What Fine-Tuning Actually Cost: A Full Audit

The Borg estate, audited 2026-08-29. Five LoRA training runs, three adapters, two memory systems, one merciless promotion gate — and every token counted.

---

Figure 1

Four honest rates, drawn on the same axis

• Allocated or list-priced
• Marginal cash on owned hardware

Training the models cost almost nothing at the token counter. Every token that touched a model during this program — teacher data generation, five training runs, all exams and canaries — prices out at about $75 at public API list rates (≈$52 with a 30% enterprise discount, under $1 in actual marginal cash because the runs sat on owned hardware, local models, and free preview windows).

That $75 is not the bill, and neither is a stack of seat fees. Prepaid cash is a fixed $214.90 after tax per Claude Max subscription and $214.90 after tax per ChatGPT Pro subscription (Utah). What this paper reports next is how much of a weekly allotment the work used, in money people already understand: one Max bar is $8,000/month ($2,000/week) of API-equivalent capacity; one Pro bar is $14,000/month ($3,500/week). Fine-tuning used 25.4% of one weekly Claude Max bar and 64.2% of one weekly ChatGPT Pro bar, which is $2,754 API-equivalent ($508 + $2,246). The whole memory-system program used 74.9% / 98.8% of those same weekly bars ($4,956 API-equivalent). If more than one account ran, those percents add (80% + 70% = 150%). The counted-token $75 versus the $2,754 allotment is the discrepancy table below: untracked session work, not a second cash bill. The reusable part is still the harness.

---

1. Claude Max prepaid cash: $214.90 after tax per subscription (Utah). Fixed monthly cost. One Max seat is treated as $8,000/month of API-equivalent allotment = $2,000/week. Tax is regional; $214.90 is this account's after-tax figure. The $8,000 figure is the square bucket for this paper, not a metered invoice.
2. ChatGPT Pro / Codex prepaid cash: $214.90 after tax per subscription (Utah). Same shape. Allotment is $14,000/month = $3,500/week (square bucket). Percents across accounts add: 80% + 70% = 150% of one weekly bar.
3. Attribution window: 2026-08-19 → 2026-08-29 (10.4 days = 1.486 weeks): memory-layer build began 08-19, pair harvesting 08-20, student training 08-27→29 (v3 was still training when this audit ran).
4. Attribution method: session-transcript byte share. Every Claude session (142 files, 757MB) and every Codex rollout (11,603 files, 9,083MB, all account homes) modified in the window was scanned in full and tagged by fine-tuning keyword frequency (mlx_lm, lora, adapters.safetensors, training pairs, canary-arm, exam-4b, …) with a ≥40-hit threshold so shared session preambles cannot false-tag. Bytes proxy tokens; JSONL overhead inflates numerator and denominator alike.
5. Token counts are measured, not guessed. Training tokens come from mlx_lm's own Trained Tokens counters summed across every resume segment. Dataset and teacher tokens come from the real Qwen3-4B tokenizer run offline over 400-line random samples (seed 42) per corpus, extrapolated by mean × line count — and used as a single cross-teacher proxy (true per-provider tokenizers differ ±10–20%).
6. Prices are public list, retrieved 2026-08-29 from official pricing pages (Anthropic, OpenAI, xAI, Alibaba, DeepSeek; sources in the appendix). GPT-5.6 Sol is in a promo ($4/$20; durable list $5/$30). Sonnet 5's $2/$10 is now standard.

Figure 2

Five LoRA runs, trained tokens as written

• Measured trained tokens
• Projected, still training at audit

Five mlx_lm LoRA runs on Apple Silicon across 2026-08-27→29:

Gross training compute: ≈78.8M trained tokens, ≈39 GPU-hours.

The crowned student scored 400/400 format-valid, 0.644 entity-agreement on the held-out exam — then failed the promotion canary 1-of-25 against the 27B incumbent because it had been trained on one of the pipeline's six call shapes and scored 0/534 on a shape it had never seen. That verdict, not any benchmark, is why v3 exists — trained on all six shapes recorded from live traffic.

Prepaid cash is fixed: $214.90 after tax per Claude Max and $214.90 after tax per ChatGPT Pro (Utah; tax is regional). This section does not multiply seats. It reports what percent of a weekly allotment the work used. Percents across accounts add. Over 100% means more than one weekly bar.

Square buckets (owner-chosen): Claude Max $8,000/month = $2,000/week. ChatGPT Pro $14,000/month = $3,500/week.

How the percents were built. The withdrawn table treated $2,000/week and $3,500/week as cash and assumed every week in the 1.486-week window was 100% burned, then took session-byte share of that window. Those byte-shares are a proxy for percent of weekly allotment, not a vendor usage dashboard. Reverse of that table:

• Claude fine-tuning share of session bytes: 17.1%. Cumulative weekly-Max equivalent: 17.1% × 1.486 weeks = 25.4% of one weekly Claude Max bar.
• Codex fine-tuning: 43.2% × 1.486 = 64.2% of one weekly ChatGPT Pro bar.
• Whole program: Claude 50.4% × 1.486 = 74.9%; Codex 66.5% × 1.486 = 98.8%.

If a second Max account had been in the tagged set, its percent would add (80% + 70% = 150% of one weekly Max). Same rule for Pro.

Prepaid cash does not scale with those percents. The percents scale the $8,000 / $14,000 allotments.

Grok ran in the bake-off and research lanes; it is not given a square allotment in this table. Ox Alpha was a free evaluation window ($0). Electricity is Rate #4.

The gap is untracked session work (engineering tokens that never entered the LoRA counters) against the $8,000 / $14,000 square buckets. Both numbers stay. Neither is cash. Cash is the $214.90 prepaid.

What the same tokens would have cost bought at list, per official 2026-08-29 pricing.

Teacher-side (data generation), measured tokens × list rates:

Training compute at hosted-LoRA rates (Fireworks $0.50/M, Together $0.48/M training tokens for ≤16B LoRA): 78.8M × $0.50 = ≈$39.4. (OpenAI's fine-tuning platform is winding down — new users cannot start — so open-model hosts are the benchmark.)

Evaluation inference (≈1,900 graded exam generations, 842 canary calls, probes): ≈$4 — generous.

Rate #2 total: ≈$75 at public list.

Cross-checks: pricing the local 27B at Alibaba's managed qwen3.5-plus rates instead of the Together flat proxy moves its line from $11.81 to $24.61 — the total stays under $90 on any defensible proxy choice. Luna at its official $0.20/$1.20 is the quiet star: the entire 7,026-pair capture corpus cost $4.28 of teacher tokens at list.

Where the 15–50% band comes from (primary sources, retrieved 2026-08-29): Google 1-year provisioned throughput = 26% off monthly PT; AWS Bedrock reserved-tier guidance = 30–50% with commitment (published provisioned examples run 44–52%); OpenAI reserved capacity ≈15% (1-year vs 3-month); batch/flex tiers at 50% are public at OpenAI, Anthropic, Bedrock, and Vertex. Azure's famous 64–70% PTU reservation savings are off capacity-hours, not per-token list, and are excluded. First-party private-deal percentages are not published; the owner's 30% sits comfortably inside the defensible band.

Owned hardware, local models, free preview windows, flat-rate subscriptions already paid: ≈$0.71 of electricity (≈39 GPU-hours × ~150W × $0.12/kWh) and nothing else.

$2,754 of agent-engineering per $75 of model tokens. Fine-tuning's cost lives in the loops around training: designing harvests, babysitting crashed runs, building exams, arguing with canaries. That is precisely the part that is now a reusable, open-sourced harness — which converts the next program's $2,754 into something much closer to its $75.

Figure 4

Pairs harvested, pairs kept

• Measured harvest
• Discarded before training

The owner's hypothesis going in: "we got ~15,000 pairs; 5,000 were enough; really 3,000." The evidence agrees — and sharpens where the real deficit was.

Volume was over-provisioned ≈4×:

• 25,270 pairs harvested from two teachers → 14,457 unique (42.8% discarded before training touched them). The second teacher (Ox) duplicated 83.7% of the first's episodes: +100% token spend for +14.8% unique pairs — 6.2× worse economics per kept pair.
• Validation loss flattened by iteration 600–900 on both model sizes — ≈3,600 pair-presentations, under 27% of one epoch. The 1.7B passed the quality bar (0.58 vs 0.53) at step 900: 13% of its planned compute. The crowned 4B is the step-2,800 checkpoint of a 6,900-step plan: 59% of its planned compute was never needed.
• Marginal quality per extra compute, measured: v2's val loss 0.393 at step 600 → 0.356 at step 3,000 (−9% for 5× the compute); exam agreement 0.58 at 13% of training → 0.644 at 41%. Classic asymptote.

Coverage was under-provisioned 6×: the pipeline makes six call shapes; the corpus covered one. The exam (drawn from the same one shape) could not see it; the canary could: 0/534 on resolve_edge, verdict DO-NOT-PROMOTE. The fix needed just 2,201 new pairs — of the right shapes, recorded free from live traffic — plus a 2× reweight, not another 13,657-pair harvest.

Right-sized plan, in hindsight: 500–900 pairs per call shape × 6 shapes (≈3,000–5,400 total — the owner's "3–5k" almost exactly), one epoch with exam-at-checkpoint early stopping (≤1,500 iterations), single teacher, recorder from day zero. At list rates that is ≈$25–30 instead of ≈$75 — but the dominant saving is upstream: roughly half the harvest-and-babysit engineering inside the $2,754 attaches to the excess (the duplicate second-teacher harvest, the over-long runs, the crash-resume cycles around them).

Capability density of open models doubles roughly every 3.3–3.5 months (Xiao et al., arXiv:2412.04315; published in Nature Machine Intelligence 7:1823–1833, 2025). Every adapter is therefore a depreciating asset with a ~one-quarter half-life: the base models it beats today are matched by something smaller and cheaper within months.

Our own program ran ahead of that curve: the crowned adapter went from exam triumph to benched in 26 hours — not by densing but by its own promotion gate. Either way the conclusion is identical, and it answers the question a reader posed while this audit ran (if density keeps doubling, doesn't the durable value sit in the harness?): yes — the integral value is the harness. The harvester, the grammar shim that turns production traffic into labeled pairs, the exam, the canary, the crash-tolerant training chain — that machinery re-targets any new base model for ≈$11 of compute-equivalent and a half-day of wall clock, forever. Train adapters like you buy milk, not like you buy a house. That is also the strongest argument for open-sourcing the adapters themselves: their moat value decays on a fast clock, while their teaching value (honest cards, reproducible gates) does not.

1. Constrained decoding before fine-tuning. The grammar shim already forces valid schema output from stock local models — the exact failure class that killed the canary. Fine-tune for judgment and speed, never for JSON shape.
2. Off-the-shelf floor first. Zero-shot joint entity/relation extractors exist on Hugging Face (GLiNER-Relex family; NuExtract-class structured extractors) — no pre-trained adapter for our schemas exists (LoRA adapters are base-and-task specific, which is also why ours only fit their exact bases), but a zero-shot floor should be measured before any distillation is funded.
3. One teacher until the pilot proves a second adds unique coverage. Measure dedupe rate on a 500-pair pilot before buying a full second harvest.
4. Recorder from day zero. 168 pairs/hour, all six shapes, $0. The corpus is a flow, not a stock.
5. Early-stop on the exam, not the plan. Checkpoint every 200–300 steps, exam each checkpoint, stop at the elbow. Compute plans written before the loss curve are fiction.
6. Semantic metrics for sentence-facts. Exact-string Jaccard scored correct paraphrases as 0.008 and nearly condemned a working adapter.
7. Batch APIs for any paid teacher (−50%, no negotiation), and cache-priced prompts for repeated harvest preambles.
8. Keep the incumbent warm. The 27B fallback is why a failed canary cost a day, not an outage.

Anthropic platform pricing page (Opus 5 $5/$25, Sonnet 5 $2/$10, Haiku 4.5 $1/$5, batch −50%, cache read 0.1×) · OpenAI developer pricing (GPT-5.6 Sol $4/$20 promo to ≥2026-11-21, durable $5/$30; Terra $2/$12; Luna $0.20/$1.20; GPT-5.3-Codex $1.75/$14; fine-tuning platform winding down) · xAI docs (Grok 4.6 $2/$6, 2× ≥200k) · Alibaba Model Studio Intl (qwen3.8-max $2/$6; qwen3.5-plus $0.40/$2.40) · DeepSeek (V4-Flash off-peak $0.22/$0.66) · Together/Fireworks fine-tuning ($0.48–0.50/M LoRA ≤16B) · enterprise-discount primaries as cited in §5. Full URL list ships in the audit evidence pack.

Every number above traces to a file: mlx_lm run logs (Trained Tokens counters, loss series), BUILD-REPORT.json dataset waterfalls, exam scorecards (exam-v1-step900 → exam-4b-full-merged), the canary pack (CANARY-2026-08-28), the teacher bake-off directory, the recorder's append-only pair file, and the session-byte census scripts. The audit ran read-only while v3 trained live on the same GPU.