Snowflake priced its coding agent at $6.34 a trial, and the mode that won stays internal
Snowflake's October 27 benchmark, built from 65 of its own merged pull requests, discloses for the first time what an AI coding agent costs per task inside its product, and shows the winning mode is not for sale. The December print now has to answer whether Snowflake can bill for the outcome rather than the compute behind it.
Vincent Jiang · 3 min read
On October 27 Snowflake's+0.92% — Snowflake, up 0.92 percent today engineers published the thing no AI vendor had put on the record: a price per attempt for an AI coding agent on real production code 1. SnowPR is a benchmark built from 65 pull requests Snowflake's own teams merged, graded by the hidden tests that shipped with each change 1. Nobody outside the company can run it, because every task comes from Snowflake's private code history. But the cost lines are the first disclosed numbers of their kind inside a listed software company's own product.
The December print has to answer this
Shares closed October 27 at $337.42 2. The next reckoning is the fiscal Q3 report, expected December 2: product revenue guided to $1,588M to $1,593M, about 37 to 38% growth, against a 15.5% non-GAAP operating margin target 5. Cortex Code, the coding agent launched in February, had passed 9,100 customer accounts by the September quarter 5. What Snowflake has now disclosed is the unit economics inside that product.
Snowflake's growth troughed at 25.7%, then reaccelerated four straight quarters
Data
| Revenue growth, year over year | |
|---|---|
| Q3 '23 | 31.8% |
| Q4 '23 | 31.5% |
| Q1 '24 | 32.9% |
| Q2 '24 | 28.9% |
| Q3 '24 | 28.3% |
| Q4 '24 | 27.4% |
| Q1 '25 | 25.7% |
| Q2 '25 | 31.8% |
| Q3 '25 | 28.7% |
| Q4 '25 | 30.1% |
| Q1 '26 | 33.5% |
| Q2 '26 | 35.1% |
The winning mode is not for sale
Cortex Code runs in two modes. REM mode exposes the agent's file, shell and search operations as Python functions inside one tool, so the model chains several operations in a single step. On SnowPR, that mode with a frontier model cost $6.34 a trial and solved 62 of 65 tasks, against $6.51 for Anthropic's$350B — Anthropic, private, latest valuation $350B Claude Code, $6.93 for Cursor and $8.19 for Snowflake's own slower code mode 1. REM mode is internal only and not yet available to customers 1. Snowflake built the cheapest agent on its own scoreboard and kept it.
The cost of the task is falling off a cliff
Put a cheap model in the same winning harness and the numbers move harder than any headline: haiku-5.5 in REM mode passed 70.8% of trials at $0.24 each, against 64.1% at $0.26 for Claude Code, while OpenAI's$500B — OpenAI, private, latest valuation $500B Codex passed 74.4% at almost nine times the cost 1. The same task in the same harness costs about 26 times more at the frontier than on the cheap model, and Snowflake prices most of its AI as consumption 1.
The same task: $0.24 on a cheap model, $6.34 at the frontier
Data
| Value | |
|---|---|
| Cortex Code REM, haiku-5.5 | $0.24 |
| Claude Code, haiku-5.5 | $0.26 |
| Cortex Code REM, opus-5 | $6.34 |
| Claude Code, opus-5 | $6.51 |
| Cursor, opus-5 | $6.93 |
| Cortex Code code mode, opus-5 | $8.19 |
The fair objection
The benchmark is 65 tasks from one company's repository, graded by that company, and its own agent wins 1. That cuts both ways: the tasks are demonstrably hard, since changes spanning Java and C++ passed only 55% of trials against 80% for Java alone 1, and no public benchmark offers anything closer to this scale. But SnowPR is evidence, not an audit. Independent coverage of it does not yet exist; the numbers rest on Snowflake's own publication.
What to watch
The same day, Oracle−0.72% — Oracle, down 0.72 percent today and Snowflake announced the AI Data Cloud will run as a native, first-party service on Oracle Cloud Infrastructure, billed through Oracle 3, with no general-availability date confirmed 4. More distribution, same question. As reported on October 26, Snowflake sells AI cheaper than the market and earns thinner margins; the December print shows whether it can charge for outcomes instead of the compute behind them 34.
Get Platforms by email, free
The cloud, data and developer software AI is built on. In our daily email, every weekday.
