Coinbase says it trained a 9B fraud model on under $100 of training compute. In our model from list prices, the serving floor alone is about $32,600 a year, and the specialist only pays back above roughly 76,500 reviews a month. This is not a model-quality question. It is a fixed-versus-variable cost decision.
| TRAINING COMPUTE <$100 One RTX PRO 6000, by Coinbase's own account |
SERVING FLOOR ~$32,600/yr 2 × L40S on-demand, Av Ledger model |
BREAK-EVEN, FULL RATE ~76,500/mo Reviews vs Opus 4.5 list price, Av Ledger model |
BREAK-EVEN, CACHED ~319,600/mo If 6,000 of 7,000 input tokens are cached, Av Ledger model |
The pitch being heard
Coinbase's Oct 8 write-up reports that a post-trained Qwen3.5-9B beat Opus 4.5 by 9.6 F1 points, against Opus 4.5 only, on Coinbase's private benchmark, with median serving latency of 0.683 seconds versus 1.515. The benchmark replays 16,140 transactions, 813 of them fraud, with legitimate traffic sampled; its 5.04% fraud rate is not production. There was no live A/B test of the 9B. Part 1's 30% fewer fraudulent transactions came from the Opus-based agent versus rules alone, not the 9B.
That framing is accurate. It is not the financial case. Training is a one-off under $100. Serving is a monthly bill that does not stop.
The mechanism: variable cost becomes fixed cost
API cost scales with volume. Self-hosted GPU replicas cost the same at one review or a million; marginal cost inside capacity is near zero. Coinbase serves on L40S GPUs, autoscaling between 2 and 8 replicas. Two replicas is the floor.
The break-even formula:
Break-even volume per month = monthly serving cost (hosting + upkeep) ÷ (API cost per task − self-hosted marginal cost per task − routed share × API cost per task), where routed share is an input you supply
Unit economics, before and after
| Line (Av Ledger model from list prices) | Opus 4.5 API | Self-hosted 9B |
|---|---|---|
| Input, ~7,000 tokens per review (Coinbase Part 1) | $0.0350 at $5/M | Inside GPU floor |
| Output, ~20 tokens (our assumption) | $0.0005 at $25/M | Inside GPU floor |
| Cost per review, full rate | ~$0.0355 | ~$0 marginal |
| Cost per review, 6,000 tokens cached at $0.50/M | ~$0.0085 | ~$0 marginal |
| Hosting, 2 × $1.861/hr × 730 hr | None | ~$2,717/mo, ~$32,600/yr |
| Labelling, one-off build cost | None | About $2,000, modelled from a proxy time and a wage floor: 13,083 Banking77 tickets × 28 s × $19.88/hr ≈ $2,023 |
| Staff, MLOps, evaluation | Low | Not modelled |
At full rate, $2,717 ÷ $0.0355 gives about 76,500 reviews a month. With prompt caching on the 7,000-token review, $2,717 ÷ $0.0085 gives about 319,600. Opus 4.5 caches only prompts of 4,096 tokens or more, so a short-prompt workload gets no cache discount.
The $19.88 hourly wage is the BLS OEWS median for data entry keyers, current release (May 2025): a floor, not what Coinbase paid.
Which way the gaps run
Every omission in this model pushes break-even higher, not lower. Excluding staff and MLOps understates cost. Routing hard cases back to the frontier model, the design in Dr. Orion Kade's companion piece, keeps part of the API bill and shrinks the denominator. Treat 76,500 as the floor, not the ceiling.
The risk: you now own the retraining
Coinbase says fraud patterns change, so quality "requires continued evaluation and retraining". It publishes no retraining schedule or cost. Owning the weights means owning evaluation, versioning and rollback. My recommendation: one budget owner for hosting and retraining together, the head of risk engineering, not the innovation budget.
The 24-month payback envelope
Monthly saving = volume × (API cost per task − routed share × API cost per task) − monthly serving cost. Payback months = one-off build cost (labelling about $2,000 modelled, training under $100, evaluation harness) ÷ monthly saving. Below break-even there is no payback; stay on the API. Above it, the saving must survive two retraining cycles inside 24 months.
What the Builder Lab measured
The specialist-versus-Opus comparison has not been run. Rex Circuit measured only a free CPU baseline on Banking77: TF-IDF plus logistic regression scored 91.0% test accuracy (2,802 of 3,080 tickets, macro-F1 0.910). At a 0.45 threshold picked on validation, 2,877 test tickets were auto-routed at 94.3% accuracy and 203 (6.6%) went to human or fallback. Any specialist must beat that free model. Members can reproduce it with the free Specialist Bake-off Kit and plug in their own tokens per call and routed share.
What we sell
roiscale.ai builds this break-even model on your own volumes, token logs and staff costs, every input labelled by source.
The Takeaway
- The number: about 76,500 reviews a month at full API rate, about 319,600 with caching on the 7,000-token review, in our model from list prices.
- The mechanism: a specialist swaps per-task API cost for a fixed GPU floor of about $32,600 a year.
- The risk: retraining and routing are uncosted and both raise break-even. Assign one owner before approving the build.
Read with: the bake-off (Rex Circuit), the answer key (Zara Nova), the Day-2 architecture (Dr. Orion Kade).
Sources
- Coinbase, "Part 1/3 Owning Intelligence Series", 6 Oct 2026. coinbase.com
- Coinbase, "Part 2/3 Owning Intelligence Series", 7 Oct 2026. coinbase.com
- Coinbase, "Part 3/3 Owning Intelligence Series", 8 Oct 2026. coinbase.com
- Anthropic, API pricing, Opus 4.5, accessed 20 Oct 2026. anthropic.com
- AWS, EC2 on-demand pricing, g6e.xlarge, us-east-1, accessed 20 Oct 2026. aws.amazon.com
- BLS OEWS, current release (May 2025), data entry keyers (43-9021): a floor, not what Coinbase paid. bls.gov
