Fine-Tuning Cost by Model (2026): Per-Token Pricing Across 20+ Models

The 30x span between the cheapest fine-tune (Llama 4 8B LoRA on Together at $9 for a typical job) and the most expensive (GPT-5 full SFT at $2,812 for the same job) reflects three different sources of cost difference.

**Model size**: bigger models are more expensive to train because forward/backward passes touch more parameters and weights consume more GPU time. GPT-5 is several hundred billion parameters; Llama 4 8B is 8 billion. The 30-100x size ratio drives 5-15x of the cost difference.

**Method efficiency**: LoRA touches ~0.1-1% of the parameters during training while full fine-tuning touches all of them. The 20-30x compute efficiency gain of LoRA over full fine-tuning is the largest single cost lever in fine-tuning. See our LoRA vs QLoRA vs full fine-tuning cost deep-dive.

**Platform margin and infrastructure**: hosted platforms charge a margin on top of raw GPU cost. Together AI runs at the lowest margin (closest-to-cost per-token rates), Fireworks at moderate margin (premium for the serving stack), OpenAI and Anthropic at the highest margin (frontier-model premium, brand value, and tight integration). Self-hosting on raw GPU hours (Lambda, CoreWeave, RunPod) is the cheapest of all but adds engineering cost.

Training cost is one-time. Inference cost compounds with traffic and often exceeds the training cost within weeks of going live.

**OpenAI fine-tuned models** are priced at approximately 2x base model input and 1.5x base model output. For a GPT-5 mini fine-tune, base model is $0.40/1M input + $1.60/1M output; fine-tune is approximately $0.80/1M input + $2.40/1M output. At 100M tokens/month inference (split 30/70 input/output), monthly inference cost on the fine-tune is approximately $192 versus $124 on the base — a $68/month premium that adds up.

**Anthropic fine-tuned models** on Bedrock charge approximately 1.5x base model price for inference, plus the provisioned throughput hourly cost. Provisioned throughput for Sonnet 4.6 runs $15-25/hour for one model unit, so always-on serving adds $11K-18K/month minimum independent of traffic.

**Google Vertex fine-tuned models** charge approximately 1.0-1.2x base model price for inference plus the endpoint hourly cost. A Gemini 2.5 Flash endpoint runs at the published Vertex AI prediction rate per hour — small but non-zero, typically $0.10-0.50/hour depending on machine type.

**Together / Fireworks fine-tuned models** charge per token at rates close to or matching the base model — no inference markup. Together Llama 4 70B is $0.88/1M in/out for both base and fine-tunes. Fireworks similar. This is one of the largest cost advantages of open-weight platforms versus closed-source frontier vendors for high-traffic production deployments.

Bedrock provisioned throughput and Vertex AI endpoint hours are charged whether or not traffic flows. This deployment floor is the single biggest cost trap in fine-tuning workflows.

**AWS Bedrock provisioned throughput** for Claude Haiku 4.5 fine-tune: approximately $8-12/hour per model unit. For Claude Sonnet 4.6 fine-tune: approximately $15-25/hour per model unit. Always-on for one month: ~$5,800-18,000 depending on model. Yes, that is per month of *idle* serving capacity. For low-traffic production deployments, this floor can dominate total spend by 10-50x over the training cost itself.

**Vertex AI endpoint hours** for Gemini 2.5 Flash fine-tune endpoints: approximately $0.10-0.50/hour depending on machine type (n1-standard, n2, accelerated). Always-on for one month: $73-365. Lower than Bedrock but still non-trivial.

**OpenAI fine-tuned models** have no per-hour serving floor — they auto-deploy and you pay only per inference token. This is the largest cost advantage of OpenAI fine-tuning for low- to medium-traffic production deployments.

**Together / Fireworks serverless inference** has no per-hour serving floor on the standard serverless tier — pay only per token. Dedicated endpoints (for high-traffic or low-latency requirements) reintroduce a per-hour cost.

**Replicate** charges per second of GPU runtime. Cold starts (30-60 seconds) add cost if the model is not warm; always-on warm pools reintroduce hourly floors.

Putting all three components together for a realistic workload: 5,000-example fine-tune (22.5M training tokens), deployed for 6 months, serving 10M inference tokens/month.

**OpenAI GPT-5 mini SFT**: $562 training + (10M × 6 months × $0.80/1M input + $2.40/1M output split 30/70) = $562 + ~$1,440 inference = ~$2,000 total over 6 months. No hourly floor.

**Anthropic Claude Haiku 4.5 SFT (Bedrock)**: $1,012 training + provisioned throughput $10/hour × 24 × 30 × 6 = $43,200 hourly floor + ~$3,150 inference (at Bedrock per-token prices) = ~$47,400 total over 6 months. The hourly floor dominates.

**Google Gemini 2.5 Flash SFT (Vertex)**: $0-180 training (depending on free tier) + endpoint hours $0.25/hour × 24 × 30 × 6 = $1,080 + ~$210 inference = ~$1,470 total over 6 months.

**Together AI Llama 4 70B LoRA**: $27 training + (60M total × $0.88/1M) ~ $53 inference = ~$80 total over 6 months. No hourly floor on serverless.

**The takeaway**: for low-traffic production deployments (10M tokens/month is light traffic), open-weight Llama 4 70B LoRA on Together is dramatically the cheapest total cost. Anthropic via Bedrock is dramatically the most expensive because of the provisioned throughput floor. OpenAI's auto-deployed fine-tunes are in the middle. Run the math at *your* traffic volume — the crossover where Bedrock becomes competitive is around 1B tokens/month.

When estimating fine-tune cost for your workload, the inputs that drive most of the variance are:

**Training tokens** = (number of examples) × (average tokens per example) × (number of epochs). Token count includes both prompt and completion. For chat-format data, tokens are typed including system messages and role separators. Underestimating this is the most common error.

**Method choice** = SFT vs LoRA vs full fine-tuning vs DPO. Multiplier is approximately 1.0 (LoRA baseline), 3-5x (full fine-tuning), 1.5-2x (DPO).

**Projected monthly inference tokens** = (requests per second) × (seconds in a month) × (average tokens per request). Include both input and output tokens.

**Deployment duration** = how long the fine-tune will be in production. Multi-year deployments accumulate inference and hourly costs that can dwarf training cost.

**Hourly serving floor** = does the platform have a per-hour deployment cost? Bedrock yes ($8-25/hour per model unit), Vertex yes ($0.10-0.50/hour), OpenAI no, Together/Fireworks no on serverless.

**Inference markup** = ratio of fine-tuned model inference price to base model inference price. OpenAI ~2x in / 1.5x out; Anthropic ~1.5x; Vertex ~1.0-1.2x; Together/Fireworks ~1.0x.

Five practical levers to reduce total fine-tune cost without sacrificing quality.

**Lever 1: Use LoRA instead of full fine-tuning unless quality forces otherwise.** LoRA is 20-30x cheaper than full fine-tuning and matches quality within 1-3 percentage points on most tasks. Default to LoRA; promote to full only when LoRA quality plateaus below your target.

**Lever 2: Pick the cheapest model that meets your quality bar.** GPT-5 mini fine-tunes are often within 5 percentage points of GPT-5 full fine-tunes at 4x lower cost. Gemini 2.5 Flash fine-tunes are within 8-15 percentage points of Gemini 2.5 Pro at a fraction of the cost. Always run a smaller-model fine-tune as a baseline before committing to the most expensive option.

**Lever 3: Avoid Bedrock provisioned throughput for low-traffic deployments.** If you are serving Claude Sonnet 4.6 fine-tune at less than 100 RPS, the provisioned throughput hourly cost will dominate. Either commit to high traffic, use batch inference (where supported), or reconsider Anthropic via Bedrock as the right platform.

**Lever 4: Use validation loss to stop training early.** Default 3 epochs is reasonable but watch the validation loss curve — if it plateaus or rises after epoch 1, additional epochs are wasted training tokens. Most platforms allow early stopping.

**Lever 5: Iterate on small samples first.** Run a 500-example baseline at $5-15 cost on any platform to validate the workflow and rough quality direction before committing to a 50,000-example $500-2,000 production run.

The cheapest fine-tune cost is the one you do not pay because you found a better lever. Three patterns where teams should not fine-tune:

**Prompt engineering has not been exhausted.** A well-engineered system prompt with 3-5 in-context examples often closes 70-80% of the quality gap that fine-tuning would close. Always exhaust prompt engineering first — it is free, fast, and iterable in seconds rather than hours.

**The real problem is factual recall, not behavior.** Fine-tuning shifts a model's behavior distribution; it does not give the model new facts. For up-to-date knowledge, factual grounding, or document-specific Q&A, RAG (Retrieval-Augmented Generation) is the right tool. See when to fine-tune vs RAG vs prompt engineer.

**You have fewer than 200 examples.** Below 200 examples, fine-tuning rarely produces measurable quality improvement on most tasks. Spend the training budget on either getting more examples or on prompt engineering.