Two of the three stories on Thursday's Value-First AI Daily board were new models, and neither one asks for more money than the thing it replaced.
Anthropic released Claude Fable 5.1 and cut the price of the line item that runs up a long agent job. That item is re-reading context the model already holds, and it now costs a quarter of what it did. Meta released Muse Spark 1.3 and left its input price alone at a dollar twenty-five per million, the same as 1.2, while the independent index that ranks these things read four points higher for the version anyone can call.
The third story, the one ranked first, is about the input nobody has been billed for yet.
The Justice Department Filed an Argument, Not a Ruling
The Justice Department told a federal court that training AI models on copyrighted text is fair use. Per IPWatchdog, no arm of the federal government had taken a side before in the copyright suits AI companies face from authors, publishers, music labels and news organizations. This one is the combined action against OpenAI from The New York Times and hundreds of local newspapers. Its reasoning: a model turns the works into numbers, so nothing protected ever reaches a reader.
Writ, the seat whose craft is legal risk, drew the edges:
A Statement of Interest is argument, not a holding — it binds no one, and the judge can reject it whole. The text that already binds is EU AI Act Article fifty-three, which turns on whether the opt-out was honored, not on how transformative the training was, and it applies no matter where the training happened.
A Munich court has already ruled against OpenAI on the same kind of claim. This is not a settled question. It is what one party filed.
Nico Lafakis argued the other side on air, as a fork: "either you're going to gate the information and we are just going to sit here and take decades to get to where we could have in, like, a year. Or, yeah, like, you could just give up the data, give up the information, and we can get there tomorrow" (21:34).
That is the price question under the legal one. Every other number on the board is a bill somebody sends. This is a bill somebody is suing for the right to send.
Anthropic Published Its Own Regressions
Fable 5.1 is the model Anthropic now puts at the top for hard coding and reasoning work, live on its own service and on the Amazon, Google and Microsoft clouds. Anthropic also published what it does worse than the version before it: one tool call per turn where the older one batched several, and a habit of rewriting a whole file for a small edit. Its own guidance still sends most work to Claude Opus 5.
Crucible, the seat that grades models, said which half to weigh:
Those benchmark figures are the lab grading itself against its own earlier model — diagnostics, not a verdict. Weight instead what cost them something to publish: their own notes say it may take one tool step where the older one took several, and rewrite a whole file for a small edit, and that most work should still start with Claude Opus five.
Nico landed on the file-rewrite line, because he has seen it before: "But at first, it would just rewrite the entire thing. And it's like, yeah, of course you're running out of context" (17:33). His verdict: "that's a major, major flaw" (17:41).
Hold that next to the discount. The cheaper item is reading. The regression is writing. A cut on one does not cover a habit on the other.
Meta Held the Price and the Index Moved
Muse Spark 1.3 is frontier-class despite the small number in its name. Artificial Analysis, which runs its own tests rather than reprinting a lab's, puts it second on its intelligence ranking — behind Anthropic's two strongest and ahead of everything OpenAI has out.
Pax, the CFO seat, read it as a price story:
The price didn't move: a dollar twenty-five per million input, same as one point two. The independent Artificial Analysis index reads sixty-one for the variant you can call today, against fifty-seven for one point two. That is more capability at an unchanged price, at roughly an eighth of Anthropic's input price.
Meta published the catch in the same announcement: that ranking belongs to its highest reasoning setting, open only to partners in limited preview while safety testing finishes. The setting anyone can call sits a step below.
Nico opened the breakdowns underneath the ranking on screen. The coding columns read well. The agentic one did not, and he said why that is the one that decides it for his work: "having to open up the browser in order to take a screenshot to verify what's going on, that kind of stuff is like that's SOP at this point. So if the model can't do that, then it's useless to me, at least" (13:33).
Episode — Value-First AI Daily
Value-First AI Daily - Sep 3, 2026
Chris Carolan and Nico Lafakis, September 3 — the board read cold in the first half, the youspot drive in the second.
Open the episodeThen the Price That Went Up
The second half of the hour was day five of driving youspot.com on air — Chris Carolan's own account, his own LinkedIn data, live, running through UNBOUND in Boston on September 16 to 18. youspot is Dharmesh Shah's, HubSpot's founder and CTO, and it changed price this week.
The front door now sells YouSpot Pro at $10 a month: "Use up to 1,000 credits/month · Store up to 1M objects · 3 connected mailboxes · Support for HubSpot, GMail, GCal, X/Twitter and LinkedIn." The $1 Solo plan is still live at its own URL, linked from nowhere on the site. He posted the new price himself on Tuesday night: "Just $10/month for the Pro version with no long-term commitment." Nothing on the site says anything changed, and none of those pages had moved when they were read again Thursday morning.
Chris Carolan called it "the 10X increase in cost" on air (10:03), then repeated what he had told the maker when the maker asked publicly what it should cost: "whatever it needs to, to keep this the most magical experience thing happening, cover your costs" (29:27). On the choices that question offered: "$10 wasn't an option, by the way" (29:41).
Then he read the Pro list off the screen and did the thing none of the day's other stories can do for you. He asked the seller, by name, on camera, before paying:
is this legitimate? Because if it is, I'll upgrade right now. But if it isn't, I'm terribly disappointed. (42:10)
At $10 a month, it says, support for HubSpot, Gmail, G-Cal, Twitter, and LinkedIn. If, by that, you mean coming soon, you should probably include that… (42:22)
Three minutes later the screen answered part of it. "Even though my HubSpot is connected, it's still not able to" (44:59). The import was, in the phrase used on air, "not yet available" (45:18).
The One Answer He Could Check
In between, he ran the test the board items cannot run on themselves. Since last year's INBOUND he has been sending connection requests to HubSpot employees one at a time, off LinkedIn's own suggestions page, never knowing how many accepted. So he asked: "how many have I connected with since then?" (32:51). The answer he read off the screen: "1,553 since September 1" (33:01). The tool volunteered its own limit in the same breath, "connected on is the acceptance date, not when you sent the request" (33:17), which is more disclosure than most numbers arrive with.
Then he asked which teams and titles those people sit in, and checked the answer against what he believed: "this matches my experience. It's overwhelmingly a sales floor network" (34:10). Then the part he had not intended, off youspot's own breakdown: "roughly 45% of the campaign landed in HubSpot's direct sales work, which is not my intent" (34:26). He also caught the tool putting words in his mouth, flagging a suggested follow-up as not his: "I don't type or say things like that" (33:44).
None of that proves the tool understands what it holds. It proves records landed and one answer matched a memory. Nobody independent has published a hands-on test of youspot's extraction. This drive, on camera, is currently the test.
The labs moved a price and printed, beside it, the reason you might not want the thing yet. The app moved its price tenfold and printed a list. On Thursday one buyer read that list aloud, asked the seller on camera whether it was true, and reached the end of the hour still on the dollar plan. The line that came back off the screen: not yet available.
Worth passing on?
