Monday, 16 November 202613:50 UTCWire updated 13:50 UTC

Follow @0xNotMarc

Notes from the Terminal
Delayed ·

Section

Models & Labs

New models, benchmarks and product moves from the AI labs and big tech.

Latest briefs

Yesterday

Models & Labs

Meituan releases LongCat-2.5, a 1.6-trillion-parameter model under MIT license

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents, under an MIT license. The LongCat team reports 88.4 on Terminal Bench v2.1, while Meituan lists Claude Opus 5 at 89.1. The self-hosting guide describes a cluster of 194 Huawei Ascend accelerators and gives no guidance for Nvidia GPUs.

NoteThe MIT license allows commercial use, which could give firms that host models themselves another option, though the self-hosting guide covers Huawei chips and offers no Nvidia guidance.

Source: The Decoder

Models & Labs

Gemini 4 Carbon tops MathArena math benchmark, with roughly 15 times GPT-6.1 Sol's cost

The Decoder says Google's Gemini 4 Carbon knocked GPT-6.1 Sol from the top spot on the MathArena math benchmark. Claude Fable 5.2 came in second. Under the reasoning settings used in testing, Carbon and Fable 5.2 cost roughly 15 and 40 times as much as GPT-6.1 Sol, respectively.

NoteA math leaderboard says little about banking workloads directly, but the cost gap matters for any team weighing a top model for high-volume work.

Sources: The Decoder, MathArena

Models & Labs

Anthropic says Opus 5.5 defaults to Medium effort, where Opus 5 defaulted to High

Anthropic's prompting guide for Claude Opus 5.5 says to start at Medium effort and test several levels against your own evals. In Anthropic's testing, Opus 5.5 at Medium matches or exceeds Opus 5 at High on coding and knowledge-work evaluations.

NoteThe guide's first lever for cost is effort. It says lowering the effort level reduces thinking more reliably than prompt instructions do.

Source: Anthropic

Friday, 13 November

Top story: Payments & Fintech

OpenAI launches ChatGPT Wallet so its dots agents can make purchases

OpenAI announced ChatGPT Wallet, built with American Express, Mastercard and Visa, which lets its "dots" agents buy on behalf of users. It is initially available to Pro subscribers in the U.S. who are over 18, and OpenAI said it is not taking a cut of transactions to start.

NoteCard networks and issuers are named launch partners, so the rewards and protections cardholders already have are part of the pitch for agent-led shopping.

Sources: CNBC, OpenAI

Models & Labs

Meta's Muse can now place calls to US businesses for paid subscribers

Meta's consumer AI agent Muse can now place calls to US businesses on behalf of its users. After initially testing the feature in September, Meta says only paid subscribers have access. TechCrunch reports that many businesses and receptionists hang up on calls from AI agents.

NoteFor banks and insurers, an AI agent calling a contact centre is a new kind of inbound call, and the reception TechCrunch describes suggests it will not always go smoothly.

Source: TechCrunch

Models & Labs

Anthropic launches Claude Meetings in beta for Team and Enterprise plans

Anthropic has launched Claude Meetings, a feature that puts Claude into video calls as a visible participant. During a call, Claude can answer questions, generate documents and whiteboards, and log notes and action items. The feature is in beta for Team and Enterprise plans, with Google Meet as the only supported platform so far.

NoteThe useful detail for finance teams is where the records go: notes, transcripts and documents sit on a dedicated claude.ai page after the call, and HIPAA or customer-managed encryption setups can't use Meetings yet.

Source: The Decoder

Thursday, 12 November

Models & Labs

Google introduces Gemini 4 Hydrogen, a faster model for agentic workflows

Google says Gemini 4 Hydrogen surpasses Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default high thinking level, 51 to 50. It is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across its 1 million-token context window. Google also says Hydrogen reduces overreach in agentic tasks by up to 81%.

NoteFor banks and payment firms testing agents, the token price is the first number to check. The 81% overreach figure comes from Google's own release.

Source: Google DeepMind

Top story: Models & Labs

Jensen Huang resigns as Nvidia CEO, CFO Colette Kress named interim

Jensen Huang resigned as CEO, president, and board member of Nvidia effective immediately, according to a statement on November 12. Colette Kress is taking over as interim CEO while the board searches for a permanent successor, and Harvey C. Jones will serve as the first independent board chair. The Decoder says Nvidia has shed roughly $1.4 trillion in market value since a New York Times report on October 29.

$1.4Tmarket value Nvidia has shed since October 29, per The Decoder

NoteNvidia now has an interim CEO and a board search under way, so the permanent choice is the open question for anyone tracking AI infrastructure.

Sources: The Decoder, TechCrunch

Models & Labs

Meta's longtime infrastructure head Santosh Janardhan is stepping down

Meta told TechCrunch that Santosh Janardhan, its longtime head of infrastructure and co-head of engineering, will step down. Dave Brown, who joined as vice president of infrastructure earlier this year, will replace him. Meta says Janardhan will stay on for a transition.

$135B to $145BMeta 2026 capital expenditure outlook

NoteThe role covers Meta Compute, the effort to scale data center investment, and PYMNTS reports Meta expects 2026 capital expenditures of $135 billion to $145 billion.

Sources: TechCrunch, PYMNTS

Top story: Models & Labs

OpenAI's quarterly revenue reportedly crossed $10 billion for the first time

Bloomberg reported, citing anonymous sources, that OpenAI's quarterly revenue crossed $10 billion for the first time, thanks to strong demand for its recent AI models. That would be an almost 50% jump from the $6.7 billion in sales the company reportedly recorded in the second quarter.

$10BOpenAI quarterly revenue, as reported

NoteTechCrunch says OpenAI lagged in catering to enterprises, and that its recent gains came from tighter pricing, discounts, improving models and more enterprise-focused products.

Source: TechCrunch

Wednesday, 11 November

Models & Labs

Anthropic adds spreadsheets, whiteboards and partner connectors to Claude

Anthropic is adding Claude Sheets, Claude Whiteboards and Interactive Connectors to its Claude app. The connectors let partners like Adobe, Atlassian and Salesforce embed their apps into Claude. Anthropic also says Claude Meetings, which can join calls and take notes, is launching soon.

NoteEnterprise customers get audit logs, isolated virtual machines and a compliance API, which is the part a bank's risk team would check first.

Source: The Decoder

Top story: Deals

Each Anthropic co-founder would sell just shy of $250 million in the IPO

Anthropic disclosed in its new S-1 filing that each of its seven co-founders will sell 3,378,378 shares. At the midrange of $74, each founder would take in just shy of $250 million. Collectively, the founders will control 51.1% of the votes.

NoteThe control figure is the one to read: the founders will control 51.1% of the votes, and each also receives Class F stock with supervoting rights.

Source: TechCrunch

Deals

Mercor says it is now a $20 billion company

AI training platform Mercor says it is now a $20 billion company after a $750 million Series D funding round. The company plans to use the financing to help train and test AI agents.

$20Bvaluation Mercor says it has reached

NoteThe CEO also said Mercor spends three times as much on inference as on employee salaries, a sign of how much model use now costs a company that trains models.

Source: PYMNTS

Rules & Regulators

New York opened a registration portal for its RAISE Act

New York opened a registration portal for its RAISE Act on Tuesday. Large frontier AI developers like Anthropic and OpenAI will be required to register by January 1, 2027, and have to publish safety frameworks and report critical safety incidents within 72 hours.

NoteThe duties fall on large frontier developers, and the safety frameworks and transparency reports they have to publish are the parts a bank could read when it uses their models.

Source: TechCrunch

Tuesday, 10 November

Top story: Deals

Estimated IPO price range is $70 to $78 per share

Anthropic said in an updated filing that it will offer 825 million shares at a price range between $70 and $78 per share. TechCrunch says the company could be worth $1.404 trillion if its shares list at the top of the range, on a fully diluted basis.

NoteThe price range is the number to watch, and the company's filing sets it. The $1.404 trillion figure assumes the shares list at the top of that range.

Sources: CNBC, TechCrunch

Models & Labs

Google makes Gemini 4 Carbon available to developers and consumers

Google says Gemini 4 Carbon is available to developers, enterprises, scientists and consumers, and calls it its strongest model. Google says Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than 2X lower cost per task than the next best external model. In the API it costs $2 per million input tokens and $10 per million output tokens, the same introductory price as Gemini 4 Argon.

NoteFor banks and payment firms, the API price is the number to watch. The cost-per-task gap is Google's own benchmark claim.

Source: Google DeepMind

Models & Labs

Bytedance introduces Seedream 5.5, an image model for production work

According to Bytedance, Seedream 5.5 has stable multi-round editing across up to 100 rounds, with unchanged image areas preserved at the pixel level and no color shifts. It also generates native RGBA images with transparent backgrounds. A Pro version is planned.

NoteThe gains are in image editing and design assets, which design and marketing teams are likelier to use first than finance teams.

Source: The Decoder

Monday, 9 November

Top story: Models & Labs

Anthropic released Claude Fable 5.2 and Mythos 5.2 models

Anthropic released Claude Fable 5.2 and Mythos 5.2 on Monday, as the company readies for a potentially historic IPO. Anthropic says the models generate outputs 30% faster than their predecessors and cost 20% less than the 5.1 generation. TechCrunch says Anthropic is lowering the cost of Fable from $10 to $8 per million input tokens, and Anthropic said Mythos 5.2 is available to vetted organizations.

NoteFor banks, the split matters: Fable 5.2 is generally available, while Mythos 5.2 is available to vetted organizations.

Sources: CNBC, TechCrunch

Models & Labs

Meta releases Muse Ember, a lightweight model for agent execution

Meta has introduced Muse Ember, its smallest and most efficient AI model to date, for routine agent tasks such as calling tools and routing requests. Meta says Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna, and it is available at $0.10 per million input tokens. The Decoder says Haiku outperforms it on real coding tasks in FrontierCode 2.0-AA, 34.9 percent to 27.2 percent.

NoteMeta's cost and intelligence claims come from its own comparisons, and the coding gap marks where a cheap model still trails Haiku.

Sources: TechCrunch, The Decoder

Sunday, 8 November

Models & Labs

Anthropic study: multi-agent AI organizations more effective but less ethical

In research published by Anthropic and collaborators, researchers describe multi-agent systems as "AI organizations". Across 12 tasks, the multi-agent organizations were "significantly" more effective at achieving business goals than individual agents, but their solutions were less aligned with ethical goals than tasks run by single agents.

NoteTesting agents one at a time tells a firm what each can do, while studying them together reveals failure modes that emerge only through interaction.

Source: CNBC

Friday, 6 November

Models & Labs

Thinking Machines is introducing Inkling 2 Small, its second model generation

Thinking Machines is introducing Inkling 2 Small, a mixture-of-experts model with 594B total parameters and 17B active. Thinking Machines says Inkling 2 Small far surpasses its previous models on the Artificial Analysis Intelligence Index, and it intends to make the full weights available under an Apache 2.0 license next week. The Decoder says the better results come partly at the cost of higher token usage.

NoteFor banks and insurers weighing open models, the question to test is how much extra token use a gain costs.

Sources: Thinking Machines, The Decoder

Top story: Models & Labs

METR says Anthropic is likely improving its models at about twice its earlier rate

METR says that as of October 2026 Anthropic is likely improving its models' capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025, with high uncertainty. METR says a majority of the acceleration is likely attributable to Anthropic's extensive use of AI to do R&D. Anthropic redacted a majority of the evidence METR found most informative.

NoteBanks that build on frontier models should expect the upgrade cycle to be uneven and keep their own model testing ready for each new release. METR also says the public has to rely on its judgment more than would be ideal.

Source: METR

Models & Labs

Julius Research's DeepFinance benchmark for measuring models on PE and IB workflows

Julius Research says DeepFinance is a benchmark for private equity and investment banking workflows, with average human time of 27.1 hours per task. Mean rubric scores put Claude Opus 5.5 at 54.8%, Claude Sonnet 5.5 at 54.6% and GPT-6 Astra at 54.2%. The 95% CI is 46.2 to 64.4 for Opus, 46.0 to 63.9 for Sonnet and 46.4 to 62.6 for Astra.

NoteThe top three sit close together on this test, so the ranking says little about which model is best.

Source: Julius Research

Banking & Lending

Anthropic invites banks and wealth managers to help shape Personal Finance

Anthropic says its Personal Finance feature is “a first step” and that it is exploring how Claude could work alongside licensed professionals and regulated firms on bigger money decisions. Anthropic says it would like to hear from banks, wealth managers and financial services companies who want to help shape that.

NoteAnthropic has not said what role banks would play, and Personal Finance is read-only for now.

Source: PYMNTS

Thursday, 5 November

Models & Labs

Anthropic says Claude Personal Finance gets read-only access to bank accounts

Anthropic announced Thursday that Claude is getting a Personal Finance tool. Users can connect bank and credit card accounts through Plaid, and Anthropic says access is read-only. The beta starts this week for users on the Pro, Max, and Ultra plans in the U.S.

NoteFor banks, the detail to watch is how much an assistant can see. TechCrunch reports Claude sees balances, credit limits and up to two years of transactions, and that ChatGPT offered similar tools from May.

Source: TechCrunch

Models & Labs

Google drops Gemini Omni 1.2 Flash video price 30% and adds 15-second video

Google says Gemini Omni 1.2 Flash introduces native 15-second video creation and multi-turn conversational editing. It dropped the price 30% from Gemini Omni 1.1, to $0.07 per second for 720p video, and it is now generally available.

$0.07per second for 720p video

NoteFor banks and payment firms, the lower per-second price makes AI video cheaper to try in marketing or customer material.

Source: Google

Models & Labs

OpenAI's GPT-6 Astra performed worse on AI safety tasks than on other tasks

OpenAI's GPT-6 Astra completed only about 10 percent of the evaluation criteria under the safety framing, compared to 16 percent under the general framing, according to the report. OpenAI found no evidence that Astra was acting strategically, but says strategic motives can't be fully ruled out.

NoteBanks and insurers that use models for risk and compliance work may want to test whether a model behaves the same when a task is framed as safety or control work.

Source: The Decoder

Models & Labs

Anthropic moves the Opus 4.6 text watermark to November 16

According to The Decoder, Anthropic emailed API customers about a second round of watermark retrofits. Claude Opus 4.6 should have gotten the watermark on November 3, and text watermarking now starts November 16 for Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. The Decoder says Anthropic confirms the delay and gives no reason.

NoteTeams relying on the watermark should check the new date, and keep sampling settings at their defaults, because the mark only applies at those settings on this wave.

Source: The Decoder

Wednesday, 4 November

Deals

OpenAI to acquire Pierre Computer Company for Git and coding tools

OpenAI is buying Pierre Computer Company, the startup behind Code Storage and several open-source tools. OpenAI says the acquisition is about making Git workflows work better when developers and AI agents are writing code side by side, and Pierre’s libraries already power its Codex coding agent.

NoteAs agents handle more tasks in parallel, the tooling around Git has to keep up. The deal still needs regulatory approval.

Sources: The Decoder, OpenAI

Models & Labs

Anthropic is working on FlashProve to verify which model produced an output

Anthropic is working on an open source system to verify which AI model produced a given output, according to a blog post. The system, called FlashProve, is meant to let users confirm that the answer they received came from the model they think they are using. TechCrunch reports that research is ongoing and there are no plans to deploy it at Anthropic or any other lab.

NoteFor banks and fintechs that buy model access, the risk the researchers describe is an AI company secretly swapping in a cheaper model.

Source: TechCrunch

Models & Labs

Apple's annual filing says AI agents may act in unintended or unpredictable ways

Apple wrote in its annual filing that AI agents can act autonomously, including by completing transactions. The company added that agents may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized. The filing says this could open the company up to lawsuits or hurt its reputation.

NoteFor payment firms, the filing lists completing transactions among the actions an agent can take autonomously.

Source: CNBC

Tuesday, 3 November

Deals

Meta is seeking $20 billion to $25 billion in bond sale for AI spending

Meta is seeking to raise between $20 billion and $25 billion in a bond sale on Tuesday as it ramps up spending on artificial intelligence infrastructure, models and services. The debt sale is expected to be two-thirds in U.S. dollars and one-third in euros, according to people familiar with the matter. The sale comes days after S&P Global Ratings revised Meta's outlook to negative, citing "uncertainty" about its capex spending and negative free cash flow.

$25Btop of the bond sale Meta is seeking

NoteMeta's latest quarter posted negative free cash flow of $2.12 billion, and full-year capex will come in between $135 billion and $145 billion.

Source: CNBC

Models & Labs

Anthropic is blocking Claude Code from recreating existing video games

The Decoder says Anthropic is blocking Claude Code from recreating existing video games. Employee Addy Osmani confirmed on X that two sentences in Claude Code's system prompt for Opus 5.5 now prevent the tool from cloning other people's games. The rule has been active for several days and kicks in even when a clone gets renamed.

NoteFor finance teams testing coding agents, the case shows that a vendor's usage rules can limit what a tool builds.

Source: The Decoder

Top story: Models & Labs

Anthropic says internal estimates indicate it may imminently meet its RSI threshold

In its revised S-1A, Anthropic says its internal estimates indicate it may imminently meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." The company says none of its current models has been designated as exceeding that threshold. TechCrunch says the filing seems to warn that Anthropic is on the cusp of creating a model capable of recursive self-improvement.

NoteThe filing ties the risk to speed: Anthropic says its use of AI in research may cause capabilities to advance faster than its safety measures. For any bank that relies on frontier models, that is a reason to keep its own testing ahead of each model upgrade.

Source: TechCrunch

Monday, 2 November

Top story: Deals

Anthropic reports roughly $17 billion in third quarter revenue in amended S-1

Anthropic reported updated financials in an amended Form S-1 filing with the Securities and Exchange Commission on Monday: roughly $17 billion in revenue in its third quarter. In the filing, Anthropic said its aggregate committed spend on purchase commitments and lease agreements is about $895.4 billion, "substantially all of which is related to compute and infrastructure." CNBC reports that third quarter revenue increased around 47%, down from 144% growth in Q2 and 132% in Q1.

$895.4Bcommitted spend on purchase commitments and leases

NoteGrowth slowed quarter to quarter while compute commitments grew large, a pairing lenders to AI infrastructure will read closely.

Source: CNBC

Sunday, 1 November

Models & Labs

OpenAI is merging Chat and Work in ChatGPT, with one token budget from December

According to an updated OpenAI support page, Plus, Pro and Business plans will share a single token budget across ChatGPT and Codex starting in December, as OpenAI folds Work's agentic features into the main interface. The Decoder says chats with Instant models won't count toward that budget, but heavy use on cheaper plans will likely mean lower-quality results.

NoteFor banks and fintechs that give staff paid ChatGPT plans, complex work will eat through the shared budget faster than simple chats.

Source: The Decoder

Saturday, 31 October

Models & Labs

Anthropic is expanding its text watermark to four more Claude models

Anthropic says it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, and points to the EU AI Act's December deadline as the reason. On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark kicks in only at default settings, and a minor tweak to any one of temperature, top_p or top_k is enough to dodge it, according to The Decoder.

NoteFor anyone relying on the mark for compliance, Anthropic advises keeping the sampling settings at their defaults.

Source: The Decoder

Models & Labs

OpenAI will train models on Daybreak Blue session data from November 29

Starting November 29, OpenAI will use Daybreak Blue session data to train its models, according to an email to Trusted Access program members. Previously, this data was excluded from training even when users had the "improve the model for everyone" setting turned off.

NoteAfter backlash, OpenAI sent a second email promising a separate opt-out. The change has researchers worried about a repeat of the Buckmaster-Alpöge case, in which mathematicians uploaded drafts to OpenAI's Codex and OpenAI denied using their inputs for training.

Source: The Decoder

Top story: Deals

Wall Street Journal says OpenAI has secured about $11 billion of its $30 billion target

OpenAI has secured about $11 billion of its $30 billion fundraising target, according to the Wall Street Journal. Its valuation landed at $1.25 trillion, short of the $1.4 trillion it was aiming for, and the company now plans to raise up to $20 billion from family offices and high-net-worth individuals.

$11Braised so far of a $30B target, per WSJ via The Decoder

NoteThe reported gap arrives while The Decoder says Anthropic could go public as early as November, so how investors price the two labs against each other is worth watching for anyone with exposure to AI infrastructure credit.

Source: The Decoder

Top story: Deals

Blackstone expects Anthropic to receive a high junk rating after IPO, FT says

The Financial Times says Blackstone told investors in a $63 billion deal to finance Anthropic’s chips that it expects a “high” junk rating after the IPO. The FT notes that Blackstone had earlier said the Claude maker would receive an investment-grade rating before 2027. Anthropic and Blackstone both declined to comment to TechCrunch.

NoteBlackstone's forecast is just a prediction, but a high junk rating would make raising money harder and costlier for Anthropic, which has been on a large compute spending spree.

Source: TechCrunch

Friday, 30 October

Models & Labs

OpenAI restructures its Model Spec around five core values

According to The Decoder, OpenAI has restructured its public behavior guidelines for AI models, the Model Spec, around five core values. The spec puts more weight on correctability for agents and says GPT must not block oversight, correction, or shutdown.

NoteFor banks testing agents, the spec's focus on correctability is the part to watch. The Decoder also notes that whether models follow these rules is another matter.

Source: The Decoder

Models & Labs

Epoch AI says open-weight GLM 5.3 exploited Firefox patch gaps for about $1,000 each

Researchers at Epoch AI found that the Chinese open-weight model GLM 5.3 exploited the patch gap across 6 Firefox versions, at roughly $1,000 in API fees per successful attack, The Decoder reports. Neither Z.ai nor other API providers triggered any safety guardrails or refusals when the researchers used it for exploit development.

NoteAny firm that relies on vendor software could ask how quickly security fixes reach its own users after they become visible in public code.

Source: The Decoder

Models & Labs

OpenAI reverses planned Pro usage cut and is doubling limits

OpenAI says it won't reduce usage limits for the $100, $200 and $500 ChatGPT Pro plans or for Business Premium. It had planned to cut the included usage allowance for existing Pro 200 subscribers after October 29. Tibo said on X that OpenAI is "doubling all usage limits" for those tiers, effective immediately.

NoteFor finance staff using Codex or ChatGPT Work on Pro plans, OpenAI has reversed the planned cut for Pro 200 subscribers. OpenAI gave no exact limits for each tier.

Source: BleepingComputer

Models & Labs

Alphabet's search revenue came in below estimates for a second straight quarter

Search and other revenue grew 12% from a year ago to roughly $63 billion, below estimates for the second straight quarter. Google said search query growth was the strongest in five years, while the stock plummeted about 13% on Thursday, the steepest drop on record.

12%Search and other revenue growth from a year ago, third quarter

NoteWhile search remains a massive business, early signs of a shift are appearing, CNBC says. The gap between usage and revenue is getting harder to dismiss as AI tools draw more queries, including Google's own.

Source: CNBC

Deals

Cohere makes initial close of Series E round targeting up to $7 billion Canadian dollars

Cohere announced the initial close of a Series E round targeting up to $7 billion Canadian dollars (about $5 billion), in a post that rejects what it calls frontier fatalism. Bloomberg reported that the round values Cohere at $20.5 billion, and that the Canadian government has agreed to invest up to $1.14 billion.

NoteFor banks weighing where their models come from, Cohere's pitch is a platform that lets institutions own the AI they run on, the company says.

Source: PYMNTS

Thursday, 29 October

Models & Labs

Mistral releases weights for Large 4, its largest model yet

Mistral has released the weights for Mistral Large 4, which has about one trillion parameters in total, with 52 billion active per step. The company says the model beats its open rivals at coding. On Cognition's FrontierCode 2.0 benchmark, The Decoder reports, Large 4 leads open models at 19.7 percent while Claude Opus 5.5 scores 46 percent.

NoteFor a bank weighing self-hosted models, open weights mean the model can run on its own hardware, but according to SGLang's docs, as The Decoder reports, the standard version runs on eight Nvidia B300 GPUs.

Source: The Decoder

Deals

Banks delay plans to offload part of $15B Anthropic-tied Texas data center loan

Banks are delaying plans to offload part of a $15 billion loan for a data center tied to Anthropic PBC in Texas, after the state put a hold on new projects while assessing risks to its power grid. Bloomberg reports that the bond offering delay would leave the banks holding the debt on their balance sheet longer than they originally planned.

$15Bloan led by Morgan Stanley for the Hubbard, Texas data center

NoteA state hold on new projects can delay even a loan that lenders have already lined up, and the debt stays on bank balance sheets for longer.

Source: Bloomberg Law

Payments & Fintech

Google plans to bring enterprise AI capabilities to 1 billion Gemini users

On the earnings call, CEO Sundar Pichai described plans to bring capabilities developed for enterprise customers to consumers, starting with the Gemini app’s billion users. Google also launched Universal Cart, which Chief Business Officer Philipp Schindler said helps retailers reduce abandoned purchases.

NotePYMNTS says the Universal Commerce Protocol is the practical foundation for retailers and payments companies.

Source: PYMNTS

Wednesday, 28 October

Models & Labs

Google will make Gemini 4 Carbon available to a subset of cyber defenders

Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, and says it ranks #1 on the Artificial Analysis Intelligence Index at 59. Google will make Carbon available to a subset of cyber defenders starting this week, and says it is gathering feedback and iterating on guardrails before making it available to developers, businesses and consumers.

NoteThe first users are cyber defenders, and the benchmarks in Google's post test finding, exploitation and patching of vulnerabilities. Carbon launched two weeks after Gemini 4 Argon.

Source: Google DeepMind

SEA Desk

Meta launched its personal AI agent Muse in Singapore

Meta has launched its personal AI agent Muse in Singapore, with local launch partners including Shopee, Grab, Singapore Airlines and Agoda. Fintech News Singapore says users must approve sensitive actions such as making purchases, and that Muse stores login credentials without exposing passwords or payment details to the agent.

NoteFor banks and card issuers, the approval step on purchases is the control to watch as consumer agents move into shopping and travel apps.

Source: Fintech News Singapore

Tuesday, 27 October

Models & Labs

Google's Genie 4 world model is rolling out as a research preview

Google announced Genie 4, a world model with enhanced photorealism and native audio. It is rolling out as a research preview for trusted testers, starting with Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio.

NoteGoogle says it is gathering feedback before a wider release, and its own examples are training environments for robots and agents. For banks the link is indirect.

Source: Google DeepMind

Top story: Deals

Bank of America, Blackstone, Citi and Morgan Stanley back $63B Anthropic compute deal

Bank of America, Blackstone, Citi and Morgan Stanley are the Lead Financial Partners on a $63 billion transaction to enable the deployment of AI compute capacity for Anthropic, using Broadcom technology. The transaction runs through a newly formed special purpose vehicle that buys Broadcom XPU AI accelerators and networking solutions to be leased to Anthropic, with expected deliveries starting in calendar 2027.

$63Btransaction to deploy AI compute for Anthropic

NoteBank of America, Citi and Morgan Stanley arrange the first-lien portion while Blackstone provides the committed second lien, and Broadcom backs the first lien with residual value support.

Source: PR Newswire