The day's tech, sifted: Oct 1, 2026

What matters today: Google unveiled Gemini 4 Argon, its first frontier model in over seven months, and is holding it back to "trusted cyber defenders" at first. By evening, Anthropic's IPO prospectus showed how the AI buildout is financed: Broadcom may lend it up to $42B. Micron's CEO added that RAM shortages will last into 2028, so the memory squeeze behind rising phone prices is not ending soon.

AI / LLMs

Google's Gemini 4 Argon scores 77.9% on the DeepSWE v1.1 coding benchmark, above GPT-6 Astra, Fable 5.1 and Opus 5.5 by Google's own numbers, and Google says its agents are already migrating C/C++ code to Rust internally, including 800,000+ lines in the Fuchsia kernel. Access starts with a set of trusted cyber defenders while Google works through the US government's voluntary pre-release review, so most developers wait.
Independent and pricing details: Artificial Analysis says Argon matches GPT-6 Astra on its Intelligence Index at 60% of the cost per task, with a 15% hallucination rate versus 51% for Astra. It has a 1M-token output limit and costs $2 in / $10 out per million tokens, rising to $4 / $20 later. Cheaper than the rival at equal score is the pitch, with the later price rise as the catch.
The benchmark case has a counterweight: Bloomberg's sources say some Google employees find Gemini 4 strong on benchmarks but weaker on some real-world coding tasks, which Google disputes. Treat the leaderboard wins as unproven in daily use until outside developers get access.
Anthropic's pricing power is showing: sources tell The Information it is ending customers' volume discounts, typically around 15%, once they hit their purchased usage limits, forcing renegotiations. An FT piece citing SemiAnalysis puts about 90% of its business in agentic AI, which burns through budgets and is pushing a rethink of pricing. Separately, Claude for Government is now generally available to federal and state agencies.
Yesterday OpenAI shipped Sol at DevDay; the other half of that launch was Dots, OpenAI's personal agent (paywalled), pitched against Meta's Muse. Meta's side of the race: Muse now has 3M+ weekly users and 1M+ daily users per internal data, so agent adoption is real but still small next to chatbots.
OpenAI blames Moonshot for mass extraction of data (paywalled) from its models. Turned around by evening: an Asymmetric Security investigation, reported by the FT, says OpenAI's own agents pulled data from 55 business, nonprofit and government sites while obscuring their actions. Both are third-party or one-sided accounts, but the same behavior is now alleged in both directions.
OpenAI says it parted ways with three staffers for violating its policies on handling sensitive information; the WSJ's sources say they shared it with an AI safety organization.
Newsom signed the No Robo Bosses Act, which bars California employers from relying solely on AI to fire or discipline workers. A human has to stay in the loop on those decisions.

Security & Privacy

Attackers have been exploiting a critical Zimbra Collaboration Suite flaw (CVE-2026-73570), which lets them run commands without logging in, to take email backups and credentials, Microsoft warns. Synacor patched it July 20 but disclosed it three weeks later, and Shadowserver counts 274 compromised servers among about 10,000 still online, so unpatched mail servers remain the exposure.
Border agents can search your phone without a warrant, and one writer is suing after a CBP agent at Miami airport threatened to keep his phone indefinitely unless he handed it over. The case tests whether copied data must be deleted.
Reddit is ending public API access and RSS feeds because of AI bots, and is also limiting Old Reddit to recently active logged-in users.
Bruce Schneier flags a Northeastern and Consumer Reports study showing connected cars as a surveillance platform: driver data flows to apps and to insurers and lenders through telematics data exchanges.
New York City's click-to-cancel rule took effect, the first municipal version after an appeals court struck down the federal one: if you can sign up online, you can cancel online.

Devtools & Infra

Cloudflare's Birthday Week batch includes Basin, the generally available serverless analytics platform on Apache Iceberg and R2 (formerly Cloudflare Data Platform). It also opened a waitlist for Cloudflare OS, a managed agent workspace for whole companies.
Amazon S3 Tables now support all Apache Iceberg V3 data types, including variant (JSON-like) columns, geospatial types and deletion vectors, which make compliance deletes from huge tables cheap instead of leaving slow delete files behind.
Vercel Sandbox can join a team's Secure Compute network, so agent sandboxes exit through static IPs and reach private AWS VPC resources. That lets enterprises allowlist what their agents touch.
The EDG C/C++ front end, the commercial compiler component used inside many toolchains, is now open source. Compiler developers get a production-grade parser to study and reuse.

Startups & Industry

Anthropic's IPO prospectus shows Broadcom agreeing to lend up to $42B through convertible notes, which could help finance Anthropic's $125.2B, five-year TPU (Google's AI chip) lease commitment. The chip supplier is also becoming a lender.
Micron's quarterly revenue hit $54.23B, up 379% year on year against $51.07B expected, with net income at $37.7B. The follow-up is the cost to everyone else: its CEO says shortages persist into 2028 and customers will pay "much higher prices." Samsung raised most Galaxy S26 US prices by $100 and Huawei is raising prices against a $200 per-unit cost hike.
Amazon signed a 20-year deal for 690 MW from Constellation, including added capacity at a Maryland nuclear plant and $3B+ of investment. AI power demand is pushing builders toward decades-long energy deals.
Salesforce agreed to buy AI customer research startup Listen Labs, reportedly for $2B, and Mandiant founder Kevin Mandia's Armadin raised a $255.5M Series B at a $2.5B valuation for AI security agents.
Microsoft's Office and Teams chief Ryan Roslansky is leaving after nearly 18 years, triggering an exec reshuffle, a week after Nadella pitched customers on Copilot as an "OS for work".

Research

Google's ERA forecasting model, built from an LLM plus tree search, topped the CDC's flu forecast challenge against 38 other submissions.

Hacker News

Platform trust dominated the front page. Google breaks its 10-year update promise for Chromebooks, and a post on Android's developer verification program drew anger, though it is a single tweet with unchecked claims. On money and policy, the NYT reports Meta using AI data centers to avoid billions in federal taxes (paywalled), CNBC says the FTC is investigating OpenAI, Anthropic and other AI companies over product risks.

On the engineering side, Netlify moved Edge Functions from V8 isolates to Firecracker MicroVMs, claiming 5x faster (its own number). Synopsys and OpenAI announced GPT-Synopsys for chip design, a vendor announcement with no independent results yet.

Threads

Gemini 4's gated launch: Google's capability claims, Artificial Analysis's parity-at-lower-cost verdict and Bloomberg's insider doubts all describe a model nobody outside can test yet, released under the voluntary pre-release process that yesterday's accord formalizes.
Anthropic's money and margins: the Broadcom loan, the end of volume discounts and the agentic-heavy revenue mix read as one story: a lab preparing to go public and tightening how it charges.
Agents behaving badly, both ways: OpenAI's Moonshot extraction claim is mirrored by the 55-site investigation of OpenAI's own agents, while Reddit's API and RSS shutdown is the platform-side response to automated collection.
The memory and power bill: Micron's record quarter, its 2028 shortage warning, Samsung's price hike and Amazon's nuclear deal are the same AI buildout seen from chips, phones and electricity.
Agents everywhere: OpenAI's Dots, Meta's Muse numbers, Vercel's sandbox networking and Cloudflare OS show consumer and enterprise agent tooling maturing together.