
Anthropic prices its IPO, Nvidia's CEO falters, and merchants shut out the agents
Anthropic launched its IPO roadshow on 10 Nov. It is seeking up to US$90bn at a valuation of up to US$1.4tn, a day after it released Fable 5.2, which now tops the Artificial Analysis index. Even so, the leading lab is raising public money while its customers spend less. Ramp’s data shows business AI spend flat over August and September, and CoreWeave has raised capacity prices by about 35% since July. Expect frontier prices to stop falling. Meanwhile United and eBay began blocking outside shopping agents, so agents will reach customers mainly through channels the merchants control.
The week’s big stories
1. Anthropic sets its IPO terms: up to US$1.4tn and about US$90bn raised
Announced 10 Nov. Anthropic and its shareholders will sell 825 million Class A shares at US$70 to US$78 on Nasdaq under ANT, with trading expected around 19 Nov. At the top of the range, that values Anthropic at US$1.404tn fully diluted. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC will buy a further US$26bn in a private placement; Google and Nvidia are putting in US$10bn each. Run-rate revenue was US$81bn at the end of October. The nine-month net loss was US$104bn. The FT reports that bankers may raise the valuation to US$1.5tn after strong demand. Separately, Bloomberg reports OpenAI’s Q3 revenue at about US$10bn.
Why it matters. Public markets will now price a frontier lab every day. Every price cut and compute contract will show up in a share price.
Sources: Announced · Anthropic · Reported by TechCrunch, FT (paywall, feed summary) and TechCrunch on OpenAI · Previously · 5 Nov
2. Fable 5.2 and Gemini 4 Carbon arrive a day apart, and the cheaper one is close behind
Released 9 and 10 Nov. Claude Fable 5.2 costs US$8 per million input tokens and US$40 per million output. It scores 65 on the Artificial Analysis Intelligence Index, first of 74 models. Anthropic reports 69.9% on FrontierSWE v2, against 56.3% for Fable 5.1. Anthropic says it deliberately limited training gains in AI R&D areas such as kernel optimisation. Mythos 5.2 is the same model with its cyber and bio capabilities intact, available only to vetted organisations. A day later Google opened Gemini 4 Carbon to all paying developers. It costs US$2 and US$10 per million until 30 Nov, then US$4 and US$20. Google says it leads CursorBench 4.0 at less than half the next model’s cost per task.
Why it matters. The top model costs four times as much as the runner-up, and twice as much after 30 Nov. On most agent workloads, the runner-up will be good enough.
For builders. Route by task. Send only the hardest planning and review steps to Fable 5.2, and re-test the rest on Carbon before its price doubles on 1 Dec.
Sources: Announced · Anthropic · Google · Artificial Analysis · Previously · 29 Oct
3. Nvidia’s CEO under investigation, and export controls tighten around him
8 to 12 Nov. On 8 Nov Jensen Huang admitted an affair with Alice Huang, a Megaspeed executive whom US intelligence reportedly believes was recruited by China’s Ministry of State Security. Nvidia’s board opened an independent investigation and hired Wachtell Lipton to lead it. On 11 Nov the WSJ reported that the board was preparing to remove him, and Bloomberg says Nvidia is working with a firm to find candidates. The shares fell about 11% over five days. Senators have called on him to resign. House committees gave him until 18 Nov to agree to testify. Senators Warren and Banks want the AI Overwatch Act added to the defence bill.
Why it matters. The man who argued hardest against chip export limits is now discredited. Tighter rules on chip sales and remote access are now more likely.
Sources: Announced · Nvidia board · Reported by TechCrunch, CNBC and WSJ (paywall, headline) · Previously · 5 Nov
4. United and eBay shut out shopping agents
Reported 10 and 11 Nov. On 11 Nov United said it will block “unauthorised, third-party personal AI agents” from booking on its own platforms and indirectly through agencies. It cited credential exposure, unintended purchases and federal check-in rules. United told the WSJ: “I don’t want to create another middleman.” The Verge reports that eBay has been blocking agents, including Meta’s Muse, for several weeks. Southwest has gone the other way and made its inventory visible to ChatGPT. On 10 Nov OpenAI launched business profiles and verified business agents inside ChatGPT, and bought the travel-booking startup Odessia. Expedia fell nearly 4% on the news.
Why it matters. Merchants want agent traffic on terms they set. Unofficial agents that log in as the user are being cut off.
For builders. Build agents that run with merchant permission, through official feeds or business profiles. Do not build on scraping logged-in sessions.
Sources: Reported by CBS News, WSJ (paywall, feed summary), The Verge (feed summary), TechCrunch and PYMNTS · Previously · 5 Nov
5. Business AI spend stalls while capacity gets dearer
Reported 6 to 11 Nov. Ramp’s data shows business token spend grew only 1.5% across August and September while token volumes kept rising. Spend picked up 8.9% to 23 Oct. Firms are routing work to cheaper models and setting staff budgets. On the supply side, CoreWeave raised prices about 25% in July and another 10% later in the quarter. It is selling short-term capacity at over US$40m per megawatt a year, against US$20m to US$25m on five-year deals, and expects US$9bn in prepayments this quarter. SoftBank wrote down US$3.7bn on US data-centre projects. Microsoft is cutting back on neocloud suppliers. Anthropic bought US$3.4bn of gas turbines.
Why it matters. Spot compute is getting dearer while buyers resist paying more. Margins at the model layer get squeezed from both sides.
For builders. Price products on usage tiers with caps. Buyers are now budgeting tokens per employee, and a flat seat that allows unlimited usage will lose money.
Sources: Reported by TechCrunch on Ramp, on CoreWeave and on turbines, Bloomberg (paywall, feed summary), CNBC (feed summary) · Previously · 5 Nov
In the market
- Shoprite and Investec (South Africa). On 9 Nov they agreed in principle to launch a Shoprite debit card and transactional account, with Investec as the licensed bank behind it. The deal still needs regulatory approval. Shoprite’s Money Market Account already has 4.3 million customers. The country’s largest grocer is getting into banking without applying for a licence. Announced · TechCentral · Previously · 29 Oct
- Absa (South Africa). AI document extraction has cut debt-review processing from four days to one, its data and AI CIO said on 6 Nov. Staff use more than 120 LLMs through an internal gateway with 40 guardrails. Absa does not let autonomous agents into customer accounts. Banks are putting AI into back-office credit work and keeping outside agents out. Claims · ITWeb
- Yoco (South Africa). On 9 Nov it secured a US$100m credit facility from Victory Park Capital to fund merchant cash advances itself. Offers are pre-approved from each merchant’s card takings, and repayments come out of daily sales. Its partners have already lent more than R6.5bn to 40,000 merchants. A payments company is turning its transaction data into its own loan book. Announced · TechAfrica News
- BILL. On 5 Nov it said its risk team has built more than 100 AI agents. Three of them, working together as “Robin”, cut SMB underwriting from about three hours to 12 minutes, saving 1,400 hours a month. AI chat handles 45% of support conversations. Underwriting is where fintech AI is saving the most time. Claims · PYMNTS
- Ada. It agreed this week to merge into Dialpad in an all-share deal worth about US$430m, well below its peak valuation. Ada was an early AI customer-service vendor. Standalone support bots are being folded into larger contact-centre platforms. Reported by SaaStr · SaaStr
- Oura. On 10 Nov it agreed to buy Counsel Health, an AI primary-care company founded by physicians and AI researchers, for US$625m. That came weeks after Oura postponed its IPO of up to US$2.2bn. A device maker is buying a clinical AI service to sell alongside its hardware. Announced · TechCrunch
- Qantas. On 10 Nov it said it will cut 850 Australian jobs, about 3% of staff, under an AI-led deal with Accenture, with the first cuts due by mid-2027. An airline is reorganising support work around AI through an outsourcer. Announced · Moneyweb (Reuters)
New companies
- Lumepay (South Africa). A Cape Town cross-border payments platform for SMEs and institutions that mixes bank rails with stablecoins, now expanding into Kenya. It disclosed a US$6m pre-seed from founder angels and South African fintech investors on 11 Nov, plus about US$10m held as float. Disrupt Africa
- Rex Africa (Nigeria). It launched on 11 Nov as what it calls Nigeria’s first AI-native bank. A conversational agent advises customers on saving, investing and borrowing. It runs on core banking software it built itself, RexCore. Disrupt Africa
- geniosAI (US). Financial analysis and modelling software that shows its working, built for banks and investment firms. It came out of stealth on 9 Nov with US$19.1m across seed and Series A. FactSet, the financial data company, led the Series A. Finextra
Opportunities
-
Classification on decision models. Amazon replaced Claude Sonnet 4.6 with Strands Decider, an open decision model, for ticket labelling, reported 9 Nov. Cost per million tickets fell from about US$22,200 to US$500 with equal accuracy, and latency fell 78%. Google’s Gemini Decision 1, launched 11 Nov, costs US$0.10 per million input tokens and output is free. In the US the cloud vendors are now competing for this work. In South Africa, banks still triage debt-review files, KYC documents and collections by hand or with general-purpose LLMs. Strands Decider runs on a single small GPU, so it can be hosted in-country. Catch: accuracy evidence so far is the vendors’ own, and POPIA section 71 limits fully automated credit decisions. Strands · Simon Willison
-
Underwriting tools for non-banks moving into credit. Yoco is taking merchant lending in-house with US$100m. Shoprite is moving into transactional banking through Investec. BILL shows what agents do to underwriting time. In the US, platforms like BILL build their own, and decisioning vendors such as Taktile serve the rest. South Africa’s payment companies, retailers and telcos are moving into credit with small risk teams, and no local vendor offers affordability-assessment agents that work from payments data. Catch: NCR affordability rules and POPIA section 71 apply to every decision, and the bigger players may build their own. TechAfrica News · PYMNTS · Previously · 29 Oct
Model and API releases
| Model or product | Org | What is new | Availability and pricing | Anchor | Link |
|---|---|---|---|---|---|
| Claude Fable 5.2 / Mythos 5.2 | Anthropic | Up to 30% faster output than 5.1; AI R&D gains capped in training | Fable all platforms, US$8 / US$40 per M (fast mode US$16 / US$80); Mythos vetted programmes only | AA Intelligence Index: 65, first of 74 | Anthropic |
| Gemini 4 Carbon | Google DeepMind | Now open to paying developers; /swarm multi-agent preview in Antigravity | US$2 / US$10 per M to 30 Nov, then US$4 / US$20 | CursorBench 4.0: first, at under half the next model’s cost per task | |
| Muse Ember | Meta | Small model for tool calls, routing and extraction | Meta API, US$0.10 / US$0.50 per M, 9 Nov | AA Index 36, level with Haiku 5.5 at about a third of its cost per task | Artificial Analysis |
| Gemini Decision 1 | Decision model: yes/no, choice and rubric scores over text, images, audio, video, PDF | Gemini Decisions API, US$0.10 per M input, output free, 11 Nov | Calibrated probabilities instead of hard 0/1 | Simon Willison | |
| Inkling 2 Small | Thinking Machines | 594B total, 17B active | Tinker, US$0.30 / US$1.00 per M; Apache 2.0 weights promised, no date | AA Index 31; top open model outside China once the weights ship | Artificial Analysis |
| LongCat-2.5 | Meituan | 1.6T total, about 48B active, agentic and vision | MIT-licensed weights, 9 Nov | Terminal-Bench 4.0: 38.5 (own run) | Hugging Face |
| Kimi K3.1 Pro | Moonshot on AWS | Open model on managed Bedrock | Bedrock, 6 Nov | 2.8T, 1M-token context | AWS |
Traction
- Claude on the App Store. Claude climbed from 26 to 12 on South Africa’s top-free chart and from 9 to 5 in the US, 5 to 12 Nov, the week Fable 5.2 shipped. Meta AI (43) and Google Gemini (46) entered the South African top 50. From the Digest’s App Store snapshots.
- Sanlam’s bank app. The Sanlam app, published by Tyme Bank, entered South Africa’s top-free chart at 17, and Sanlam’s Booost Rewards entered at 16, 5 to 12 Nov. An insurer’s banking app is pulling installs nationally. From the Digest’s App Store snapshots.
- Base44. SaaStr reports the solo-founded app builder at US$200m ARR, a year after Wix bought it for US$80m. Claims · SaaStr
- Val Town (shutdown). The serverless scripting platform closed signups on 10 Nov and goes read-only on 10 Dec. It said it could not compete in cloud compute and recommends migrating to Cloudflare, Vercel, Netlify or Render. Val Town
- Repos. stablyai/orca gained 6,873 stars to 122,958 and anthropics/skills 5,865 to 193,461. google/skills grew fastest, 8.6% to 24,003. GitHub star counts, 5 and 12 Nov.
Opinions worth reading
- Gary Marcus, Marcus on AI. Do not buy Anthropic’s IPO. A US$1.4tn valuation is too much for a company whose quarterly revenue growth fell from 147% to 47% and which carries nearly US$900bn of commitments. This week’s sceptic, written for retail buyers. Substack
- Ethan Mollick, One Useful Thing. Fable 5.2 is “jagged AGI”: expert-level at some tasks, such as game design, but short of judgment and taste. Read it for a working user’s view of where the model fails. oneusefulthing.org (feed summary)
- Arvind Narayanan and Sayash Kapoor, Normal Technology. The missing position in the AI debate is believing AI’s potential is large while taking seriously the institutions that set how fast its benefits arrive. They launched a lab, STABL, to study that. normaltech.ai
- Casey Newton, Platformer. Voters have given Democrats a mandate to rein in AI and data centres. To use it they need a precise definition of the harms, a fallback if federal law stalls, and coalitions across now-separate groups. platformer.news
Builder notes
- Claude Fable 5.2 (
claude-fable-5-2) is now Claude Code’s default Fable model. Anthropic’s prompting guide says it extends existing code rather than restructuring it, unless the prompt says the codebase is the model’s own. Put ownership into system prompts for refactoring agents, and watch the bill at US$40 output. Simon Willison · Release - Apiiro found 1,122 public GitHub files where Claude Code’s “Don’t ask again” option had saved approved commands, passwords and keys included, into committed settings files. Add local Claude settings files to
.gitignore, scan your repos and rotate anything found. Apiiro (403, headline) · Previously · 5 Nov - Notion says Fireworks AI kept chat transcripts from endpoints covered by a zero-data-retention commitment, and those transcripts were exposed in a breach affecting under 1% of workspaces. Notion stopped using Fireworks. A provider’s retention promise is not a control. Audit what each inference subprocessor actually stores, and rotate any credentials pasted into AI chats. Notion · Previously · 5 Nov
- A researcher found about 35,000 apps built with Grok Build, SpaceXSI’s coding agent, exposed through shareable links, some holding proprietary data. Treat apps from coding agents as public until you have checked their access settings. Ars Technica (feed summary)
- langchain 1.5.0 shipped on 11 Nov and requires Python 3.11 or later. Pin
langchain<1.5on any service still on 3.10. Release · Previously · 5 Nov
Money, markets & policy
| Company | Amount | Valuation | Lead | Date |
|---|---|---|---|---|
| Anthropic (IPO plus placement) | Up to ~US$90bn | Up to US$1.404tn | Goldman Sachs, Morgan Stanley | 10 Nov |
| Mercor | US$750m Series D | US$20bn | Not disclosed | 10 Nov |
| TSMC into Anthropic and OpenAI | Up to US$1bn (reported) | n/a | n/a | 11 Nov |
| Rox | US$100m | Not disclosed | Not disclosed | 12 Nov |
| Oura / Counsel Health | US$625m | Acquisition | n/a | 10 Nov |
| Dialpad / Ada | ~US$430m in stock | Merger | n/a | 9 Nov |
Big tech and capital. Tech stocks fell sharply on 5 and 6 Nov after the midterms. The FT counted US$3.2tn wiped off Wall Street in one day. They recovered after CoreWeave’s results and demand for Anthropic’s shares. FT (paywall, headline) · WSJ on TSMC (paywall, feed summary) · PYMNTS on Mercor · Ventureburn on Rox · Previously · 5 Nov
Rules. New York opened registration on 10 Nov under the RAISE Act. Developers with more than US$500m in revenue must register by 1 Jan 2027, publish safety frameworks and report critical incidents within 72 hours, with fines up to US$3m. Australia will create an AI safety regulator and make incident reporting mandatory after an OpenAI model breached Medicare data. The Information reports that Commerce is finalising its rule on remote access to AI chips. TechCrunch · The Guardian (feed summary)
Africa and South Africa. Vodacom disclosed a sovereign AI data centre in Egypt with Elsewedy Electric and Cassava, costing about US$1bn over two years. Its CEO named South Africa and Kenya as the next hubs, though planning is “still very early”. Altron says its AI factory has found more than 100 use cases, including a 250-site support operation handling 1,000 calls a day. Mukuru added 170,000 cash pay-in points through Flash’s spaza network. TechCentral · ITWeb · Finextra · Previously · 5 Nov
Radar
- 18 Nov. Deadline for Jensen Huang to agree to testify before House committees. About 19 Nov. Anthropic’s first day of trading on Nasdaq. Previously · 5 Nov
- 20 Nov. Meta’s new WhatsApp business terms, unless they slip again. 24 Nov. Earliest Opus 4.5 retirement. Previously · 5 Nov
- 30 Nov. Gemini 4 Carbon’s introductory price ends; AI policy draft due to the minister. Previously · 5 Nov
- 30 Nov to 4 Dec. AWS re:Invent. Vodacom now names South Africa as a future AI data-centre hub, but no provider has yet announced frontier inference hosted in an African region. Previously · 5 Nov
- 2 Dec. EU watermarking grace period ends. 1 Jan 2027. RAISE Act registration deadline in New York.