2026 Managed Fine-Tuning Pricing: GPT-5-Class Models on OpenAI, Azure, Vertex AI, and Bedrock

What to Expect from Managed Fine-Tuning in 2026 Managed fine-tuning remains a cornerstone for customizing large language models (LLMs) like GPT-5-class variants to enterprise needs, such as domain-specific RAG pipelines or agentic workflows on platforms like LUMOS. In 2026, the landscape has shifted: OpenAI's direct fine-tuning is largely legacy, with new users directed to cloud-managed services on Azure, Google Vertex AI, and AWS Bedrock. These providers handle infrastructure, compliance, and scaling, but introduce structured pricing around training compute, dataset preparation, and post-tuning inference. Key trends as of May 14, 2026: - Training charges : Billed per token processed or compute hour, with GPT-5-class models (e.g., high-parameter flagships) commanding premium rates due to resource intensity. - Dataset minimums : Vendors enforce thresholds to ensure viable jobs, typically

100–10,000 examples, impacting small-scale pilots. - Inference surcharges : Fine-tuned models often incur 1.2x–2x base rates for hosted inference, critical for ongoing ops costs in RAG/agents. This roundup draws exclusively from official vendor documentation (cited per section), focusing on methodology for reading tiered SKUs rather than static leaderboards. Always verify via provider calculators for your workload. OpenAI GPT-5-Class Fine-Tuning: Legacy Pricing and Wind-Down OpenAI's fine-tuning API, once dominant for models like gpt-4o and early GPT-5 previews, entered wind-down by 2026. Per OpenAI's platform status page as of May 14, 2026 (https://platform.openai.com/docs/guides/fine-tuning), new accounts cannot initiate jobs, though existing fine-tuned deployments continue with support until at least 2027. For legacy GPT-5-class SKUs (e.g., gpt-5.4-pro-2026-03-15): - Training charges

: Input tokens at $30 per 1M, output at $60 per 1M, per OpenAI pricing docs (https://openai.com/api/pricing). Jobs require JSONL datasets with a 10-example minimum. - Dataset minimums : 10 examples (real-world viable at 1,000+ for convergence). - Inference surcharges : Hosted fine-tuned models bill at 1.5x base rates, e.g., $180 input / $540 output per 1M tokens for gpt-5.4-pro equivalents. Enterprises with pre-2026 tunes benefit from continuity but face migration pressures to Azure for scale. For RAG customization, this means evaluating transfer learning to avoid retraining. Azure OpenAI Fine-Tuning: Charges, Minimums, and Inference Azure OpenAI Service provides managed fine-tuning for OpenAI models, including GPT-5-class like gpt-5-turbo-2026-02, integrated with enterprise governance. Pricing follows a pay-as-you-go (PAYG) or provisioned throughput units (PTU) model, detailed in Azure'

s pricing calculator as of May 14, 2026 (https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/). - Training charges : Token-based, mirroring OpenAI but tiered by region/commitment. For gpt-5-turbo, $25–$35 per 1M training tokens (input/output split 1:2 ratio typical); minimum job size 100 examples. - Dataset minimums : 100 high-quality examples required, uploaded via Azure Storage with validation tools. - Inference surcharges : Fine-tuned deployments use PTU for predictable latency, at 1.2x–1.8x base model rates (e.g., $15–$75 per 1M tokens depending on tier). No per-token overage in reserved capacity. Azure excels for hybrid setups, with fine-tuning jobs spinning up in minutes. For agent workflows, factor in data connector costs ( $0.50/GB processed). Google Vertex AI: Training Costs and Dataset Requirements Vertex AI's supervised fine-tuning supports Gemi

ni and PaLM successors (GPT-5-class equivalents like gemini-2.0-pro-2026-01), emphasizing hyperparameter tuning and evaluation. Pricing is compute-centric, per Vertex AI documentation as of May 14, 2026 (https://cloud.google.com/vertex-ai/pricing). - Training charges : Billed per node-hour on custom accelerators (e.g., A3 instances at $3.50–$10/hour per node for 8x setups). A 1B-token GPT-5-class job might total $500–$2,000 based on epochs. - Dataset minimums : 500 examples minimum for tuning jobs, with auto-splitting (80/20 train/eval) and quality checks via Vertex Data Labeling. - Inference surcharges : Tuned endpoints match base model rates (no explicit markup), e.g., $0.50–$5 per 1M chars for gemini-2.0-pro, but add endpoint hosting ( $0.10/hour idle). Vertex suits data-heavy RAG fine-tunes, with built-in drift detection. Use the pricing calculator for multi-node scaling. AWS Bedrock

Customization Pricing: Surcharges Explained Amazon Bedrock's model customization (fine-tuning) covers base models like Anthropic Claude 4 or Titan equivalents to GPT-5-class (e.g., titan-text-g5-2026-04). It's provisioned instance-based, per AWS pricing page as of May 14, 2026 (https://aws.amazon.c