Our email agents now run on GPT-6 Luna.
Five of our email agents now run on OpenAI's GPT-6 Luna. Three moved on 28 September, following the comparison described below: the agents that track commitments, summarise threads and draft replies. Two further agents, which draft outbound emails and sort follow-ups, were introduced on 29 September and have run on GPT-6 Luna from the start.
The comparison
GPT-6 Luna reached our production platform on 23 September. That day we replayed real work from our own mailbox through three of the agents, once on the model they were then using, GPT-5.6 Luna, and once on GPT-6 Luna, and scored every output against each agent's own checks.
| Agent | Sample | GPT-5.6 Luna | GPT-6 Luna | Quality check |
|---|---|---|---|---|
| Thread summaries | 28 | 10.5 s | 5.4 s | Fully cited: 93% to 100% |
| Reply drafts | 10 | 8.8 s | 4.4 s | No invented commitments in either |
| Commitments | 40 | 4.3 s | 3.1 s | Usable output: 93% to 100%; every commitment quoted from its source in both |
Times are the average per answer. At list prices, each answer cost between 38% and 49% less on GPT-6 Luna, even though the model writes more tokens because it reasons before answering.
An issue found on the first run
On the first run, three of the 28 summaries came back empty. GPT-6 Luna spends more of its output allowance on reasoning, and in those three cases the allowance ran out before any summary was written. We raised the allowance from 1,500 to 4,000 tokens and ran the comparison again: all 28 summaries were complete, with no loss of speed. The higher allowance went to production together with the model.
What remains open
The reply-draft sample was small, at ten drafts. On it, GPT-6 Luna suggested a next step in eight drafts, against nine for GPT-5.6 Luna. We are monitoring this in production rather than treating ten drafts as conclusive either way.
The models our agents use, and how the work is divided between them, are listed on Models.