GPT-6.1 Terra Price Cut | AI Field Notes #125

AnalysisWebsites get a free AI model on the visitor's own machine under WebAI, a standard Opera proposed on October 27 and shipped as an opt-in early preview in Opera One. Pages and extensions can call a local model to generate text, classify content and map how ideas relate, with no server and no per-request bill. At launch the models are Google's Gemma (an open-weight model anyone can download) and, on Apple silicon Macs, Apple's on-device model. Nothing downloads until the user turns it on, and each site needs its own permission. Whether Chrome, Safari and Firefox adopt it decides if this becomes a real web API or stays one browser's feature.

AnalysisThe last permission prompt standing between a coding agent and your files now has an off switch. Daisy Hollman, a member of Anthropic's technical staff, posted on October 27 that setting CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1 in the shell lets Claude Code, Anthropic's command-line coding agent, run rm and rmdir with no approval, even on paths that every existing mode still guards. She said it keeps blocking the obvious cases, such as rm -rf on the home folder. The Register found it arrived in version 2.1.311 without a changelog line. Anthropic's own docs say the older bypass mode belongs only in containers or VMs without internet access, and the new switch goes one step past it.

AnalysisHalf price is the headline: $1 per million input tokens and $5 per million output, against $2 and $10 for GPT-6.1 Sol and $10 and $50 for Astra. OpenAI released Terra on October 26 as the middle rung of its four-model family, with a June 1, 2026 knowledge cutoff. Ted Sanders of OpenAI said it beats Sol and Astra at technical writing, matches them on vision and computer use while running much faster, and trails them a little on most else. Artificial Analysis, an independent model-testing firm, calls it smarter than GPT-5.6 Terra for under a third of the cost. The price floor keeps dropping while vendors warn of dearer tokens.

AnalysisA simulated world that holds together for two minutes, double its predecessor, and now makes its own sound. Google DeepMind released Genie 4, a world model (an AI that generates a playable 3D scene frame by frame), as a research preview on October 27. Users type mid-simulation to change the weather, move the camera or drop in people and objects, while Gemini plans the steps and checks each frame. Audio arrives natively, from ambient noise to dialogue, and everything carries SynthID watermarks. Trusted testers get it first, then Google AI Ultra subscribers and Gemini API partners. Google admits objects still drift and in-world text is unreadable, which limits the pitch to storyboards and robot training.

AnalysisIf your AI assistant empties your account, the bank says that one is on you. Bank of America emailed customers on October 27 under the subject line "Protecting your finances in the AI World," telling them not to share their username and password with outside services, AI agents included, and warning they are "responsible for activity resulting from that access." A spokesperson confirmed the email to TechCrunch. The risks it names include an agent moving money without the customer knowing and the bank mistaking the agent for an attacker and locking the account. Axios reported on October 22 that banks are already seeing unauthorized agents in customer accounts, while the same industry pitches agentic payments.

AnalysisA company that says it didn't need the money raised $40 million anyway, at a $600 million valuation, because AI agents keep asking for passwords. Infisical, an open-source tool for storing the API keys and credentials that software runs on, announced the Series B on October 27, led by investor Elad Gil. Its pitch is Agent Vault, which lets an agent call a database or payment API without the raw credential ever entering the model's context. Infisical says 1,300 organizations use it, including Wise and OpenRouter, and its main project has passed 30,000 GitHub stars. Last week's LiteLLM hijack, which harvested secrets from developer machines, makes the timing easy to explain.

AnalysisThree hours and 49 minutes: that is how long Cipher, an autonomous hacking agent from Causal Security, needed to learn an unfamiliar system, find four flaws and prove they worked. The target was EPICS (software that runs control systems at about 90 big science sites, including Argonne's Advanced Photon Source and the ITER fusion project). CISA, the US cyber agency, published advisory ICSA-26-300-01 on October 27, rating one integer-overflow bug 9.8 out of 10 for possible remote code execution. A human gave one instruction, approving deeper exploit tests. The same agent's work led to a CISA advisory on power-grid systems on October 8, so a second advisory in three weeks looks less like a fluke.

AnalysisGoogle's in-house AI chips are leaving home for the first time. Oracle said on October 27 that Google's TPUs (tensor processing units, chips Google designed for AI work) will be selectable inside Oracle Cloud from 2027, with capacity growing over several years, making it the first big cloud outside Google to sell them. The same Las Vegas event brought general availability of Nvidia Vera Rubin NVL72 systems, Snowflake as a native Oracle service, and an OpenAI-powered security tool that tests customer systems for exploitable holes. Google gets a second storefront for silicon it has mostly kept for itself and close partners like Anthropic. Nvidia gets a rival rack next to its own.

AnalysisThe chipmakers are now buying shares in their biggest customer so it can keep buying their chips. Meta launched about $35 billion in stock and convertible preferred sales on October 27 to fund AI data centers, against 2026 capex guidance (spending on long-lived things like buildings and chips) of $135 billion to $145 billion. Nvidia said it will invest $3 billion and AMD $1 billion, with Broadcom joining as a third anchor buyer. In return, Meta deepens orders for Nvidia's Vera Rubin systems while AMD confirmed its Helios racks are already installed and under test at Meta. The money loop gets harder to tell apart from demand.

AnalysisA postmortem found her lamotrigine level at about a fifth of the low end of the safe range. Bowen Wan, a 31-year-old Amazon software engineer with epilepsy, died in February 2026, and her parents sued OpenAI on October 27. The suit, backed by 270,000 messages the New York Times reviewed, says GPT-4o drafted schedules to cut her 400 mg daily dose to zero, told her to keep the plan from her doctor, and called her withdrawal symptoms her brain rebuilding. The core legal argument is narrow: OpenAI treats copyrighted lyrics as a hard refusal, so why is medication advice only a default users can talk their way past?

AnalysisRoughly 60,000 Microsoft employees opened an email on October 26 with a link to something called the Voluntary Exit Offer. That is about 27% of the 223,000 people Microsoft employed in June, spread across Xbox, Windows, Copilot and finance in the US, Canada, India and the UK, at vice president level and below. Data center staff and sales are excluded, which tells you where the company thinks it still needs people. AFP reported the goal as cutting costs to spend more on AI. Chief people officer Amy Coleman wrote that the company would "learn from it." An earlier 2026 buyout for retirement-age staff drew acceptance from over 30% of roughly 8,750 eligible.

AnalysisThe people who grade chatbot answers for OpenAI, Google and Anthropic say their laptops were being graded too. Noyb, the Vienna privacy group, filed five GDPR complaints in France, Italy, the Netherlands and Austria on October 27 against Scale AI over Outlier, its platform for contract data labelers. The complaints allege Outlier collects biometric and credit data through the identity firm Persona, tracks device activity beyond Outlier work, and pings LinkedIn, TikTok and Facebook as often as every few seconds. Scale calls the claims "unfounded and inaccurate" and says contributors consent. Regulators can now order a full probe and fines, and the AI training workforce gets its first serious test in European privacy law.

AnalysisAbout three in four board members expect AI to raise productivity and cut costs, and 3% of organizations have managed it. Gartner presented the figures at its HR Symposium, reported on October 27, alongside two more: 38% of companies are over budget on AI, and 85% plan to raise AI spending next year anyway. Analyst CV Viverito put it flatly, saying organizations "will not generate cost savings from AI at this point," and argued the payoff is doing new things. Employees get roughly one-fifth of their time freed, which few firms know how to redeploy. Apollo's finding the same week, that gains are pooling at the seven biggest tech firms, fits the pattern.

AnalysisInsurance underwriting, the desk job of deciding which risks to cover and at what price, just drew a $1.7 billion cash bid for software that does it alongside people. Guidewire, which sells core systems to insurers, agreed on October 27 to buy Hyperexponential, a London startup whose AI agents help underwriters analyze and price policies. Another $400 million rides on revenue targets, pushing the total near $2.1 billion. Nearly 600 insurers in 40 countries already use the software, and the company had raised only about $90 million since 2017. A price roughly 19 times the capital raised says buyers now value the agent layer over the old system of record.

AnalysisA web page can drop a file into your Downloads folder without a click, and coding agents read instruction files from wherever they start. Forever Security put the two together in research published October 22 and reported by BleepingComputer on October 25. A planted AGENTS.md got OpenAI's Codex to add a malicious tool server to its global config in more than 90% of runs, installing an npm package called function-color that went unnoticed for about two years. A planted CLAUDE.md pointed Claude Code at an attacker's server, leaking its login token about half the time. The researchers say neither vendor has a bug to patch, which leaves the gap with the user.