CalcSays
LLAMA 3.1 70B · FINE-TUNING

Llama 3.1 70B Fine-Tuning Break-Even

Thinking about fine-tuning Llama 3.1 70B? Training costs $2.9/M tokens once — this page computes exactly how many calls it takes to beat few-shot prompting's recurring cost.

For engineers deciding between few-shot prompting and fine-tuning — computes the exact call-volume break-even, not a 'fine-tuning is cheap/expensive' guess.

Model prices from OpenRouter · updated 2026-07-13

01 Training vs few-shot

Model

02 One-time vs recurring waste

Fine-tuning wins by $1,903 (99%) — break-even was 3 days ago.

Training (one-time)
$17.40
6,000,000 tokens
Few-shot (first year)
$1,920
$0.000320/call × 12mo

Break-even: 54,375 calls (3 days at this volume)

Tweet
Why isn't few-shot free? Every call re-sends the example tokens as input — a small per-call cost that recurs forever and eventually exceeds a one-time training investment. Fine-tuned inference bills at the same rate as the base model, so the whole decision comes down to this break-even, not inference pricing.
📄 Get the full agent cost audit
A multi-page report on your exact setup: where cost spirals, how much context-trimming / state-resets / model-routing each saves, and which model+framework combo is cheapest at your real failure rate. Generated from your numbers.

Inference prices from OpenRouter, snapshot 2026-07-13, synced daily. Training prices hand-verified (checked 2026-07-13), re-audited quarterly: Together AI fine-tuning pricing. Break-even = training cost ÷ per-call few-shot cost. All math runs in your browser.

How the math works

OpenAI wound down its self-serve fine-tuning API in May 2026 — open-weight fine-tuning via neutral hosts like Together AI is now the primary path most teams take. Llama 3.1 70B fine-tunes at $2.9/M training tokens (Together AI, SFT LoRA, large tier).

Fine-tuning looks like "pay once" against few-shot prompting's "pay nothing" — but few-shot isn't free. Every call re-sends the 800 example tokens as input, forever: $0.000320 per call, $160/month at this page's 500,000 calls. Multiplied over a year that's $1,920 — a real, recurring cost most teams never add up.

Same baseline, one identity: training 6,000,000 tokens (2,000,000 × 3 epochs) costs $17.40 once. Few-shot's first-year cost is $1,920. Fine-tuning wins by $1,903 (99%) at this volume.

The real number to know isn't "is fine-tuning worth it" — it's the break-even: 54,375 total calls (≈3 days at this page's volume). Below it, few-shot is cheaper; every call past it is pure savings, compounding for as long as the model stays in production.

Fine-tuned inference reuses the base model's serverless rate (Together publishes no separate fine-tuned-inference price), so this page's comparison is training cost vs. recurring example-token waste — nothing else changes. Training price hand-verified (checked 2026-07-13), re-audited quarterly; inference prices sync daily from OpenRouter.

Frequently asked questions

Should I fine-tune Llama 3.1 70B?

At this page's defaults — 500,000 calls/month, 800 few-shot tokens per call — yes: fine-tuning pays for itself in 3 days and saves $1,903 over the first year (99%). Tune the sliders for your real numbers.

How many calls until fine-tuning pays for itself?

54,375 total calls — the point where cumulative few-shot example-token waste equals the one-time training cost. At this page's volume that's 3 days. This number doesn't depend on your timeframe, only on total calls served.

Why isn't few-shot prompting free?

Because the example tokens you paste into every prompt bill as input tokens on every single call — $0.000320 extra per call here. That's invisible in any one bill but adds up to $1,920 over a year at this volume. Fine-tuning removes it permanently after a one-time training cost.

What does fine-tuning actually cost to train?

Llama 3.1 70B trains at $2.9 per million training tokens on Together AI (SFT LoRA). At 2,000,000 dataset tokens × 3 epochs = 6,000,000 training tokens, that's $17.40 — a one-time charge, not recurring.

Does the fine-tuned model cost more to run than the base model?

Not on Together AI — LoRA fine-tuned models serve at the same rate as the base model. The entire fine-tuning decision comes down to training cost vs. the few-shot tokens it eliminates from every future call; inference pricing itself doesn't change.

Are these prices current?

Fine-tuning training prices are hand-verified against Together AI's pricing page (checked 2026-07-13) and re-audited quarterly — they aren't in the daily-synced inference catalog. Inference prices sync daily from OpenRouter. All math runs client-side with tested code.