BriefTechNews
GPT-6.2 Astra and Sol hit the API as Huang dangles a Nvidia comeback

GPT-6.2 Astra and Sol hit the API as Huang dangles a Nvidia comeback

7 min read · 15 sources

TL;DR
  • GPT-6.2 Astra costs $8 per million input tokens, 20% less than GPT-6 Astra, and beats it by 5.2 points on Agents' Last Exam.
  • Jensen Huang says he will return as Nvidia CEO if asked, with a board faction pushing despite the board's denial.
  • Google signed a $500M, 50MW geothermal deal with Quaise Energy to vaporize rock for data center power.
  • Anthropic's chaos loop held Claude Docs to 60 server errors across 2 billion requests on 50 pods.
  • Uber and Zipline launched commercial drone delivery in Spring, Texas, targeting one million deliveries per day by 2029.

OpenAI dropped a two-model professional tier this morning, and the pricing is the headline. GPT-6.2 Astra and GPT-6.2 Sol are live in ChatGPT Work, Codex and the API as gpt-6.2-astra and gpt-6.2-sol, and Astra undercuts its predecessor by 20% on output tokens while beating it on the benchmarks that matter. If you run agents in production, the cost-per-task math just shifted.

The bigger drama is in Santa Clara. Jensen Huang broke his silence and said he would return as Nvidia CEO if the directors ask - and a faction of the board is pushing for exactly that, despite the company’s official statement that it isn’t considering it. The man who resigned over extramarital affairs, one involving a suspected Chinese spy, is now the subject of an employee and investor revolt. That story has legs.

Astra leads Agents’ Last Exam by 5.2 points over GPT-6 Astra at less than a third of the cost.

GPT-6.2 Astra: frontier results at a discount

Source: testingcatalog.com ↗

The model pair splits cleanly. Astra is the frontier workhorse: it leads Agents’ Last Exam by 5.2 points over GPT-6 Astra, and Terminal-Bench-Science 0.1 by 9 points. Sol is the efficient sibling, coming within 0.5 points of Astra on FrontierCode 2.0 at one-fifth the cost.

The API pricing is where this gets concrete. Astra: $8 per million input tokens, $0.40 cached, $40 output. That’s 20% cheaper than GPT-6 Astra on output. Sol keeps GPT-6.1 Sol’s rates. Both models produce better results with fewer output tokens, which means the same usage limits go further in ChatGPT Work and Codex.

For engineers, the practical effect is in review cycles. Astra’s claimed capabilities in document generation, coding and CAD tool use mean fewer tokens spent on iterations. If the benchmark gap holds in your workload, you are looking at a straight cost reduction for the same quality bar.

Huang wants back in, and the board is split

Source: wsj.com ↗

The WSJ report is unambiguous: Huang says he is willing to return if asked. The outpouring from employees, investors, business partners and customers is real, and President Trump has said he would support reinstatement.

The Semafor follow-up shows the board is not unified. Chairman Harvey Jones says the resignation acceptance was unanimous, but other directors are narrowing the search for a replacement - considering ARM CEO Rene Haas and Marvell CEO Matt Murphy - while Huang has met with tech executives and spoken to at least one director. Haas has expressed concerns about the optics of taking the job while the calls for Huang’s return grow louder.

This matters beyond the soap opera. Nvidia is in the middle of a mega-merger with Lumentum. Leadership instability at that scale affects partnership decisions and strategic direction. If you are building on Nvidia’s roadmap, the uncertainty is a real operational risk.

Google bets $500M on vaporizing rock

Source: wsj.com ↗

Quaise Energy is going to vaporize stone to drill at superhot depths, and Google is paying for it: a deal valued at more than $500 million for 50 megawatts of power. The site is in central Oregon, feeding Google’s existing data center operations in the state.

The technology uses directed energy to melt rock rather than mechanical drilling, which lets it reach depths where temperatures produce supercritical steam. That is baseload clean power, not intermittent. For data center operators, 50MW of firm power is the difference between a new region and a stalled one. Watch whether the Oregon site hits its drilling milestones - that is the gating factor for the whole thesis.

Drones are delivering lunch in Texas

Source: techcrunch.com ↗

Uber Eats and Zipline went commercial in Spring, Texas. The drones carry up to eight pounds, fly at 300 feet and 70 mph, and lower packages via tether - no landing gear needed. Participating restaurants at launch include Jersey Mike’s, Zaxby’s and local Houston spots.

The ambition is one million deliveries per day by the end of 2029, with Dallas and dozens of other cities to follow. The engineering challenge is not the drone; it is the orchestration. Airspace deconfliction, noise complaints, weather windows and regulatory compliance at that scale is a distributed systems problem with a physical component. If they hit the volume, it changes the economics of last-mile delivery.

Anthropic's chaos loop: 60 errors in 2 billion requests

Source: claude.dev ↗

Anthropic published the postmortem for the Claude Docs launch, and the numbers are striking. The service launched September 16, 2026. By September 22, it had seen 4.6 million docs created across 2 billion requests - with only 60 server errors, no rollbacks, running on about 50 memory-optimized pods.

The technique was a “chaos loop”: a staging cluster where Claude drove over a million synthetic docs a day while the team broke core components on purpose, for two weeks. The key decision was avoiding runtimes with garbage collection to keep telemetry legible to Claude. That is a specific, transferable lesson: if you want an AI to debug your system, give it clean signals about memory behavior.

The week of outages was not an attack

Source: cnbc.com ↗

Four consecutive business days of outages hit Microsoft Outlook, the FAA (grounding flights), JPMorgan Chase, United Airlines, the NSA and Apple. Director of National Intelligence Jay Clayton dismissed cyberattack suspicions, blaming “scheduled routine maintenance” and “internal updates.”

Security researchers have a different explanation: AI is accelerating vulnerability discovery, so businesses are patching more aggressively than ever, and the patching itself is breaking things. For engineers, the lesson is operational. If you are running deeply embedded services, the outage plan needs to assume the patch itself is the failure point.

Copilot is now the OS, and Office is the plugin

Source: arstechnica.com ↗

Microsoft at Ignite inverted the application model. Word, Excel and PowerPoint are now generally available inside Copilot, which CEO Satya Nadella called “the new OS for work.” Agents drive tasks; apps are plugins. Copilot Autopilot, a persistent proactive agent, will use usage-based billing via Copilot Credits - pricing not yet announced.

This is a governance change, not just a UI change. Agents accessing corporate data through an organizational knowledge layer means your security model has to account for agent-driven, cross-application access. Microsoft IQ, including Work IQ, is generally available as the context layer. If you are a Microsoft shop, the integration surface just got much larger.

ABFS: Android checkouts go from 36 minutes to 8 seconds

Source: developers.googleblog.com ↗

Google open-sourced ABFS (Another Build File Service), a lazy-loading, content-addressable cache for source and build files. For Android 17 - 1,084 Git repositories, 1.62 million files, 300 GB of built workspace - checkout times drop from ~36 minutes to under 8 seconds. Full builds go from ~80 minutes to just over 2 minutes.

The consequence is parallel agentic workflows. If the checkout is that fast, you can spin up many agents working on different parts of the tree simultaneously. For anyone running massive monorepos, this is the pattern to steal.

eve 1.0 goes GA

Source: vercel.com ↗

Vercel’s eve hit 1.0 after 230+ releases since June. It is an open-source framework for production agents: a harness for calling models, running tools and delegating to subagents, plus a Workflow SDK for durable execution and multiplayer sessions. Managed hosting goes through Vercel Workflows and Sandbox.

The proof points are in the numbers: Rillet tripled its shipping rate with a software factory, and Clay’s monty agent is used by 30-40% of the company weekly. For teams building agents, eve offers an observable, customizable alternative to the closed platforms.

The tail

Source: daniel.haxx.se ↗

The curl rewrites in Rust are real - one November rewrite converted 185,000 lines of C into 340,000 lines of Rust for $8,600 - but the original project merged over 300 commits and 200 bugfixes in the same period. The rewrites are snapshots; curl is a living project. Don’t bet production on a static copy.

The recursive self-improvement piece cites a METR report showing AI R&D progress roughly doubled at Anthropic since October 2025, driven by a $10bn data-generation industry selling RL environments to labs. Caveats: the report redacted data and Anthropic’s IPO may bias projections. The acceleration is real; the jaggedness of ASI means progress will be uneven.

George Hotz’s essay argues AI requiring exponentially more resources for linear returns breaks down at some point, and that nature didn’t collapse to a singleton. A useful counterweight to the scaling-is-everything crowd.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like