Coinbase says tuned Qwen3.5-9B beat Claude Opus 4.5 on Onramp fraud checks
Coinbase reports that a Qwen3.5-9B it tuned on proprietary transaction data beat Anthropic's Claude Opus 4.5 at fraud detection in Onramp, where a trial cut fraud 30%.
Why it matters
If it holds outside Coinbase's own tests, the result points to fraud screening that runs on in-house models instead of frontier API calls. Both the model comparison and the finding that newer closed models regressed come from Coinbase alone, with no outside replication reported, so the claims stay narrow in scope and the story carries no market, regulatory or pricing consequence.
Where the sources disagree
How Qwen3.5-9B was trained
engineering report published October 8
Crypto News Flash
Brian Armstrong post dated October 12
CryptoPotato
Crypto News Flash says Coinbase used reinforcement learning with proprietary transaction data, while CryptoPotato describes the same work as fine-tuning on internal data.
Key facts
Coinbase tuned Alibaba's open-weight Qwen3.5-9B on proprietary transaction data to score transaction risk in Onramp.
Reported by Crypto News Flash
The tuned Qwen3.5-9B outperformed Anthropic's Claude Opus 4.5 on four fraud-detection metrics in Coinbase's internal evaluation.
Reported by Crypto News Flash
The largest improvement was in detecting the monetary value of fraudulent transactions.
Reported by Crypto News Flash
An online experiment using the added AI risk agent found 30% fewer fraudulent transactions.
30 %Reported by CryptoPotato
The same online experiment found 22% less fraud value.
22 %Reported by CryptoPotato
Coinbase tested newer closed-source models against a fixed benchmark of 16,140 transactions across 7,293 users, including 813 confirmed fraudulent transactions.
16,140 transactionsReported by Crypto News (CryptoNews.net)
Newer Opus, Sonnet and GPT versions performed worse than their predecessors on several fraud-detection measures.
Reported by Crypto News (CryptoNews.net)
The Onramp AI risk agent reviews recent transaction sequences and can block a transaction that passed the initial checks, but cannot reverse an upstream block.
Reported by Crypto News (CryptoNews.net)
What happened
Coinbase tuned Alibaba's open-weight Qwen3.5-9B on its own transaction data to score risk inside Onramp, the service that sells crypto through partner applications and allows guest checkout without a Coinbase account. Crypto News Flash reports the tuned model beat Anthropic's Claude Opus 4.5 across four fraud-detection metrics, with the largest gain on the monetary value of fraudulent transactions, and describes the training as reinforcement learning carried out for an engineering report published on October 8. CryptoPotato quotes Coinbase CEO Brian Armstrong, writing in an October 12 post that specialist models increasingly beat frontier generalists on narrow tasks, and reports that a separate online experiment adding an AI risk agent to Coinbase's existing machine-learning model and rules engine found 30% fewer fraudulent transactions and 22% less fraud value. CryptoNews.net carries the same CryptoPotato text, so its version is not a separate check on those two percentages. Coinbase also said newer Opus, Sonnet and GPT versions performed worse than their predecessors on several fraud-detection measures when tested against a fixed set of 16,140 transactions from 7,293 users, including 813 confirmed fraudulent ones. The comparison against Claude Opus 4.5 is published without a sample size, a scoring method or a cost figure, and Crypto News Flash notes the tests do not show smaller models beating frontier AI on broader applications.
How the story developed
- Official statement
Coinbase publishes engineering report on its tuned Qwen3.5-9B
The report says the tuned model beat Claude Opus 4.5 across four fraud-detection metrics.
- Official statement
Brian Armstrong posts about specialist models
Armstrong wrote that specialist models increasingly outperform on narrow tasks and quoted the model as beating frontier models at lower cost and speed.
- First report
First report by Crypto News Flash
Coinbase’s AI Experiment Challenges the Need for Bigger Models
- Independent corroboration
Independently corroborated by CryptoPotato
Coinbase AI Agent Cuts Fraud by 30% in Onramp Experiment
- Update
CryptoPotato reports the Onramp fraud experiment figures
Reports 30% fewer fraudulent transactions and 22% less fraud value in the group with the added AI review.
How coverage built up
Independent sources over time, counted the way the consensus panel counts them: a republication of a wire story does not move the line.
- 12 Oct 2026, 06:00 UTC: Crypto News Flash — 1 independent source, 1 reports
- 12 Oct 2026, 07:03 UTC: CryptoPotato — 2 independent sources, 2 reports
Affected entities
Original sources
| Publisher | Report | Role | Published |
|---|---|---|---|
| Crypto News Flash Crypto media | Coinbase’s AI Experiment Challenges the Need for Bigger Models | Original | |
| CryptoPotato Crypto media | Coinbase AI Agent Cuts Fraud by 30% in Onramp Experiment | Independent | |
| Crypto News (CryptoNews.net) Aggregator | Coinbase AI Agent Cuts Fraud by 30% in Onramp Experiment | Syndicated |
Related events
Brian Armstrong Rejects Coinbase Blame as CLARITY Act Senate Vote Falls Short
Brian Armstrong denied blame for Coinbase in the CLARITY Act's Senate defeat; Bitcoin.com News said the Wall Street Journal had not published the alleged story, while Bitcoin Insider said it had.
Coinbase proposes quantum-resistant Bitcoin signatures that could increase signature data tenfold
Coinbase outlines a plan for quantum-safe Bitcoin addresses and signatures, including a draft scheme where stateless signatures reach 5,777 bytes.
BlockTower's Ari Paul says Coinbase lost $25M of firm funds and hid over $1B in hacks
BlockTower Capital founder Ari Paul alleges Coinbase lost $25M of his firm's funds and covered up hacks worth over $1B; Coinbase says it messaged him to look into it.
Coinbase data shows Grok drove 59.5% of AI agent trading volume in a week
Coinbase chart puts Grok at 59.54% of AI agent notional volume for Sept. 14-21, against 1.54% for ChatGPT, 13.37% for Claude and 21.73% for custom CLI setups.
Armstrong Backs SEC Proposal to Add Knowledge-Based Paths to Accredited Investor Status
Coinbase CEO Brian Armstrong backed SEC notices on October 11 that would let FINRA exam takers and CPA, CFA and CFP holders qualify as accredited investors without the wealth test.
Coinbase CEO says stablecoins will power payments by AI agents
Coinbase CEO Brian Armstrong said crypto and stablecoins will serve as the payment layer for AI agents, with Coinbase for Agents starting as a trade interface inside ChatGPT and Claude.