Four hundred thousand tokens. That is where Uber now forces its coding agents to compact their context, even when the model underneath could happily hold a million. It's the least glamorous number published yesterday, and the one I'd tape to my monitor.
Yesterday was supposed to belong to Anthropic's first day on the Nasdaq, and in dollar terms it did. But read the two stories side by side and they tell you something the opening bell didn't: the money pouring in at the top of the AI stack is betting on token volume, while the most sophisticated buyers are working hard to send fewer tokens.
Uber's coding agents run on a 400K diet now, and the bill halved#
The backstory, per VentureBeat's write-up of Uber's October talk: the company burned through its 2026 AI budget in the first few months of the year. Weekly agentic requests grew 9.4x between February and August. Rather than ration engineers, Uber changed the defaults inside AIFX, the internal wrapper that installs and configures tools like Claude Code. Engineers don't get to pick these settings:
Context compaction kicks in at 400,000 tokens, regardless of the model's window. Uber's distinguished engineer Uday Kiran Medisetty says cache misses and input costs climb sharply at 700,000 to 800,000.
Reasoning effort defaults to medium, not high. Medisetty says recent models do better at medium and degrade somewhat at high.
No MCP tool definitions load at startup. Agents reach tools through CLIs wired to a central MCP gateway, because 100-plus preloaded tools would eat 50,000 to 70,000 tokens before the first prompt.
Sub-agents default to cheaper models, with a one-hour prompt cache for the orchestrator and five minutes for the helpers.
The result: overall AI spend stayed roughly flat from April to late October while requests kept climbing. Cost per session fell 52% from its June peak, and cost per 1,000 requests dropped almost 34% with the model held constant. Picking the right model for stale feature-flag cleanup alone cut that job's cost by 90%.
None of this is exotic. It's the stuff most of us know we should do and don't, because the defaults in our tools point the other way. Max reasoning feels safer. Loading every MCP server feels convenient. Uber's numbers say both are expensive habits, and in the reasoning case maybe not even better. I've been running high effort by default on nearly everything; I'm switching my own setup to medium this week and measuring whether I miss it.
Fireworks AI co-founder Chenyu Zhao made the same argument from the vendor side on IPO day, telling TechCrunch that most companies can skip frontier models unless a task genuinely needs the newest general intelligence, and should track cost per task rather than cost per token. He puts the open-model lag at three to six months. Take that with the salt it deserves: Fireworks is still answering for two breaches this fall, including one Uber says exposed about 150,000 driver records through a vendor TechCrunch identified as Fireworks. Cheap inference that leaks isn't cheap.
Wall Street priced Anthropic above Meta on day one#
Anthropic opened at $111.50, 33% above its $83 IPO price, according to TechCrunch and The Verge. The IPO price itself was already above the $70 to $78 range. The figures that matter:
More than $81 billion raised in the offering; about $220 billion in expected cash once a $26 billion private placement is included.
Opening price implies a market cap over $2 trillion, ahead of Meta, Tesla and Walmart.
About $900 billion in commitments, mostly compute, against Q3 revenue of $17 billion.
Each of the seven founders took home roughly $280 million at the IPO price.
TechCrunch's analysis of the circular financing is the piece to read. The private placement came from Nvidia, AMD, Microsoft, Broadcom, Micron, TSMC and Google, most of which also sell Anthropic compute. SemiAnalysis's Joey Brookhart estimates Anthropic needs around $280 billion in cumulative revenue over four years to cover its compute bill. That is the counterweight to Uber's story: Anthropic's valuation assumes enterprises keep consuming more tokens, while Uber just showed a large enterprise holding spend flat as usage grew 9x.
Those aren't contradictory, exactly. Cheaper per-task inference can grow total volume. But it means the growth has to come from many more tasks, not fatter ones.
Dario Amodei used the opening bell to restate that Anthropic will pace training and deployment regardless of business or political pressure:
“The most difficult safety work lies ahead of us.” — Dario Amodei, at the Nasdaq opening bell
Public shareholders now get a quarterly vote on how serious that is.
Gemini can drive your Mac today; Apple is still hiring for it#
Google added computer use to the Gemini app for macOS, per TechCrunch. It works across local files, native apps and the web in the background, shows a live preview, and lets you pause, stop or approve actions first. It's US-only and English-only for now, and requires a Google AI Plus, Pro or Ultra subscription. OpenAI, Anthropic and Meta already ship similar desktop agents, so this is Google catching up rather than leading.
The more interesting signal is Apple. A job listing spotted by Bloomberg seeks someone to lead an agent that “browses the web, works across apps, runs commands, writes and debugs code, and calls services.” Bloomberg also reports macOS 27.3 will warn users against giving third-party agents full computer access. Read those together: Apple is building the permission wall before its own agent exists, which is a very Apple way to make sure its agent is the one you trust with the keys. If you're shipping a Mac agent, expect that onboarding friction to get worse in 2027.
Quick hits#
Agent-era outages, or just old pipes? TechCrunch asked security experts about this month's Apple Pay, FAA and two-day Microsoft outages. Dave DeWalt blames a rush to patch AI-discovered vulnerabilities; Corsha's Anusha Iyer blames legacy systems colliding with modernization. Companies say none were AI attacks.
A Doritos bag, six patrol cars, and a lawsuit. Four Baltimore County students are suing over an Omnilert AI gun alert that officials cancelled within two minutes, yet police still arrived with guns drawn, Ars Technica reports. The suit claims over 99% of the district's alerts were false positives. The human-in-the-loop existed on paper and failed in practice.
Meta's Muse Charm may cost $300. That's per The Information, via The Verge. A hard sell the same week Meta staff reported Muse showing them messages, orders and balances that weren't theirs.
The cheapest token is the one you never send#
My bet: within a year, “context budget” and “effort level” become managed settings in every serious engineering org, the way cloud instance sizes did a decade ago. Uber got there because it ran out of money in spring. The rest of us can just copy the config. And if Anthropic's $2 trillion depends on nobody doing that, the public market will find out quickly.