AI Coding Agent Costs | AI Field Notes #95

AnalysisA federal judge in San Francisco found the Defense Department broke the law when it branded Anthropic a national-security supply-chain risk, ruling on August 27 that the label was retaliation dressed up as risk assessment. Judge Rita Lin wrote that the Pentagon punished the company for refusing to let Claude run autonomous weapons or mass surveillance of Americans, and for defending that stance publicly. She found violations of the First and Fifth Amendments and threw out the designation. A government can choose its vendors. It cannot blacklist one for disagreeing where people can hear.

AnalysisChatGPT stopped waiting to be asked. On August 26 OpenAI added webhook-triggered tasks to ChatGPT Work, so an agent can wake up when a new Gmail message lands, a Slack channel gets a post, or a GitHub pull request changes, then act without a human refreshing anything. Free users get three scheduled tasks; webhook triggers stay on paid tiers. Tasks can be shared so a colleague clones the setup with their own accounts. The quiet part is the attack surface: an agent that reads your inbox and acts on it is also an agent a crafted email can steer.

AnalysisPerplexity put its entire agent stack onto hardware you own. On August 26 it launched Portable Computer, running the planner, tool router, scheduler, and search index directly on Nvidia's DGX Spark, a desktop machine with 128GB of unified memory. Work done on the device burns no billing credits, and any trip to the cloud needs separate approval. It ships first on Linux for Pro and Max subscribers, with Windows due in September and no Mac plans. Moving serious agent work off the cloud attacks the one cost that scales with every task: the metered token.

AnalysisRun the same model two ways and the price splits. Databricks tested coding agents on its own multi-million-line codebase and found that Pi, an open-source harness (the software scaffolding that feeds a model context and runs its tools) built by a startup called Earendil, beat OpenAI's Codex by 5.5 points while using the same GPT model. Pi sent about a third of the context per turn and finished in fewer runs, and in some matchups the cost per task differed by more than two times at equal quality. The model is becoming the commodity. The wrapper is where the money leaks.

AnalysisMarvell posted a record quarter and the stock still dropped about 8%. On August 28 the chipmaker detailed an expanded deal to design custom silicon for Google, worth up to $120 billion in purchases through fiscal 2033, but told investors the money will not turn meaningful until fiscal 2029. Revenue hit $2.74 billion, up 37%, with data-center sales up 46%. None of that satisfied a market that wanted the Google windfall now. Custom chips are the hedge against paying Nvidia's margins, and even the winners of that shift are on a slow clock.

AnalysisGrok 4.6 showed up where developers already work. On August 27 xAI put its latest model into Microsoft Foundry, Microsoft's catalog for testing and deploying frontier models under enterprise controls, and into the Cursor editor and its own Grok Build tool, with double the usage free for the first week. The model carries a 500,000-token context window and adjustable reasoning effort from low to xhigh, aimed at long-running agents. Distribution is the play here: a model nobody can reach in their editor loses to one already sitting in the tool they open every morning.

AnalysisAustralian police charged two men, aged 21 and 23, over the Shai-Hulud worm, malware that spread itself through open-source software packages, stealing credentials and using them to infect the next package automatically. Investigators tie the pair to TeamPCP, the group behind the March compromise of widely used developer tools including the Trivy security scanner and the AI gateway LiteLLM. More than 1,000 organizations are believed hit. The two appeared in a Perth court on August 27 after police seized their devices. The supply chain that ships your dependencies also ships whatever climbed into them upstream.

AnalysisChina is building the money plumbing around brain implants before most countries have approved one. In August a large state-owned insurer, PICC Property and Casualty, launched the country's first commercial policy covering the surgery to implant an invasive brain-computer interface, a device wired directly into brain tissue to restore lost function. It follows the March approval of NEO, a hand-movement implant from Shanghai's Neuracle and Tsinghua University, which received a medical insurance code within 48 hours. Reimbursement, not just research, is how a technology reaches ordinary patients. Beijing is laying that track early.

AnalysisPhysical AI got a cheap entry point. Hugging Face and Pollen Robotics opened pre-orders on August 27 for Microduck, a 25-centimeter bipedal robot with 15 motors, a camera, a small lidar, and an articulated beak, for $399. The whole software stack is open source, and the duck learns new tricks through reinforcement learning, the trial-and-error method where a system improves by being rewarded for good moves. It can walk, pick things up, and get back up when it falls. Robotics research has run on five-figure arms; a sub-$400 platform widens who gets to try the hard problems.

AnalysisAndreessen Horowitz, a firm built on software's high margins, just raised $1.1 billion to buy into the physical layer of AI. Announced on August 28, the Machine Age Fund targets chips, memory, networking, storage, data centers, and robotics, an unusual turn for investors who spent a decade avoiding hardware's capital costs. The pitch is that every layer of the AI stack is hitting the limits of today's supply chain and physics, so the returns move to whoever rebuilds the plumbing. When the software crowd starts funding factories, it is conceding that compute, not code, is the current bottleneck.

AnalysisThe next round of chip tariffs may reach the finished machine. The Trump administration is weighing duties that extend past bare semiconductors to laptops, game consoles, and the servers that fill AI data centers, according to reporting on August 27, and the data-center exemptions granted in January could be scrapped. Commerce Secretary Howard Lutnick favors tying relief to how much chip manufacturing a company builds on US soil. Tech firms are lobbying hard, warning the duties would raise the cost of the exact hardware the AI buildout depends on. The bill for reshoring may land on the buildout it means to protect.

AnalysisA Wall Street Journal opinion piece criticizing Treasury Secretary Scott Bessent turned out to be written with AI, and the paper decided that was fine. Readers flagged the August column by investor Stanley Druckenmiller using Pangram, an AI-detection tool, and Druckenmiller confirmed it: of course he used AI. Editorial page editor Paul Gigot called AI a fact of modern life and said no policy was broken. The line between assisted and authored just moved inside one of the most prominent opinion pages in the country. A byline now means the name approved it, whoever or whatever drafted it.

AnalysisAbout 80 British actors signed an open letter this week asking the government to make a person's voice a protected part of their identity. The campaign, Save Our Voices Now, backed by Hugh Bonneville, Nicola Coughlan, and Matt Lucas, wants everyone in Britain to own their voice in law, with a petition to match. Founder Peter Caulfield notes that three seconds of audio is now enough to clone a voice and put words in it the speaker never said. Denmark is already reforming its rules to give people rights over their own face and voice. The technology to copy a voice arrived before the right to refuse.