Skip to content

Cost reference

What it costs to fine-tune gpt-oss-20b

The cheapest setup we compared is LoRA on 2 × L40S: about $0.030 per million training tokens, so $0.30 for 10 million and $3.01 for 100 million. Full fine-tuning costs about $0.45 per 10 million tokens on 8 × L40S. As a mixture of experts it computes like a 3.6B model, so it is cheap per token for its size.

By Will Schulz · Checked 15 November 2026 · Prices as of November 2026

GPU-hours and cost by GPU and method

Per epoch over the token count shown; multiply by your number of epochs. GPU-hours are the same however many GPUs the job is split over.

GPU · methodGPUs for memoryTokens/s per GPUGPU-hours per 1M1.0M tokens10M tokens100M tokens
H100 SXM · LoRA$4.39/hr127,4720.01$0.044$0.44$4.44
H100 SXM · QLoRA$4.39/hr121,4280.01$0.057$0.57$5.69
H100 SXM · Full fine-tuning$4.39/hr518,3150.02$0.067$0.67$6.66
B200 · LoRA$7.99/hr162,5000.00$0.036$0.36$3.55
B200 · QLoRA$7.99/hr148,7500.01$0.046$0.46$4.55
B200 · Full fine-tuning$7.99/hr341,6670.01$0.053$0.53$5.33
L40S · LoRA$1.09/hr210,0560.03$0.030$0.30$3.01
L40S · QLoRA$1.09/hr17,8430.04$0.039$0.39$3.86
L40S · Full fine-tuning$1.09/hr86,7040.04$0.045$0.45$4.52
RTX 5090 · LoRA$1.39/hrDoesn't fit5,8190.05———
RTX 5090 · QLoRA$1.39/hr14,5390.06$0.085$0.85$8.51
RTX 5090 · Full fine-tuning$1.39/hrDoesn't fit3,8800.07———

Compared with fine-tuning APIs

OptionPer 1M tokens10M tokens100M tokens
Rented GPU: LoRA on L40S (our estimate)$0.030$0.30$3.01
Together AI, gpt-oss-20B LoRA$0.40$4.00$40.00
OpenAI gpt-4.1-mini fine-tuning (closed model; closed to new users)$5.00$50.00$500

API prices are per training token processed (dataset tokens × epochs), seen on 15 November 2026. Together AI charges a minimum per job ($4 for this model). OpenAI's pricing page says fine-tuning is closed to new users. Rented GPU estimates leave out setup, failed runs and idle time, which an API absorbs.

Your run

Pre-filled for gpt-oss-20b: 10M tokens, 3 epochs, LoRA on an H100.

Estimated cost

$1.33

H100 SXM-hours

0.30

Time on 1 GPU

18 minutes

Tokens per second per GPU

27,472

30M tokens processed. Fits on one H100 SXM at 2,048-token sequences. GPU-hours assume perfect scaling across GPUs.

Assumptions

  • Compute per token: 6 FLOPs per active parameter for full fine-tuning (forward plus backward) and about 4 for LoRA and QLoRA, which skip the frozen weights' gradients. gpt-oss-20b uses 3.6B parameters per token.
  • Utilisation: 40% of each GPU's dense BF16 peak. torchtune measured 3,083 tokens a second for Llama 3.1 8B LoRA on one RTX 4090, about 60% on the same accounting, so well-tuned code can beat these figures.
  • QLoRA runs at 78% of LoRA's speed, the ratio torchtune measured (2,413 vs 3,083 tokens a second), because the 4-bit weights are dequantised on the fly.
  • Memory decides the GPU count, from our gpt-oss-20b VRAM estimates at 2,048-token sequences. Consumer cards are counted as single GPUs only.
  • Prices: on-demand Runpod Secure Cloud rates (page dated 12 November 2026). Community Cloud and spot capacity cost less.

Frequently asked questions

How much does it cost to fine-tune gpt-oss-20b?
About $0.030 per million training tokens at the cheapest setting we compared (LoRA on 2 × L40S, $1.09/hour per GPU), so about $0.30 for 10M tokens and $3.01 for 100M. Budget for several runs: the first rarely wins.
How long does it take to fine-tune gpt-oss-20b on an H100?
LoRA processes about 27,472 tokens a second on one H100 at 40% utilisation, so 10M tokens take about 6 minutes. Full fine-tuning needs 5 H100s for memory and about 1.5× the compute.
Is it cheaper to fine-tune gpt-oss-20b through an API?
Together AI listed gpt-oss-20B LoRA fine-tuning at $0.40 per million tokens (minimum $4 per job) in November 2026, about $4.00 for 10M tokens, against roughly $0.30 of rented GPU time by our estimate. The API includes the setup, retries and idle time a rented GPU doesn't.
Should I use LoRA or full fine-tuning for gpt-oss-20b?
Start with LoRA or QLoRA. Full fine-tuning gpt-oss-20b needs about 339 GB, so 5 × 80 GB GPUs with sharded optimiser state, and about 1.5× the compute per token. LoRA usually gets close for format, style and task adaptation; move to full fine-tuning only when evaluation shows LoRA has plateaued.

Sources

  1. [1]Hugging Face. openai/gpt-oss-20b config.json
  2. [2]OpenAI. openai/gpt-oss-20b model card
  3. [3]OpenAI. gpt-oss-120b & gpt-oss-20b Model Card (arXiv:2508.10925)
  4. [4]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  5. [5]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  6. [6]Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
  7. [7]PyTorch. torchtune: memory and throughput benchmarks (Llama 3.1 8B on RTX 4090)
  8. [8]Together AI. Fine-tuning pricing (seen 15 November 2026)
  9. [9]OpenAI. API pricing: fine-tuning (seen 15 November 2026)
  10. [10]Runpod. GPU pricing (updated 12 November 2026)

Try the cheaper levers first.

Prompt tuning and model search in Kojen cost API calls, not GPU-hours, and often close the gap before any fine-tuning.

Start training free