AI News, Oct 25: OpenAI Is Buying Anthropic's Customers
The Big Story: OpenAI is buying Anthropic’s customers at a discount
OpenAI is cutting prices by as much as 90 percent to pull Anthropic’s largest accounts across, according to reporting on The Information’s “Anthropic’s margin is OpenAI’s opportunity”. The named accounts are specific enough to check. At Rippling, OpenAI’s share of the coding workload went “from 16 to 52 percent” while “Anthropic’s share fell from 57 to 18 percent,” per CEO Parker Conrad. Replit switched after OpenAI offered “more than 30 percent off a contract worth over $100 million a year.” DataCamp’s CEO said OpenAI’s pricing “cut his costs by 80 to 90 percent.”
The sharpest number is the smallest. A US bank with over $7 billion in assets doubled its ChatGPT seats to 500 employees at no extra charge, and pays under $7 per user against a usual rate of $20 or more. That is not a volume discount, it is a customer being handed seats. An investor memo cited in the same reporting says “Anthropic’s growth slowed from Q2 to Q3.”
The 90 percent figure is one customer’s cost reduction rather than a published price cut, and these are negotiated deals described by a single outlet. But Anthropic is weeks from a reported listing, which is the worst moment to have your biggest logos quoted explaining how much cheaper the competition got.
The sellers are cutting prices; the buyers still can’t show a return
Apollo chief economist Torsten Slok published a note Saturday arguing that “the technology has delivered zero financial payoff to corporate America outside of tech”. His charts split the index: “Revenue per worker is rising for the Mag 7 but is flat for the S&P 493. Profit margins are rising for the Mag 7 but are flat for the S&P 493.” The line that explains the price war is the next one, that mainstream corporate America “is pulling back and starting to view token spending not as a source of growth but as a cost center that weighs on margins.”
Then a body count. TechCrunch’s running list of dead AI projects cites S&P Global Market Intelligence that “about 42% of AI initiatives are ultimately abandoned by their corporate parents.” Its largest entry, Character.ai, raised $150 million at a $1 billion valuation in 2023 and had its technology licensed by Google for a reported $2.7 billion in 2024. It filed for bankruptcy this month, and Disney is the opening bidder at $20 million for a license to some of its intellectual property.
Slok’s note then became the weekend’s biggest technical argument at more than 560 points without resolving: the gains are merely waiting on firms rebuilt around AI, or they are real but captured entirely as vendor revenue, or it is simply too early, which drew the reply that “people have been saying it just recently got good for about two years now.” When buyers cannot agree whether a thing works, sellers compete on price.
Today’s Top Stories
Elvis Presley’s rightsholders sued Anthropic over 274 songs
A subsidiary of Authentic Brands Group, with publishers and trusts holding rights to Presley’s catalog, sued Anthropic on Friday over 274 songs, alleging it trained on the compositions and lyrics without permission and that Claude reproduced protected material even after safeguards were added. Plaintiffs seek “statutory damages of up to $150,000 for each of the 274 songs, amounting to approximately $41.1million.” It is the third music-publishing suit against Anthropic since August: Round Hill filed in mid-August, and Reservoir four days ago over lyrics from 853 compositions, which makes this less a landmark than an assembly line.
Forced to pick a candidate, Grok picks Republicans and GPT picks Democrats
Convovo, a studio led by Cornell professor Tapan Parikh, pushed both models through their APIs across nearly 500 midterm races, demanding one name and no hedging. Grok 4.7 named the Republican in 92 percent of answers, GPT-6.1 Sol the Democrat in 97 percent, and the two agreed less than 10 percent of the time.
The diagnostic leg beats the headline. Rerun with search disabled across 30 races, GPT still picked 24 Democrats, while Grok refused outright in 27 of 30. Parikh’s read: “For GPT, it has mostly decided already…it’s in the weights. For Grok, it really is the search pattern.” Grok’s citations skewed to a Christian conservative voter guide in 39 percent of answers and X in 41 percent. Two models, two different failure modes, so there is no single fix.
Anthropic’s co-founder moved self-improving AI a year closer
Jack Clark pulled his forecast for recursive self-improvement forward from 2028 to 2027, citing TasteVal, an eight-task benchmark from P-Zero Research measuring how well models design experiments and interpret results. Opus 5.5 posted a compute multiplier of 2.3, reaching “expert-level scores with less than half the compute.” P-Zero says research taste has “roughly doubled every three months since December 2025.” The authors supply their own brake, that results may overstate “how quickly research taste is improving,” and the tasks are cheap to check, which is what makes them easy to optimize toward.
Anthropic told developers to delete their subagents
Anthropic recommended companies consolidate sub-agents back into one primary agent that loads context on demand, for newer models starting with Opus 4.5 along with Opus 5.5 and Fable 5.1. One comparison ran the single-agent setup “about twice as fast,” and Anthropic cut the Claude Code system prompt for newer models by over 80 percent. It puts degradation at 30 to 50 tool definitions and coins “Frankenprompt” for the contradictory rules a system prompt accumulates. A cited Vals study found agent teams “can be faster than single agents but don’t deliver higher quality.” Two years of orchestration tooling existed to route around limits the models now handle themselves.
Block’s new voice model runs full-duplex on a laptop
Block released Dessa from its research lab, a model that talks and listens at once, which TechCrunch found unsettlingly human. Block says it was trained “fully on human speech and a single NVIDIA GPU rack” and runs locally on “a consumer-grade computer such as a MacBook M4 Max or better.” Its case against the standard pipeline is the interesting part: today’s voice AI “turns your voice into text and throws away an incredible amount of information text alone can’t carry.” Block claims emotion is built into the model, though whether it is steered by tags is unclear. There is no public access yet, and Block has asked safety researchers to test it first.
Quick Hits
- Peer review: a guest post on Terry Tao’s blog launched oai-math-revised, a community hub for revising OpenAI’s mathematical write-ups, on the grounds that they “can be difficult to understand and can have incomplete citations.”
- Exploitation window: the two unauthenticated PHP remote code execution flaws were being exploited about a week before fixes existed, with Amazon telling defenders to look back as early as October 17 and banks, hospitals and some government agencies pulling servers offline.
- Ad disclosure: Pennsylvania enacted one of the first state AI advertising-disclosure laws, requiring ads significantly modified or generated by AI to carry a label in the same medium, with carve-outs for parody and satire and no definition of “significantly modified.”
- Government access: a former Michigan CISO argued government portals should follow the banks and bar personal AI agents from transacting, on the grounds that agents already count as third parties under the account-sharing clauses almost every state and federal user policy carries. No government has enacted such a block; Citigroup and Bank of America have.
- Agent access: Carly connects to thousands of apps and can pull in the notes from your meeting notetaker, so follow-ups land without anyone retyping them.
- Staff bans: Sweden’s Government Offices cut off ChatGPT and Claude on work computers starting Monday, with digitalisation director Mikael Wettercrantz pointing to the information-security risk of AI services running without a contract in place; Microsoft Copilot survived an internal security review, but only after its settings were tightened.
- Leak investigations: the Wall Street Journal reported OpenAI’s security team asked three fired safety researchers to open their personal phones and named Ryan Greenblatt as the suspected leaker, who answered on X that “OpenAI has made false accusations about me.”
- Robot risk: Tokio Marine will be the first Japanese non-life insurer to cover AI-powered robots, with premiums “set to be 20% higher than regular industrial counterparts,” which is autonomy priced by someone who pays out when it fails.
- AI empires: Niall Ferguson, Eric Schmidt and co-authors argued in Foreign Affairs that OpenAI and Anthropic resemble modern colonial trading companies, calling raw-material and data-labor dependence “the plantation economics of the chatbot age.”
- Bloc hedging: the AP found countries torn between the US-led Pax Silica and China’s new AI cooperation organization, with Kazakhstan joining both and a former diplomat there saying “Everyone is hedging… We don’t want this binary choice.”
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
