đź’¸ Three models tied at 97%. Now pick on price.
The top three finished level, and the cheap one came within three points.
Three-minute read. The lead is what a fresh set of evals does to your model bill.
PRICING
Three models tied at the top, so price is the only vote left
What changed. Guillermo Rauch published fresh Next.js evals: Opus 5.5, GPT-6 Sol and Fable 5.1 all landed at 97%, and Grok 4.7 came in at 94% while being called two to seven times cheaper.
What it means at your size. When the options finish level, the thing you actually feel is the invoice, not the capability. A three point gap on someone else’s coding test is not worth a bill several times larger on your support replies.
The first step. Take the highest volume, most repetitive AI job you already run, the boring one: support replies, summaries, data entry. Run a week of it through the cheaper model alongside the one you use now, and read the two sets of output side by side before you compare the two prices.
The limitation. This is one person’s test on one kind of work, Next.js code, and code is not your quotes or your support queue. GPT-6 Sol and Opus 5.5 both launched today, so nobody has run either at volume yet. We have not tested any of them on your work.
Rauch on the fresh Next.js evals (x.com)
COSTS
The same capability gets about 47% cheaper every quarter
Epoch AI Research puts the fall in cost for a given level of AI performance at roughly 47% per quarter since 2023, four to fifty-four times faster than any earlier transformative technology. The steepest falls are at the top: around 66% a quarter while a capability is new, slowing to around 32% a quarter after two years. Read that as a rule about contracts rather than a headline. The quote you are given today is the most that job will ever cost, so a long lock-in at today’s price is the expensive way to buy, and anything you priced this year and walked away from is worth pricing again.
Epoch AI Research on falling AI costs (x.com)
SUPPORT
A support spike absorbed without hiring, if the numbers hold
Min Choi passed on a claim from SpaceX and xAI that a Grok Bot may have taken a 175% jump in support volume with no new hires, at $0.20 to $0.30 a resolution. The hedge is the source’s own, and there are no tickets and no invoice behind it. So hold the figure as a reference price rather than a result: it is the number to put next to what one of your own tickets costs you, the next time somebody quotes you for a support tool.
Min Choi on the Grok Bot support claim (x.com)
Just shipped
Make added MCP tools for managing webhook queues, live now and written up in Make’s own community post. If your automations run on Make and a queue has ever backed up without anyone noticing, this is the part an agent can now inspect and clear for you. (community.make.com)
The rest of the day
GPT-6 Sol and Opus 5.5 both launched today, and Dan Shipper says Every has been testing both. (x.com)
Elon Musk calls Grok 4.7 a good showing for a smallish model, which is the evals result seen from the other side. (x.com)
Peter Yang has posted his own run at Opus 5.5, if you would rather watch someone use it before you try it. (x.com)
Too early
Running models on your own hardware. Yohei Nakajima ran the same job locally and reports slower inference but no network latency. The tradeoff is real, and it buys privacy and predictability at the cost of speed, but it also needs somebody in your building who keeps that machine alive. Nothing to act on this year. (x.com)
🍕 One for the road
Airtable put the question to its friends: is a calzone a pizza? You already have an answer, and you are already a little annoyed that anyone disagrees. (x.com)