How much does OpenAI fine-tuning cost in 2026?

OpenAI supervised fine-tuning (SFT) on GPT-4.1 costs $25 per million training tokens. Reinforcement fine-tuning (RFT) on o4-mini costs $100 per hour of training time, with a $5,000 per-job cap. RFT is billed by the hour regardless of dataset size; a small dataset (10K examples) and a large dataset (1M examples) can cost the same hourly rate. Hosting for fine-tuned models is billed separately at roughly $1-$2 per hour per deployed model regardless of traffic (kensink.com, September 2026).

How much does Vertex AI fine-tuning cost?

Vertex AI charges per training token at a rate that varies by base model. The tuned model then bills at 1.5x the base inference rate for the lifetime of the deployment. For example, a fine-tuned Gemini 2.5 Flash at 1.5x the standard $0.30/$2.50 rate costs $0.45/$3.75 per million tokens in production. Vertex also offers adapter-based tuning (LoRA) which trains a small adapter while keeping base weights frozen — the cheapest fine-tuning path on the platform (kensink.com, September 2026).

How much does Bedrock fine-tuning cost?

AWS Bedrock supports managed fine-tuning for select models plus Custom Model Import for any Hugging Face model. Custom Model Import costs $0.0785 per minute per Custom Model Unit (CMU) for hosting. Fine-tuned models and imported models both require Provisioned Throughput to serve, billed hourly. Bedrock's managed fine-tuning catalog is the smallest of the three hyperscalers — only Claude 3 Haiku has first-party fine-tuning support; other models require Custom Model Import or SageMaker (kensink.com, September 2026).

How much does Azure OpenAI fine-tuning cost?

Azure OpenAI fine-tuning mirrors OpenAI's pricing: $25 per million training tokens for SFT and $100 per hour for RFT on o-series models. Hosting is billed separately at roughly $1-$2 per hour per deployed model regardless of traffic. Fine-tuning spend is the line item teams most often lose track of on Azure — Azure ML compute, monitoring, and storage add 15-40% on top of headline training cost (technologymatch.com, June 2026).

When is fine-tuning worth the cost?

Fine-tuning is worth the cost when behaviour, format, or latency still falls short after prompt engineering and RAG. The standard sequence is: start with prompt engineering, add RAG when the model needs your data, fine-tune only when accuracy, latency, or output format still fails. Fine-tuning is most justified for: domain-specific terminology, structured output formats (JSON schemas), low-latency workloads where RAG lookup overhead is too high, and tasks where the base model consistently fails at >5% of attempts. It is NOT justified for: tasks solvable by better prompting, knowledge that fits in context (use RAG instead), or one-off classification tasks (kensink.com, September 2026).

Can Claude be fine-tuned in 2026?

Claude 3 Haiku supports fine-tuning via AWS Bedrock, which Anthropic made generally available in November 2024. As of September 2026, no other Claude model supports first-party fine-tuning on Anthropic's direct API. Bedrock fine-tuning requires Provisioned Throughput for serving. Newer Claude models (Opus 5, Sonnet 5, Fable 5.1) do not support fine-tuning through any platform. For Claude customisation beyond prompt caching, the options are: better prompting, RAG with tool use, or building a fine-tuned open-weight model and routing through Claude for evaluation (kensink.com, September 2026).

What are hidden costs in LLM fine-tuning?

Hidden costs across all platforms include: data preparation labour (typically $5K-$50K for a quality 50K-example dataset), evaluation infrastructure (held-out test sets, human review, automated metrics at $1K-$10K/month for hosted services), hosting during development (sandboxed fine-tuned deployments cost $1-$2/hour even at zero traffic), retraining as base models update (when GPT-6 Astra succeeds GPT-5, fine-tuned models often need to be retrained), and ongoing maintenance (drift detection, dataset refresh). The headline training cost is typically 30-50% of total TCO in the first year (kensink.com, September 2026).

LLM Fine-Tuning Cost Comparison 2026: Full Guide

Quick Answer

LLM fine-tuning cost 2026: OpenAI GPT-4.1 SFT $25 per million training tokens, RFT on o-series $100/hour with a $5,000 per-job cap; Vertex AI per-token training plus 1.5x base inference for the lifetime of the tuned deployment; Bedrock $0.0785/min/CMU for Custom Model Import; Together AI $0.48-$2.90 per million training tokens on open-weight models. Hidden costs add 15-40% on top of headline training rates.

Last verified: Sep 16, 2026.

At a glance

OpenAI SFT: $25 per million training tokens (GPT-4.1)
OpenAI RFT: $100/hour training, $5,000 per-job cap (o-series)
Azure OpenAI: mirrors OpenAI; $25/MTok SFT, $100/hr RFT, hosting $1-$2/hr
Vertex AI: per-token training + 1.5x base inference for tuned model
AWS Bedrock: $0.0785/min/CMU Custom Model Import; only Claude 3 Haiku native fine-tune
Together AI: $0.48-$2.90 per million training tokens on open-weight models
Anthropic: no direct fine-tuning — only via Bedrock, Claude 3 Haiku only
Hidden costs: data prep $5K-$50K, eval $1K-$10K/mo, retraining as base models update

OpenAI fine-tuning: two flavours, two price models

OpenAI offers two fine-tuning flavours as of September 2026: supervised fine-tuning (SFT) at $25 per million training tokens, and reinforcement fine-tuning (RFT) at $100 per hour with a $5,000 per-job cap.

SFT is the standard approach: you provide a dataset of input-output pairs, OpenAI trains a new model on your data, and you get a model ID that can be served via the standard Chat Completions or Responses API. Pricing is per training token, so a 10M-token dataset costs $250 to train. Hosting is separate: deployed fine-tuned models cost roughly $1-$2 per hour per model regardless of traffic. The bill climbs quickly when development requires multiple training runs (kensink.com, September 2026).

RFT is the more expensive option designed for o-series reasoning models. The trainer runs a verifier (your custom grading code) on each output and reinforces the model's behaviour toward higher-scoring completions. The $100/hour rate is paid regardless of dataset size — a small 10K-example job and a large 1M-example job both run for hours of compute. The $5,000 per-job cap prevents runaway training jobs from blowing budgets. RFT is the fastest path to a measurable reasoning lift if your data fits the o-series (kensink.com, September 2026).

Vertex AI: per-token training plus 1.5x inference

Vertex AI's fine-tuning model charges per training token at a rate that varies by base model, then bills the tuned model at 1.5x the base inference rate for the lifetime of the deployment.

For Gemini 2.5 Flash at $0.30/$2.50 base inference, a tuned deployment costs $0.45/$3.75 per million tokens in production. For Gemini 2.5 Pro at $1.25/$10, a tuned deployment costs $1.875/$15 per million tokens. The 1.5x multiplier is permanent — it does not amortise over time, so a tuned deployment is always 50% more expensive than the base model in production (kensink.com, September 2026).

Vertex offers adapter-based tuning (LoRA), full fine-tuning, and a distillation service that trains a smaller student model from a frontier teacher. Adapter-based tuning is the cheapest: the adapter is small (typically 1-10% of base model size), trains in minutes, and can be served alongside the base model. Full fine-tuning retrains the entire model — useful for behaviour change but expensive and slow. Vertex's catalog of fine-tunable models includes Gemini 2.5 Pro, Flash, and Flash-Lite (kensink.com, September 2026).

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.

Photo: Jhedengren, CC BY, via Wikimedia Commons (https://upload.wikimedia.org/wikipedia/commons/b/b5/Artificial_Neural_Network_Example.png?utm_source=commons.wikimedia.org&utm_campaign=imageinfo&utm_content=original)

#llm fine tuning cost #openai fine tuning #vertex ai fine tuning #bedrock fine tuning #together ai fine tuning #fine tuning pricing

Written by

Fazlur RahmanAI, Entrepreneurship, EdTech

Fazlur Rahman is the founder of Tutorsbot, building AI-powered tools for learning and career growth. He writes about applying AI in real products and the practi… Read more

Fazlur Rahman is the founder of Tutorsbot, building AI-powered tools for learning and career growth. He writes about applying AI in real products and the practical side of building an ed-tech startup.