On Tuesday, September 22, Anthropic and OpenAI shipped new models about ninety minutes apart, and both led with the same number: the price. Anthropic put Claude Opus 5.5 at $4 per million input tokens and $20 per million output — 20% below Opus 5, and around 40% cheaper on a typical workload. OpenAI then halved its own: GPT-6 Sol dropped from $4 and $20 to $2 and $10, and the small model, GPT-6 Luna, went to 10 cents and 50 cents per million.
That is the second round of cuts on frontier models in under a month. Nothing about the trend is new; what is new is how far it has now gone, and what the cheaper models are explicitly built to do.
The other half of the announcement was accuracy
Price cuts usually come with a catch. This time both companies claimed the opposite. OpenAI says Sol makes about half as many mistakes as its predecessor on its internal factuality tests — while noting those tests draw on conversations already flagged for errors, so they are not a picture of ordinary use. Anthropic says 16 of 18 of Opus 5.5’s research reports cleared a bar where a single invented figure or quote counted as a failure, and that the model is less likely to act outside the boundaries it has been given.
Read what the cheap model is aimed at: “high-volume jobs with a clear goal, such as summarising documents or answering short questions.” That is a description of a front desk. “They are also half the price per token, and even less per task,” Sam Altman wrote.
Why this lands differently for a church office
Most businesses weigh an answered question against revenue. A church cannot. There is no ticket price on “what time is the 9am service, and where do I park?” — so the cost of answering has always been the whole argument, and it has always been paid in someone’s Tuesday afternoon.
The volume is real and it is lopsided. Service times, campus locations, parking, kids check-in, accessibility, whether the building has a ramp, when the newcomers’ lunch is, how to sign up to serve — a handful of questions account for most of what arrives, and a large share of it arrives on Saturday night and Sunday morning, when the office is shut and the staff are working. Multi-campus churches carry it twice, because the right answer depends on which building the person means. Congregations that speak several languages carry it again. None of that is complicated work. It is repetitive work at inconvenient hours, which is precisely the category that just got cheaper and steadier. What that looks like for a church and its campuses is not a new idea — it is an old one that stopped being expensive.
Where to start before the next big Sunday
Pull the last two weeks of voicemails, emails to the office and messages to the church Facebook or Instagram page, and count. You will almost certainly find five or six questions that make up most of them, and a cluster of them landing outside office hours. That list is the entire build.
Then write the second list, which matters more: what the assistant must never answer. Prayer requests, bereavement, counselling, anything about a family in crisis — those get recognised and passed to a person quickly, not handled. A system that answers the routine questions on every channel and hands the rest straight to your team is the version worth running; one that tries to counsel someone at 11pm is not.
Do the counting this week. The price of leaving those messages until Monday hasn’t changed. Everything else about the maths has.
Working in churches & communities? See how TalkRev works for Churches & Communities.
The TalkRev blog
One short email per new post
No ads. Unsubscribe anytime.
