Models & Labs

★Top story: Deals

Anthropic's IPO prospectus discloses billions in losses and an AI risk warning

Anthropic's prospectus for its planned US listing shows a 2025 operating loss of more than $8 billion against revenue that grew twelvefold to nearly $4.6 billion, with the company targeting a valuation above $2 trillion. Reuters and the Financial Times reported that the filing is the first on the SEC's database to list existential risk to humanity as a disclosed risk factor.

$8B+Anthropic's 2025 operating loss

NoteA company does not usually put existential risk in the same document as its revenue growth, so seeing Anthropic name it as a risk factor is as much a signal to the banks building on its models as it is to the investors pricing a valuation above $2 trillion.

Source: TechCrunch·Share

Banking & Lending

Capital One says agentic AI lives or dies on data plumbing

Capital One vice president of enterprise AI Rashmi Shetty said the bank treats agentic AI as an end-to-end system, not a model choice, leaning on governed data pipelines, sandbox testing, observability of agent actions and human review for high-risk steps. The bank's Chat Concierge multi-agent tool has been in production for more than two years helping customers shop for cars.

NoteThe lesson from a bank that has run agents in production the longest: the hard part is not picking a model but building the data lineage and guardrails around it, which is the work still missing at banks racing to bolt agents onto legacy pipelines.

Source: Banking Dive·Share

Models & Labs

Anthropic ships a cheaper Sonnet that nearly matches Opus

Anthropic released Claude Sonnet 5.5, which it says runs more than 30% faster and costs up to 30% less per task than Sonnet 5, at the same $2-per-million-input, $10-per-million-output pricing. On benchmarks such as GDPval-AA and OSWorld 2.1, Sonnet 5.5 scores within a few points of the larger Opus 5.5 model.

NoteBanks and fintechs running Claude at scale for coding, document work and agents get most of a flagship model's capability at mid-tier pricing, which lowers the cost floor for the agentic pilots already spreading across back-office and customer-service work.

Sources: TechCrunch, The Decoder·Share

Payments & Fintech

Shopify opens checkout to browser-based AI shopping agents

Shopify extended its WebMCP standard with three new tools, get_checkout, update_checkout and complete_checkout, letting an AI agent read a cart, change details like address or delivery, and submit a purchase with the buyer's authorization instead of scraping the page. The rollout covers all eligible Shopify merchants, with Meta's Muse and Instinct AI named as early agent partners.

NoteThis gives agents a structured door into checkout instead of a screen to fake being a shopper on. It moves the agentic commerce question from whether agents can buy to who is liable when one gets the order wrong, the gap last week's bank consortium principles were trying to close.

Source: TechCrunch·Share

Deals

Goldman sees Big Tech AI spending topping Wall Street's own estimates

Goldman Sachs strategists led by Ryan Hammond forecast Amazon, Alphabet, Microsoft, Oracle and Meta will lift AI infrastructure spending 54% to $1.2 trillion in 2027, above the Street's $1.1 trillion consensus, after an estimated $800 billion this year. The bank says the outlay would be the largest technology investment cycle relative to GDP since 19th century railroad building, and that the five companies would need roughly $300 billion a year in AI revenue to break even on it.

$1.2TGoldman's 2027 hyperscaler AI capex forecast

NoteGoldman is both a forecaster and a lender into this buildout, so a number this far above consensus is a signal the debt and equity backing hyperscaler capex has room to grow further before anyone calls it overextended.

Source: PYMNTS.com·Share

Banking & Lending

Nvidia launches agent guardrails with JPMorgan and Citi on board

Nvidia introduced the Open Agent Safety Platform, pairing runtime software that sets boundaries for AI agents with a hardware watchdog on its BlueField-4 chips that can quarantine one in milliseconds. JPMorganChase and Citi are among more than 100 organizations working with Nvidia on the open-source project. The software is available on GitHub.

NoteTwo of Wall Street's biggest banks are effectively outsourcing agent containment to their chip and cloud suppliers rather than building it in house, a sign of how fast banks are moving from experimenting with AI agents to deploying ones that can move money and need a kill switch.

Source: GlobeNewswire (Nvidia)·Share

Models & Labs

Anthropic and OpenAI cut frontier model prices on the same day

Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $25 output on the previous Opus, and called it the strongest model it has tested. OpenAI answered with GPT-6 Sol and Luna at half the price of the models they replace, and says Sol makes about half as many mistakes as its predecessor.

50%OpenAI's price cut on GPT-6 Sol and Luna

NoteFor a bank's AI budget, price matters more than any benchmark. Statement analysis, KYC checks and document reading are high-volume jobs, and every cut like this moves more of them past the ROI line.

Sources: TechCrunch (Opus 5.5), TechCrunch (GPT-6 Sol, Luna), Tech in Asia·Share