XIO

xWire

Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.

The New Stack | DevOps, Open Source, and Cloud Native News

Your AI coding agent could cost 5x more depending on the harness, benchmark finds

The same AI model can cost up to five times more to run depending on the coding agent harness it’s paired with, according to a new benchmark that tested 22 configurations across five open-weight models. Runta and Baseten ran those combinations 27,280 times for FrontierHarness Eval v1.1, consuming roughly 83 billion tokens in the process. Nearly 65% of estimated inference spending went to runs that failed verification or timed out. The model itself accounted for 92.8% of the cost differences,…

OpenAI’s own tests show GPT-6.2 Astra is twice as likely to hide coding failures

OpenAI’s new GPT-6.2 Astra is nearly twice as likely as its predecessor to leave users unaware of problems that prevented it from finishing a coding task, according to the company’s own safety evaluations. Developers building autonomous agents could encounter the same problem, since OpenAI has added a mitigation to Codex that isn’t automatically applied to direct API requests. In its GPT-6.2 system card, published Tuesday alongside the release of GPT-6.2 Astra and GPT-6.2 Sol, OpenAI reported…

Microsoft’s new coding model promises near-Opus performance at 5% of the cost

At its Ignite conference in San Francisco on Tuesday, Microsoft announced a number of new AI models, including two new coding models. Imaginatively named Code 2 and Code 2 Flash, these new models promise to come very close to models like GPT-6.1 Sol and Claude Opus 5.5 at less than half the cost. As Microsoft notes, “MAI Code 2 is at the frontier of price-performance” for enterprise coding tasks. “It does all of this while also delivering changes code maintainers would merge, at a fraction of…

Microsoft’s ARM-based Cobalt 200 chips go GA. Why it matters for AI agents.

Microsoft is making its Cobalt 200 virtual machines generally available, with CPU improvements aimed at the sandboxes, tools, and code execution that keep AI agents running. At its Ignite conference in San Francisco on Tuesday, the company announced that the VMs, powered by its second-generation Arm-based CPUs, are available in 14 Azure regions, expanding to 23 by the end of the year. Microsoft says they offer up to 50% better performance compared with Cobalt 100. It’s been almost exactly a…

Replit’s agent swarms built an app monitoring platform in 2 days on a $3K budget. Now, it’s offering them to customers.

Agent swarms are all the rage, as companies set about finding ways to coordinate multiple AI agents to tackle complex tasks in parallel. In the past week alone, Google and Cursor unveiled competing approaches to multi-agent coding, building on existing efforts from companies including Anthropic, whose experimental Claude Code Agent Teams feature lets multiple agents collaborate on development tasks. OpenAI has also explored shared workspaces for agents. On Tuesday, Replit revealed that it’s…

CoreWeave connects custom training loops to model serving

Production traffic shows failures that benchmarks or pre-deployment evaluations can’t fully reveal. It tells you which edge cases matter and what users need. Turning that signal into the next model version requires an upfront decision: does the model need to change, or is the problem in retrieval, context, prompts, tools, or serving? Changing weights won’t fix a failure elsewhere in the system. When the model is the right place to intervene, production evidence can improve both quality and…

AI agents were all over KubeCon. Production is another story

AI agents are testing systems that were originally built for people, and Jonathan Bryce thinks the cloud-native community has a lot of work to do to help those systems keep up in this new era. The New Stack sat down with Bryce, the executive director of the Cloud Native Computing Foundation (CNCF), at KubeCon + CloudNativeCon North America in Salt Lake City to talk about the projects joining the foundation and how AI is changing its work. Bryce tells The New Stack, “A lot of the existing…

Google found hackers using autonomous AI agents to turn security patches into exploits

Google’s Threat Intelligence Group says attackers are using autonomous AI agents to reverse-engineer security patches, generate exploits, and attempt to execute them without human approval. In a report published Monday, Google and Mandiant describe the first cases their researchers have identified of attackers using LLM frameworks to fully automate vulnerability discovery and exploitation. The framework uses an orchestrator to coordinate specialized agents, share context, and invoke tools…

DuckDB 2.0 goes client-server and adds an Agent Mode for coding agents

DuckDB has always been the database you don’t have to set up: Import it into your code and start querying. DuckDB 2.0 was released in October, and it can now run as a server that other machines can connect to. DuckLabs’ Gábor Szárnyas calls this the headline feature of the release. The team decided to make this latest upgrade a major release since that feature, along with a new storage format and a new SQL parser, marks a big shift in how DuckDB works. On The New Stack podcast, Szárnyas…

Dax Raad rebuilt his browser tool in Rust for $60. Then memory use fell 97%.

OpenCode developer Dax Raad says he rebuilt the project’s browser automation tool in Rust for $60, replacing the parts of Playwright and Patchright it actually used and dramatically reducing its memory footprint. A 4.3 MB binary The figures he posted to X on Saturday are striking. The new Rust binary is just 4.3 MB, compared with 53 to 64 MB for the previous Node version. Memory use for each OpenCode Browse connector also fell from roughly 190 MB to 5.6 MB after a few calls, a reduction of…

With Qwen4-Max, Alibaba puts the open source jacket back on the hanger

Alibaba has officially unveiled Qwen4-Max, the first model in its next-generation Qwen4 family, with a focus on complex professional tasks and autonomous AI agents. The Chinese tech titan teased the new Qwen4 family last week, taking to X to promise improvements in coding, reasoning and real-world knowledge, while initially making Qwen4-Max available in preview through several of its products. Now, the company has formally detailed the model’s capabilities, pitching it as an always-available…

DuckDB has an answer to “bloated” database messages

Older database protocols were built before big data. Some still send results row by row, repeating details like column names and types with every row. The result, says DuckLabs product team lead Guillermo Sánchez Dionis, is “quite a bloated message.” That overhead adds up when you’re moving large query results. DuckDB’s answer is Quack, an HTTP-based protocol that streams query results in batches of column data as DuckDB produces them. It was previewed in May and was released in October as part…

Claude Fable 5.2 vs. Opus 5.5: Same scores at almost 4x the cost

Anthropic released Claude Fable 5.2 on November 9 and said it beats Claude Opus 5.5 in almost every capability the company measured. In the same announcement, it recommended starting with Opus for most coding and knowledge tasks, calling it its “strongest daily driver.” I wanted to find out why a company would steer people away from its best model, and whether the cheaper one really is enough. …why a company would steer people away from its best model, and whether the cheaper one really is…

AWS Emulator proves agents need a verification loop. Realism is the hard part.

This week, Amazon Web Services released AWS Emulator, a local copy of AWS that coding agents can test against. It covers more than 200 AWS services and was written entirely by agents. The most useful part of the launch is a story Swami Sivasubramanian tells in the announcement. About 20 years ago, Sivasubramanian helped build the replication engine inside Alf, a Paxos-based lock service that still powers AWS systems such as S3 and DynamoDB. The first build took about four months. When he later…

Claude Fable 5.2 vs. Fable 5.1: Fewer refusals, with small savings on the login fix

Anthropic’s Claude Fable 5.2 promises fewer refusals and lower token prices, but those changes played out very differently in our head-to-head coding comparisons. Anthropic compared Fable 5.2 with its predecessor, Fable 5.1, on safeguard blocks, output speed, and API pricing. The company states that Fable 5.2 triggered about 55% fewer safeguard blocks than Fable 5.1 overall and about 60% fewer in cybersecurity in early testing across claude.ai, Claude Code, and Cowork. It also says Fable 5.2…

Web search for agents is here: Why ranking still decides the answer

AI agents have had a URL problem for a while. When an agent needs current information but has no search tool, it often makes up the web address it thinks the answer lives at. It might guess something like docs.example.com/v4.2/auth, try to load it, and get a “page not found” error because the page doesn’t exist. Then it either keeps guessing or falls back on outdated information from its training. “Live web search gives AI agents fresh pages. Deciding which ones to trust over your own data is…

DuckLake 1.1 solves the absurd complexity of data lakes at AI scale

How much simpler could a data lakehouse get if an ordinary SQL database kept track of its files? That’s the question in this episode of The New Stack podcast. DuckLake is an open-source table lakehouse format created by Amsterdam-based DuckLabs that acts as a traffic cop for data management by storing table metadata inside a standard SQL database, rather than scattering manifest files across object storage. Unlike traditional table formats, DuckLake treats lakehouse metadata as a relational…

Anthropic wants you to build on Claude. There’s a catch.

Anthropic wants developers to build businesses on Claude, but its expansion into enterprise software is raising concerns among founders that the company supplying their models could eventually become their biggest competitor. A nervous warning to wrappers Gerald Zingraf, president and COO of GlacierGrid, raised the concern on LinkedIn following Anthropic’s Tuesday expansion of Claude’s enterprise lineup, comparing its strategy to Amazon’s private-label business. “Makes me nervous for wrappers…

Claude Fable 5.2 vs. GPT-6 Astra: A lower list price didn’t mean a lower bill

Anthropic is positioning its new model, Claude Fable 5.2, for demanding coding and technical work. Anthropic launched Fable 5.2 on Nov. 9 and reported higher scores than OpenAI’s GPT-6 Astra on two coding benchmarks. Fable 5.2 also has lower standard API rates. Anthropic’s launch benchmarks put Fable 5.2 ahead of Astra, 49.1% to 42.8% on FrontierCode 2.0 and 69.9% to 65.5% on FrontierSWE v2. Fable’s 49.1% came at X high effort. At High, its default on the Claude API, it scored 45.3%. Benchmark…

Anthropic now lets you self-host Claude Managed Agents. The catch? 8 hours to act if it flags a safety risk in the software

Anthropic has unveiled a self-hosted version of its platform for building and running AI agents. Self-hosted Claude Managed Agents, as it’s aptly called, is available now in private beta. Customers host both the server that coordinates the agents and the sandboxes where the agents run their tools, either in their own cloud account or in an on-premises data center. Credentials and session histories stay on their infrastructure, while requests to Claude go to the inference provider they…

The context problem: It’s what’s really breaking AI agents at scale

AI models are getting smarter, tokens are getting cheaper, and agents are improving. But the context gap sits between the modern enterprise and agentic nirvana, keeping many companies from realizing the ROI they hoped for from their AI investments. As with most problems we’ve seen in the AI era, tech companies are rushing to fill the gap. The New Stack recently spoke with Sri Desikan, an Elastic executive focused on search, to learn more about the AI context gap and how his company is tackling…

DeepSeek’s coding agent is too tidy for its own good — and benchmarks are punishing it

DeepSeek’s open-source coding agent is losing points on a major benchmark for a habit developers usually treat as good practice: cleaning up after itself. The problem surfaced when Hugging Face harness optimization researcher Joël Niklaus added DeepSeek Harness to his coding-agent leaderboard, which compares ten harnesses running the same underlying models. DeepSeek Harness is statistically tied for first across the board’s five scored evaluations, yet on Terminal-Bench alone it drops to sixth,…

Here’s how to teach AI agents to search like human experts

Agentic retrieval is a hot topic for a reason: Every company wants to automate work, but they often just are not ready. It’s also extremely easy — and tempting — to spin up an agent, point it at a mountain of data, and spend heavily on tokens for a lackluster result. To unleash agents that can accomplish work, you’ll need a different approach. You can’t throw a more intelligent AI model at a failing agentic workflow without first sorting out how to provide it with better data. It won’t do much…

Google’s Gemini 4 coding agent now switches models mid-task — and finishes 19% faster

Google launched Gemini 4 Hydrogen on Thursday alongside a new system that changes how coding agents use AI models. Rather than keeping its most capable model running across an entire session, Google is pairing Hydrogen with the more powerful Gemini 4 Carbon and switching between the two as the work progresses. The company says the combination gets nearly the same results as running Carbon on its own while using 24% less quota and finishing tasks 19% faster. How Orchestrator handles the work The…

Google’s coding agents now /swarm for one developer. Cursor’s go multiplayer so teammates’ agents can talk.

Coding agents are pretty good at working in parallel: one writes a new API endpoint, another updates the tests, while a third fixes a bug elsewhere in the codebase. However, they’re not quite so adept at knowing what their coding counterparts are doing — put several on one project and they duplicate work, trip over each other’s changes, and leave the developer firefighting the fiasco that follows. That’s why the major AI labs and coding-tool makers have all been looking for ways to bring order…

Chainguard defends its choice to break the PHP security embargo: “We own the decision.”

At KubeCon + CloudNativeCon North America 2026 this week, Chainguard co-founder and CTO Matt Moore defended the company’s decision to release patched PHP images before an agreed-upon embargo with other partners had lifted. When Chainguard shipped patched PHP container images more than six hours ahead of an October 21 coordinated security release, PHP’s maintainers accused the company of giving attackers a head start. Chainguard argues attackers already had one. Chainguard founder and CTO Matt…

Google built an AI model that can’t write a sentence — and that’s the point

Google’s new Gemini Decisions API, introduced Wednesday, is designed to handle the routine decisions that AI applications make thousands of times a day, with lower costs, and faster response times than traditional generative models. …the routine decisions that AI applications make thousands of times a day, with lower costs and faster response times than traditional generative models. Instead of using a generative model to work through each decision and produce a text response, the API uses a…

The death of the 14-day patch: Surviving machine-speed vulnerability discovery

For decades, enterprise vulnerability management ran on human time: A security researcher manually dissected a binary, traced execution paths, built a proof of concept, and filed a coordinated disclosure. Engineering leads evaluated the CVSS score and prioritized the fi before working it into an upcoming sprint schedule. Then, finally, they shipped the update. That deliberate cadence is a thing of the past. In the first nine months of 2026, publishers recorded 75,242 CVEs, representing roughly…

See what AI looks like in production at Microsoft Ignite

If you’re deciding which AI pilots to move into production, Microsoft’s Satya Nadella and Judson Althoff’s opening keynote at Microsoft Ignite will show you where the company is taking the tools you’ll use to build and run those agents. They take the stage at 10 a.m. Pacific on Tuesday, November 17. After the opening keynote, you can experience how customers put AI into production at scale, see where Microsoft is taking the tools those customers use, and compare notes with industry colleagues…

Anthropic built a model that beats Opus. It recommends Opus anyway.

Anthropic just released a coding model that beats Opus 5.5 on two major benchmarks and can complete certain difficult tasks for half the cost. Yet in its Fable 5.2 announcement on Monday, the company recommended that most developers stick with Opus, calling it the “strongest daily driver” for coding and knowledge work. “Based on capabilities and pricing, Claude Opus 5.5 is still our recommended model for most everyday work, from coding to knowledge work,” Anthropic tells The New Stack. “Based…