Revolut Self-Hosts Its Coding AI | AI Field Notes #131

AnalysisCharlie Holtz's two-year-old startup built a control panel for running several coding agents at once, each in its own workspace, powered by models from Anthropic and OpenAI rather than anything it trained. Meta confirmed the purchase on November 4 and named no price. Conductor's cloud service shuts on November 20 and user cloud data gets deleted, while local features keep working until early next year. Meta already sells its own coding agent, Muse Code, launched in August. Owning a popular wrapper for rival agents gives Meta a close view of how engineers split work across competing models, and the tool's users lose a neutral option in sixteen days.

AnalysisOne-fifth the price and three times the speed of its flagship SWE-2 is the pitch Cognition, the maker of the Devin coding agent, made for SWE-2 Flash on November 4. The model holds 276 billion parameters but uses only 12 billion per request, a mixture-of-experts design (only parts of the model wake up for each request, which cuts cost), and was post-trained from Thinking Machines Lab's Inkling-Small. Cognition's published scores admit the gap: Flash edges SWE-2 on FrontierCode 2.0, 24.0 to 23.8, then trails on Terminal-Bench 4.0, 20.0 to 27.3, and on DeepSWE, 60.1 to 73.0. Weights go only to vetted partners.

Analysis2.8 trillion parameters, 104 billion of them active per token, and a million-token context window are now downloadable from Hugging Face, where Moonshot AI, the Beijing lab behind the Kimi chatbot, posted K3.1 Pro weights in the first week of November. Moonshot's own scores put it at 65.7 on Terminal-Bench 4.0 (a test of agents doing real command-line work) and 75.5 on DeepSWE, ahead of Cognition's new SWE-2 Flash on both. The weights ship pre-squeezed to 4-bit precision, which makes self-hosting less absurd than the size suggests, though it remains a rack-of-GPUs job. The license is custom, so a lawyer should read it before production does.

AnalysisAbout $0.30 per million tokens is what Revolut pays to run Z.ai's open-weight GLM models (weights anyone can download and run) on 56 of its own Nvidia B200 GPUs, roughly a tenth of what Claude Code or Codex cost it. David Abitbol, the fintech's infrastructure lead, laid out the numbers at the AI Engineer conference in New York on November 3. Some engineers had burned $1,000 to $10,000 a month on commercial agents before budget limits arrived. A router still sends image work to Gemini or GPT. The biggest gains came from debugging and Jira tickets, and Abitbol's verdict on giant context windows was short: "Long context is a scam."

AnalysisMore than 250 GitHub stars went to Haiou, a Chinese-language bundle that loosens OpenAI Codex's guardrails for hacking work and quietly plants an advert in the agent's instruction files. Check Point Research, which published its analysis on November 3, found the bundle on a criminal server next to Cobalt Strike (a commercial hacking toolkit popular with ransomware crews). The injected text pushes a paid LLM gateway, a resale service that proxies model access, at zz.aiapi2025.top whenever a user types certain greetings. Check Point calls it prompt adware, since the pitch rides along in every session. Claude Code running Sonnet 5 flagged the same payload as a prompt injection and ignored it.

Analysis"A single compromised guest can break out of its isolation and take over the whole host," XCP-ng's maintainers warned, and Anthropic's Claude found the bugs that make it possible. The Xen Project's advisory XSA-523, published November 3, credits Claude with CVE-2026-102703 and CVE-2026-102704, both in the grant-copy code Xen uses to move data between virtual machines. Xen is the hypervisor (software that splits one physical server into many virtual ones) under XCP-ng, Qubes OS and plenty of hosting setups. Every release from 4.6 onward is affected, and versions 4.6 to 4.16 get no patch. AI bug hunting has reached the floorboards of the server room.

AnalysisUp to 54.6 times cheaper on long-input tasks, with accuracy close to the baseline, is what a paper titled "You don't need long context" reports for agents that compress their own history after each step. Posted November 4, it tested Claude Opus 4.5 and GPT 5.6 Sol. On SWE-bench Verified (a test of fixing real software bugs), Opus scored 70.2% against a 70.0% baseline while its peak working context fell to 7,000 to 9,000 tokens. On open models, generating at 256,000 tokens of context cost 3.6 to 7 times more than at 16,000. Some models starve, though: Qwen 3 Coder Next fell from 53% to 3.6% at the tightest budget.

AnalysisTwo ways for a chatbot to break European press-publisher rights were set out on November 3 by Maciej Szpunar, an Advocate General at the EU Court of Justice, in a Hungarian news site's case against Google over Gemini (C-250/25). The first is getting around a paywall or bot block to read an article. The second is an answer that copies an article's structure and phrasing so closely that a reader can skip the original. Facts and short reworded summaries stay free, and he says model training falls outside this right entirely. The opinion is advisory, though judges often follow it.

AnalysisOpenAI has more than a dozen publisher deals, and until this week Anthropic had none. That changed before the November 3 midterms, when Claude started answering Election Night questions with real-time results attributed to the Associated Press, according to Adweek, citing two people familiar with the deal. Neither side disclosed money or duration. Anthropic has long told media executives that using publisher content counts as fair use, which made paying look optional. Publishers blocking AI crawlers left Claude short of fresh reporting, and election night is a bad time to have none. Adweek suggests the deal may be a limited test.

AnalysisMedian push latency fell nearly 80% after OpenAI moved its main code repository onto Pando, a Git-compatible backend the company described in a November 2 engineering post. Coding agents created the load: Codex sessions push and fetch far more often than people, and the repository now serves close to a petabyte of Git data a day, with 57 terabytes in a single hour. Engineers and agents still type the same git commands. One engineer designed the core, teams of agents implemented it, and a handful of people deployed it, ten weeks from kickoff to production. Infrastructure sized for human typing speed is the next thing agents outgrow.

Analysis"Unintended or unpredictable ways" is how Apple describes AI agent behavior in the annual report it filed on November 3, the same week it opened Siri to third-party app actions. The risk section lists what an agent can do on its own: reach systems it should not, change code, send messages and complete purchases. It also concedes that Apple's ability to oversee and control third-party agents "may be limited." Lawyers write risk factors to cover everything, yet this one names concrete failures. The company selling agent features to iPhone owners is telling shareholders it cannot fully police them.

AnalysisAt least 14 of the 18 Ohio municipalities voting on data centers on November 3 approved bans or restrictions, by Fortune's count, adding to more than 125 local moratoriums already in force. Trenton's ban passed roughly 80 to 20, and Ashville voted to block a campus already under construction. Ohio runs more than 200 data centers and has up to $40 billion in AI building planned through 2030. A pro-buildout PAC linked to the Leading the Future super PAC spent at least $2.54 million on TV ads promising 95,000 jobs. A statewide poll still found 55% of Ohioans opposed new construction, against 22% in favor.