day free inference credits. Third, the cost trajectory of reinforcement fine-tuning (RFT) — the $100/hour rate has held since launch; if it drops to $50/hour with the next OpenAI pricing refresh, RFT becomes viable for many more use cases.