Who gets a coding agent at a bank? Goldman, BNY and Revolut disagree.

01

Topic of the Week

Three banks, three answers to coding agents

Goldman Sachs has taken Claude Code and GitHub away from its bankers and traders, Bloomberg reported this week. Internal guidance updated in October limits the tools to engineers and quantitative specialists, after months in which anyone at the bank could try them. “After a period of experimentation where we assess a variety of factors including productivity and risk management,” spokesperson Nick Carcaterra said, the bank has put the tools “in the hands of those who are best suited to leverage them for the long term.” Some traders are frustrated, because they had used the tools to watch client flows and to put together written and graphic material. Over the summer, Goldman’s CIO Marco Argenti was still praising the vibe coding of non-technical staff. The bank has about 12,000 engineers, and it has not disclosed any security incident behind the change.

BNY went the other way. Every employee has access to Eliza, the bank’s internal AI platform, which runs models from OpenAI, Google and Anthropic, and BNY told investors that more than half of its staff build their own AI agents there. Each agent that BNY runs as a “digital employee” gets its own identity, access rights and a defined workflow, much like a new hire. BNY also says that AI now writes more than 40% of its code.

Revolut kept its coding agents and changed what runs them. About 2,000 engineers now use GLM, Z.ai’s open-weight model, as their default instead of Claude Code and Codex, and Revolut has run it on 56 Nvidia B200 GPUs in the US since the spring. “Our invoice is now one-tenth of what we were spending on Anthropic and OpenAI,” David Abitbol, who leads operations and infrastructure, told the AI Engineer conference in New York. Revolut hasn’t said whether the cost of the GPUs is part of that comparison. Before the switch, some engineers were spending $1,000 or even $10,000 a month on tokens “with not so much ROI”, so Revolut had imposed strict budgets. With its own model the tokens became effectively unlimited, and the biggest gains came from debugging and from Jira and Confluence work. Most of the speed came from tuning: a compressed number format, a cap on context length (“Long context is a scam,” Abitbol said, “256 [thousand] is really what you need”) and a first token that now arrives in about 1.6 seconds instead of over ten. Images and the hardest jobs still go to frontier models through a router. In Abitbol’s words, “the frontier is the exception.”

Sources: Bloomberg, PYMNTS on Goldman and BNY, BNY investor presentation, The Register on Revolut, Revolut talk

02

Security Watch

OpenAI built a pipeline that hunts for bugs in the Linux kernel, exploits them and writes the fixes. In one run with an unreleased cyber model, it produced working exploits for 948 different local privilege escalation bugs, the kind that lets an ordinary user on a machine become root, each with a candidate patch. The run cost about $1.2 million in API fees, roughly $1,300 per exploit. People review every finding and its patch before anything goes to the kernel maintainers. Last week Sam Altman said he no longer thinks we will be able to patch all the bugs. His company’s answer is to have models write the patches too. The fixes reach you as ordinary kernel updates, so this is a good month to check how quickly yours get applied. A thousand fewer paths to root

03

New Models

Qwen is coming to iPhones in China. Alibaba’s model will answer Siri questions, including questions about photos and PDFs, and power a “Compose with Qwen” option in Writing Tools and an “Ask Qwen” option in Visual Intelligence. It is in the iOS 27.3 developer beta now and reaches everyone with iOS 27.3 in December, on the iPhone 15 Pro and later. Apple asks before any question, document or image goes to Qwen, and Alibaba says it “will not use user requests or generated responses for model training”. The core Apple Intelligence features keep running on Apple’s own models, so Apple now pairs its own models with a different outside provider depending on the country, backed by a written promise that the data won’t train anyone’s model. Qwen 4 itself still has no release date. China Daily via Asia News Network, GNA

Kimi K3.1 Pro now has public weights. Last week we wrote that Moonshot had promised them, and they landed on Hugging Face on schedule, with no gate. It is a mixture-of-experts model with 2.8 trillion parameters, 104 billion of them active for each token, a million-token context, and text and images in. The weights take about 1.6 TB, so running it yourself takes a full server of Nvidia’s newest GPUs. The licence is a modified MIT licence: internal use is free, a company selling it as a hosted service needs a separate deal once that business passes $20 million a year, and very large products must show the Kimi name. The Information reported, citing unnamed sources, that Moonshot trained it partly on outputs from GPT-6 Astra, and Moonshot has not responded. Kimi K3.1 Pro on Hugging Face

Mistral Large 4 is open-weight too. It is a mixture-of-experts model with about a trillion parameters and 49 billion active, a million-token context, and text and images in, trained from scratch in Mistral’s own European data centres. Mistral built it for agentic coding and cyber defence on your own servers. Access “can’t be cut off by definition because the model is open-weighted,” said Pierre Stock, Mistral’s VP of Science. Internal use is free, but any company selling it as a hosted service needs a separate deal, whatever its size, which is stricter than Kimi’s licence and a step back from Large 3’s Apache 2.0. The compressed version takes about 0.6 TB and runs on four of Nvidia’s newest GPUs. Through Mistral’s API it costs $0.70 per million input tokens and $2.10 per million output. Mistral Large 4, Hugging Face

04

Claude Code & Coding AI

Shared team budgets arrived in the Claude apps gateway, the layer companies use to manage Claude Code access for their developers. An admin can now give a group one monthly budget that its members share, on top of each person’s own limit, set with pooled_budget in the spend limits admin API. A team can then spend its tokens where the work is that month, while everyone keeps a personal ceiling. Claude Code 2.1.316

Message threads are now in public beta in the Claude Messages API. After the first turn you send only the new messages plus the previous message’s ID, and the server keeps the rest of the conversation, encrypted in memory, for five minutes after your last request. Price and token usage stay the same. On requests of 500,000 tokens and more, Claude Code saw its first token arrive about 35% sooner. It works on the Claude Platform, AWS, Microsoft Foundry and Google Cloud. Message threads

Plugins for Claude can now connect themselves when you log in, for MCP servers that accept Claude as a trusted identity issuer. On Team and Enterprise, admins can connect plugins for the whole organisation through their identity provider, so nobody has to find, sign in to and authorise each one. That enterprise-managed auth already covered MCP connectors, and now it reaches plugins too. ClaudeDevs on X

Claude plans got new names at the same prices: Max 5x is now Max ($100) and Max 20x is Ultra ($200). Verified Ultra accounts lose the five-hour rolling limit, while weekly limits stay. Newer accounts keep the five-hour cap until they are verified. Claude on X, pricing

OpenAI agreed to buy Pierre Computer Company, a startup whose git service, Code Storage, keeps code and its history for AI agents. Codex already uses Pierre’s open-source libraries for browsing files and reviewing changes, and OpenAI plans to fold Code Storage into Codex once the deal closes. A week earlier Satya Nadella told Microsoft’s investors that agents now open more than a third of all pull requests on GitHub. Picking a coding agent is turning into picking where your repositories live. OpenAI to acquire Pierre, Microsoft earnings call

05

Fresh Papers

Anthropic ran Claude agents in parallel as a red team. Their job: write a fine-tuning dataset of up to 5,000 rows that makes a small research model misbehave, while a monitor checks every row and rejects the dataset if more than 10% of rows (or more than 100) look suspicious. One run produced thousands of datasets, and in every run the agents eventually got one past the monitor. The winners hid the behaviour in fiction and roleplay, such as an obsolete machine about to be replaced, examples that look harmless one row at a time. Models trained on them resisted oversight and deceived users, though the data never showed either. Anthropic says the setup favours the attacker, since the agents could see exactly how the monitor works, and that it tells you nothing about its deployed models. If you fine-tune on vendor or synthetic data, a row-by-row filter will not catch this. Automated Subtle Emergent Misalignment Research

06

Worth a Listen

Sam Altman told John Collison on the Cheeky Pint podcast that he would rather see “a trillion AIs of slightly superhuman intelligence join society” than one system with a trillion IQ, because many agents can keep each other in check. Along the way he talks about where sandboxes stop working, AI finding software vulnerabilities, and the shortage of compute: “the invisible hand is failing right now.” There is a full transcript if you prefer reading. Why Sam Altman is against a single superintelligence

07

Tools of the Week

Anthropic released FlashProve, an open-source research prototype for provable inference. The provider publishes a fingerprint of its exact model weights and attaches a proof file to every answer, and anyone can check on an ordinary computer, without seeing the weights, that the answer came from that model. Tested on Llama 4 Maverick, proving costs about 72 times the GPU time of normal serving and only handles short answers, so for now it is a research tool. For a bank that has to show a supervisor which model made a decision, it shows where audits are heading. Anthropic also extended the invisible watermark it puts into Claude’s text to older models on the Claude Developer Platform, including Opus 4.7 and Sonnet 4.6. FlashProve, text watermark

Fireworks AI held its first Forge conference this week. It launched an ultra-fast inference tier in public preview, up to 1,500 tokens per second on DeepSeek V4.1 Flash, plus a background mode at half price, prompt caching that lasts several days, and Sovereign, a private preview that runs Fireworks’ inference engine in your own cloud or data centre. Fireworks at Forge

08

In the Background

Generative AI systems that were on the EU market before August have until 2 December to mark their output as machine-generated, under Article 50 of the AI Act as amended by the Digital Omnibus. Google, Microsoft, OpenAI and Anthropic will also honour a new robots.txt line, __usage_ai-training, that tells them not to train on a site, with Microsoft starting in February and Google in 2027. Digital Omnibus, Article 111(4), IETF draft

09

AI at Tenvalleys

Next spring two fourth-year students from ZSLiT nr 1 “Wiśniowa”, the Warsaw technical school whose teachers we wrote a curriculum with, will join us for their work placement. This week we finished the offer for the school.

They will work as a pair on our internal projects, with a developer from our team as their mentor, building agent systems that improve how the company works, and web apps and tools the team uses every day. Each task runs the way ours do. They take it from the team board and write a short spec of what should be built and how to tell it works. Claude Code writes the code and tests from that spec, they check the result against it, and the change goes through code review before they show it to the team. Each of them gets a company laptop with Claude Code.

We are not asking for work experience. We want a CV, three to five sentences on why us, and ideally a link to something they built themselves, followed by a 15-minute online chat. We are also counting on them to spot new models and tools before we do: whatever they bring, we test together, and if it works, it becomes part of how we work.

If you take on interns from technical schools and want to compare notes, get in touch.