How Much Does It Cost to Fine-Tune an LLM in 2026? LoRA vs Full, GPU vs API

Short answer: fine-tuning an open 8B model on a hosted API costs about $10 to $20 for a 10M-token dataset trained for 3 epochs. The same job on a 70B model costs $60.90 to $201.60, and tuning a Gemini model runs $90 to $750. Those are our calculations from list prices on Together AI, Fireworks and Google Cloud, read on 30 October 2026.

The compute is the cheap part now. What decides your real bill is which provider you pick (the same 70B run varies more than threefold), whether you need full fine-tuning at all, and what it costs to serve the model afterwards, which can dwarf the training run within a week.

Fine-tuning costs at a glance

Scenario: a 10M-token training set, 3 epochs, so 30M training tokens billed. Supervised fine-tuning (SFT), no validation set.

All totals are our arithmetic: per-1M-token rate × 30. The mid row uses Qwen3.5 27B on Together, the 16.1B to 80B bracket on Fireworks and Qwen 3 32B on Google Cloud. Rates from Together AI pricing, Fireworks pricing and Google Cloud Agent Platform pricing, read on 30 October 2026.

Two patterns matter for your choice of vendor. Together prices full fine-tuning only about 10% above LoRA, so the method is not where the money goes there. Fireworks doubles the rate for full-parameter training, and its 16.1B to 80B bracket charges a 27B model the same as a 70B one, which punishes mid-sized models.

Dupple data: on the table above, the most expensive hosted 70B run costs 3.3 times the cheapest ($201.60 on Google Cloud vs $60.90 on Together AI), our arithmetic from list prices on 30 October 2026. The base model, data and epochs are identical, so the gap is pure pricing policy.

How is fine-tuning priced?

Hosted fine-tuning bills per million training tokens, and every provider in this guide counts them the same way: dataset tokens × epochs, plus validation tokens if you evaluate during training. A 10M-token dataset trained for 3 epochs is 30M billed tokens.

That makes your dataset the budget. If you are new to the unit, a token is roughly a word fragment; what a token is in AI explains how text turns into billable tokens. Three habits move the bill most:

• Epochs multiply everything. Going from 3 epochs to 5 adds two-thirds to the cost of the run.
• Validation is billed. Together counts validation set size × number of evaluations on top of the training tokens.
• Padding can inflate the count. Together's docs say that with packing turned off, training tokens are computed as dataset length × maximum sequence length, so short examples get billed as if they were full length.

GPU rental is the other route. You pay per GPU-hour, and the cost depends on how long your run takes on your hardware, which no price list can tell you in advance.

LoRA vs full fine-tuning: what is the price gap?

LoRA trains a small set of adapter weights on top of a frozen base model, while full fine-tuning updates every weight. On hosted APIs that difference shows up as a rate multiplier, and the multiplier varies a lot by vendor.

Ratios are our arithmetic. Together lists full fine-tuning for a shorter set of models than LoRA: it includes Llama 3.1 8B, Llama 3.3 70B and several Qwen and Gemma models, but not the largest ones such as DeepSeek or Kimi.

On self-managed GPUs the gap is much wider than any rate card suggests, because full fine-tuning needs far more GPU memory. The weights of a 70B model alone take about 140 GB in 16-bit precision (70 billion × 2 bytes, our arithmetic), more than one 80 GB H100 holds, before gradients and optimizer state are added. LoRA on a smaller model can fit on one card; a full 70B run means a multi-GPU node.

Preference tuning costs more again. Together's DPO rate for Llama 3.3 70B is $5.08 per 1M tokens on LoRA, so the 30M-token scenario becomes $152.40 instead of the SFT figure.

What does it cost to fine-tune Gemini or GPT models?

Closed models are where 2026 changed the most. You can still tune Gemini, but only through Google Cloud, and OpenAI is closing self-serve fine-tuning in stages.

Google Cloud prices Gemini tuning per 1,000 training tokens, which makes it easy to underestimate. Here is the same 30M-token scenario:

Google lists these as $0.003, $0.005, $0.01 and $0.025 per 1,000 tokens; the per-million figures and totals are our arithmetic. Serving costs more too: for Gemini 3 and later, a tuned endpoint is billed at 1.5 times the base model's prediction price in non-global regions, and at the base price in the global region.

The consumer-facing Gemini API is not an option. Google's Gemini API tuning page says that since Gemini 1.5 Flash-001 was deprecated in May 2025, no model in the Gemini API or AI Studio supports fine-tuning, and it has no immediate plans to bring it back.

OpenAI's deprecations page sets out the wind-down of its self-serve fine-tuning:

• 7 May 2026: organizations that had never run fine-tuning can no longer create jobs.
• 2 July 2026: job creation closed for organizations with no inference on a fine-tuned model in the previous 60 days.
• 6 January 2027: active existing customers can no longer create new fine-tuning jobs. Existing fine-tuned models keep serving until their base model is deprecated.

So if you are starting a fine-tuning project today, OpenAI is off the table. Mistral is a weak bet as well: its fine-tuning docs now sit under deprecated features and say the feature is no longer actively supported.

Is renting GPUs cheaper than a fine-tuning API?

Sometimes, but the honest comparison is "how many GPU-hours does the API price buy?", since run time depends on your model, sequence length and code. Here are the H100 rates on the vendors' own pages:

8-GPU totals are our arithmetic except Replicate's, which it prints. Rates from Together AI pricing, Lambda pricing, fal pricing, Replicate pricing and Fireworks pricing on 30 October 2026. fal also advertises H100s "as low as" $2.49 an hour without listing the terms; Lambda's prices exclude sales tax or VAT.

Now set that against the hosted runs. The cheapest hosted 70B LoRA run in our scenario buys about 1.9 hours of an 8x H100 node at Lambda's rate; the Google Cloud run buys about 6.3 hours (our arithmetic). For the 8B model, the hosted Together run equals about 2.4 hours of a single Lambda H100, or about 5.1 hours of a single 40 GB A100 at $1.99 an hour.

The practical rule follows from that. Small and mid-sized LoRA jobs are cheaper on a hosted API once you count setup, failed attempts and the hours you spend babysitting a box. Renting wins when your dataset is large, when you need a training method the APIs do not offer, or when you run many experiments and already have a working training script; at that point cloud cost optimization tools help keep idle GPUs off the bill. Preemptible GPUs at half price suit checkpointed training, but they can be taken back mid-run.

Together also sells post-training compute by the hour: B200 time is on a promotion at $9.00 instead of $11.00 per GPU-hour until 14 December 2026, so budget at the higher rate for later runs.

How much does a bigger dataset cost?

Costs scale linearly with tokens, so a tenfold dataset is a tenfold bill. At 100M tokens × 3 epochs (300M billed tokens), Llama 3.3 70B LoRA costs $609 on Together, $900 on Fireworks and $2,016 on Google Cloud (our arithmetic). That Together figure equals about 19 hours of an 8x H100 node at $3.99 per GPU-hour, which is where renting starts to look tempting if your runs are faster than that.

At the other end, minimums kick in. A 1M-token dataset trained for 3 epochs on Llama 3.1 8B works out to $1.02 at Together's LoRA rate, but each Together job carries a $4 minimum charge, so you pay $4. The minimum rises with model size: $6 for gpt-oss-120B, $22 for Qwen3.5 397B A17B and $60 for the Kimi and GLM-5 models.

The costs nobody puts in the headline

• Serving the model. Together's docs say a fine-tuned model is served on a dedicated endpoint billed per minute; one H100 costs $5.49 an hour, or $3,952.80 for a 30-day month left on (our arithmetic). Fireworks says it serves fine-tuned models for the same price as base models.
• Idle time. Replicate bills private models for setup and idle time as well as active time, except "fast booting fine-tunes", which bill only while processing requests.
• Reasoning traces. Fireworks warns that tuning with reasoning content unrolls multi-turn conversations, so the billed token count grows; its estimate multiplies by half the average number of turns.
• Failed and cancelled runs. Together refunds jobs that fail, including completed steps, and bills cancelled or early-stopped jobs only for steps completed.
• Your time. Data cleaning, evaluation and iteration are not on any price list, and most projects run several training jobs before one is good enough to ship. Tracking those experiments is what LLM observability tools are for.

Toolradar data: Toolradar tracks 11 AI fine-tuning tools as of 30 October 2026, and 45% of them are free, freemium or open source (AI fine-tuning tools on Toolradar).

Should you fine-tune at all?

Often not first. Fine-tuning teaches a model a format, tone or narrow task; it is a poor way to add knowledge that changes. If your problem is "the model does not know our documents", retrieval is cheaper and easier to update: compare the best RAG tools before you pay for a training run.

Fine-tuning earns its cost when a small tuned model can replace a large general one on a high-volume task, because inference savings repeat every month while the training run is a one-off. If you go that route, an open model you can also run yourself keeps your options open: how to run an LLM locally and the best free local LLM tools cover serving it without per-token fees. Teams that call several providers can route traffic through one of the best LLM gateways, and if the tuned model will drive an agent, how to build an AI agent and the best AI agent frameworks are the next stops.

How we worked out these costs

We read the fine-tuning and GPU rates on Together AI pricing and its fine-tuning pricing docs, Fireworks pricing, Google Cloud Agent Platform pricing, Lambda pricing, fal pricing and Replicate pricing, plus OpenAI's, Google's and Mistral's own pages on fine-tuning availability, on 30 October 2026. The 30M-token scenario, the GPU-hour equivalents and every total are our arithmetic from list prices, before taxes and discounts. We did not run training jobs, so we give no run times.

FAQ

Bottom line

Budget the training run in tens to low hundreds of dollars, then spend your real attention on serving costs and on whether retrieval would solve the problem without training at all. For most teams starting now, a hosted LoRA job on an open model is the cheapest first experiment; how to reduce LLM API costs covers the levers once the model is in production.

Cite this: Dupple, "How Much Does It Cost to Fine-Tune an LLM in 2026? LoRA vs Full, GPU vs API", November 2026.

Pricing out the AI stack for your team? Start the Dupple X trial.