Free Tools · LLM Cost + Fine-Tuning Last updated: June 2026
Enter your real volume. Watch the cheapest path sort itself out.
The defaults below assume a testing-heavy month across a few small projects. Replace them with your actual numbers — the ranking re-orders instantly. Costs are per month.
Your usage
Find these in the Anthropic Console usage dashboard, or estimate.
million
million
Reused / cacheable input60%
Batchable (non-interactive) calls40%
If you self-host instead
hours
RTX 4090 $0.44/hr
A100 80GB $1.30/hr
H100 $2.50/hr
If you buy hardware
$ total
months
$ / mo
Option
Input $/M
Output $/M
Claude Opus 4.8
Claude
5.00
25.00
Claude Sonnet 4.6
Claude
3.00
15.00
Claude Haiku 4.5
Claude
1.00
5.00
DeepSeek V3.2
Open
0.14
0.28
Llama 3.3 70B (Groq)
Open
0.59
0.79
Llama 3.1 8B (Groq)
Open
0.05
0.08
RTX 4090 rental
Infra
~$0.44 / hr
A100 80GB rental
Infra
~$1.30 / hr
H100 rental
Infra
~$2.50 / hr
Owned rig
Infra
rig ÷ months + power
Per-million-token rates. Claude rates from Anthropic's June 2026 pricing. Open-model rates are representative serverless host prices (Groq / DeepSeek / Together) and vary by host. GPU rates are mid-range on-demand figures (RunPod / Vast / Lambda).
Prompt caching is modeled as the cached share of input tokens costing 10% of the base input rate. Real caching adds a small write premium and applies only within the cache window — treat it as a close estimate, applied to Claude here.
Batch (50% off) is applied to the batchable share of Claude spend. Open-model serverless hosts mostly don't offer a batch tier, so it isn't applied there; some (e.g. DeepSeek) do offer their own caching.
Open-model quality gap: on chat, RAG, classification, and summarization the gap to frontier is small. On hard coding and multi-step agentic work it's real — so the cheapest row isn't always the right row for that task.
GPU rental is flat: it costs the same whether you send one token or a billion — you pay for the hours the box is up. The discipline is spinning it down when idle. Best when fine-tuning or hosting an always-on endpoint.
Owned hardware only wins at sustained high utilization. A consumer card (4090-class) runs ~7B models well or ~13B quantized; matching frontier quality locally needs far more VRAM.
Not financial advice — these are infrastructure cost estimates to help you compare, not a recommendation on spend.
Pick a model and a method. See the VRAM you'll need — and what runs it.
Fine-tuning memory is dominated by optimizer state, not the model itself. Moving from full fine-tuning to LoRA or QLoRA can cut the requirement 10–20×. Set your config; the GPU fit re-sorts instantly.
Your fine-tune
Open-weight models only — Claude can't be fine-tuned locally (see Options guide).
1B
3B
7–8B
13B
34B
70B
405B
QLoRA 4-bit base
LoRA 16-bit base
Full FT 16-bit
8-bit Adam
Adam (fp32)
Batch size1
Sequence length2048
On
Off
Estimated VRAM
— GB
Fits on which card?
RTX 4090 24 GB
RTX 5090 32 GB
A100 / H100 80 GB
H200 141 GB
Method
VRAM
Trainable
VRAM for the selected model at the current batch / sequence / optimizer / checkpointing settings. "Trainable" is the share of parameters whose gradients and optimizer state are stored — the real driver of the gap between methods.
Full fine-tuning — updates every weight. Highest quality ceiling, but you store weights + gradients + optimizer state (~16 bytes/param with Adam, ~10 with 8-bit Adam). A 7–8B needs an 80GB card or multi-GPU; 70B needs a node.
LoRA — freezes the base (16-bit) and trains small low-rank adapters (~0.1–1% of params). Optimizer state shrinks to almost nothing, so you mostly pay for the frozen weights + activations.
QLoRA — LoRA with the frozen base quantized to 4-bit (NF4), roughly quartering base-weight memory. This is what puts 70B fine-tuning on a single 80GB GPU and 7–8B on a 12GB consumer card.
Preference tuning (DPO / ORPO / RLHF) — runs on top of LoRA/QLoRA. DPO/ORPO add a reference pass (~1.5–2× the activation memory of plain SFT); full RLHF (PPO) holds policy + reference + reward models and needs substantially more.
Activations scale with batch × sequence length. Gradient checkpointing recomputes them in the backward pass — a large VRAM saving for ~20–30% more compute time. Leave it on unless you're memory-rich.
Claude can't be fine-tuned locally — it's closed-weights. Anthropic's only managed fine-tuning path has been Claude Haiku on Amazon Bedrock (AWS hosts it; you never touch weights). For most "make it know my domain" goals, prompt caching + RAG beats fine-tuning anyway.
Estimates, not guarantees — real usage shifts with framework (DeepSpeed / FSDP), attention kernel (FlashAttention), and fragmentation. Treat the number as a planning floor and add ~15% headroom.