The New Stack
Articles about The New Stack
“A major overhaul”: OpenCode 2.0 puts its terminal, desktop, and web apps on one shared service
OpenCode 2.0 entered public beta in July. Now, Anomaly explains the architectural overhaul, including a rebuilt API, an embeddable SDK and Code Mode.
GPUs on Kubernetes are bring-your-own-everything, starting with the bill
CPU and memory come with per-pod metrics, rightsizing and cost splits built in. Shared GPUs come with none of it, so platform teams end up rebuilding visibility for the most expensive line on the bill.
Google paid $90,000 for this bug, then waited two months to fix it
An OAuth secret doubled as a signing key on Google’s developer docs servers, enabling remote code execution. Google paid $90K but took two months to patch.
Coding agents made building easy. Platform teams that build from scratch are 3.5x more likely to lack a strategy
Platform teams that build from scratch are far more likely to report a lack of a clear strategy. As AI makes building easier, the reason behind the choice matters more than ever.
Andrej Karpathy says only 5,000 people have seen what AI can really do
Andrej Karpathy says the AI capability gap is widening. Box CEO Aaron Levie says agent swarms are next, but most enterprises aren’t ready to use them yet.
Claude Code found my app’s shared contract in two MCP calls. Then the records ran out
Claude Code’s redesigned Projects give agents a shared memory of a project’s decisions. I tested what a brand-new session with no conversation history or local code could recover from versioning records, and where it hit a wall.
Claude found 29,000 possible bugs in open source. Only 516 have been fixed
Anthropic’s OSS Scanner sends unreviewed Claude vulnerability reports to open source maintainers, but finding bugs is proving far easier than fixing them.
How to turn AI production feedback into better agents
Your AI agent’s service is healthy. Its answers might still be getting worse.
Microsoft skipped OpenAI’s decision model and built its own on Alibaba’s Qwen
Microsoft’s Decision-1 matches Jev’s price and skips OpenAI’s model. Here’s who it’s built for and what it still leaves out for developers building agents.
Elon Musk’s Grok Bot will pick Claude over Grok when it’s better. One-model loyalty is dead
Even Elon Musk’s Grok Bot now picks the best model for each task. Here’s why companies are doing the same, and how cheap models are taking on most of the work.
Kubernetes on cgroup v1 is dead. Here’s what comes next
Welcome back to another edition of Road to KubeCon, where we’re tracking the latest movements in the cloud-native ecosystem as
Cloudflare acquires Node.js creator’s startup that copied its serverless playbook
Ryan Dahl’s Deno built Celld, a self-hosted take on Cloudflare Workers. Now Cloudflare is absorbing the team, and Deno’s own runtime is being wound down.
Amazon ECS now auto-repairs failing GPUs and instances. Here’s why it matters for SREs
Amazon ECS now auto-repairs failing GPUs and instances, mitigating outages and simplifying cloud application resilience for SREs.
AWS, Upstage and Ollama agree on a decision-model API. OpenAI hasn’t signed on
Decision models are converging on a startup’s API. Developers gain cheaper, portable AI judgments, but shared code doesn’t guarantee reliable answers.
“Don’t use ‘open weight’ and ‘open source’ interchangeably”: Percona CEO on why AI terminology matters
Peter Farkas warns of the threat of “open washing”. And he thinks the AI industry could set a damaging precedent for software more broadly.
Harness bought Augment’s coding agents. The best feature hasn’t shipped yet
Harness bought Augment Code’s Cosmos to link AI coding agents to its delivery platform, but the part that learns from failed deployments hasn’t shipped.
Claude can now build your dashboards
Anthropic’s answer to OpenAI’s Data agent shows the query behind each number.
GitHub Copilot is going local — but Microsoft won’t say what gets sent to the cloud
GitHub Copilot will soon route coding tasks between local and cloud models, but Microsoft hasn’t said what context leaves your machine or how to stop it.
You can outgrow vanilla Postgres without abandoning Postgres
Postgres holds up well until live data keeps growing. See why performance slips at scale, what tuning can still solve, and when TimescaleDB offers a better path.
Most AI incident tools still leave a key decision to humans
In our October 22 webinar, Traversal CTO Raaz Dwivedi explores what it takes to trust AI agents to decide when production incidents need attention.
Your AI agent just provisioned a resource. Who owns it?
Stop runaway cloud costs from AI agents. Use queries, policies, and TTLs to enforce human ownership on all provisioned cloud resources.
The audit log says my name: what an agent inherits when you hand it your credentials
Discover what AI agents inherit when borrowing human credentials, from collapsed audit logs to wide production database permissions.
AI data lakes are driving new storage demands
AI data lakes are driving new storage demands. Learn how ultra-high-capacity QLC NAND flash solves critical data center bottlenecks.
The API tax: Why AI agents stall without infrastructure context
An AI agent saw a single EC2 instance and flagged nothing. It couldn’t see the policy it was breaking.
AI changed how you respond to incidents. Your tabletops haven’t caught up
AI changed incident response, but tabletop exercises haven’t caught up. Learn why and how to test AI tooling alongside human responders.
Claude Sonnet 5.5 vs. Opus 5.5: 42% cheaper and perfect on every run
Opus 5.5 failed two of 15 coding runs. It thought until it ran out of tokens.
Building scalable, agent-friendly APIs for AI applications
Apply microservices principles like bulkheads to scale agent-friendly APIs, handle bursty AI traffic, and prevent lost conversation context.
Redis by proxy: Percona targets the last major hurdle to Valkey adoption
Valkey-proxy lets companies stuck on Redis migrate without having to rewrite their apps. The code is set to go public this month.
OpenSSH 10.6 deliberately breaks two features in the name of security
OpenSSH 10.6 weakens SSH compression and rejects $ and in command-line usernames to close a plaintext leak and a shell injection risk in automation.
Claude’s cyber safeguards are getting more flexible — but not for everyone
Anthropic expands verified cyber access, but fewer safeguards require deeper review. Offensive tests leave the impact on everyday defensive work unclear.
Decision models are suddenly everywhere. OpenAI’s is now public
Decisions API enters public beta with image inputs as Perplexity, Cloudflare, and Amazon push open-weight alternatives.
Kubernetes won the orchestration wars. Its teams are still fighting
Kubernetes won the container orchestration wars. The CNCF reports that 82% of container users run Kubernetes in production. Despite its market dominance, Kubernetes teams aren’t always happy. It turns out that keeping developers and operations teams happy takes more than agreeing on which technology to use. Developers appreciate Kubernetes self-service capabilities, which allow them to scale apps and tweak configurations without ex
OpenAI’s former safety lead says it’s shipping “new capability and risk every Tuesday”
Former OpenAI safety lead David Robinson tells Ezra Klein that reasoning updates, new tools and coding agents now let OpenAI ship new AI risk every week.
Anthropic launches Haiku 5.5 at a much lower price
Haiku 5.5 adds effort controls and posts big gains on computer-use and command-line benchmarks as Anthropic pitches its smallest model for agentic work.
OpenSearch veterans launch Infino. Here’s why it matters for agent builders
Infino emerges from stealth with an open source retrieval layer that lets AI agents query one Parquet copy of data, replacing a fragmented data stack.
GPT-6.1 Sol vs. GPT-6 Astra: Same accuracy at 18% of the cost
OpenAI says its cheaper model nearly matches Astra. I tested it.
What’s up, Docsy? Google’s docs project joins the Linux Foundation as AI agents become readers
Docsy already generates Markdown and llms.txt for AI tools. Next, the open source project wants to score how easily agents can read your docs.
The CNCF is graduating projects faster than ever. AI agents are helping with the due diligence
OpenAI scaled on Kubernetes years before ChatGPT. The CNCF’s Jonathan Bryce says AI and open source now feed each other, and agents are speeding it up.
Your phone’s vector index might be bigger than the AI model running it
Google’s EmbeddingGemma 2 brings text, code, image, video and audio search to one on-device model, so developers can skip captioning and transcription.
Copilot tops GitHub’s own AI code review benchmark. An independent one tells a different story
With ReviewBench, GitHub wants a common yardstick for AI code reviewers. But the team behind it ran every rival’s initial test itself.
Mistral’s new AI tried to escape its test environment. In three weeks, anyone can download it
Mistral’s 1-trillion-parameter Large 4 tried to go beyond its testing environment. Experts get three weeks to probe it before open weights ship Oct. 27.
SAP acquires TechWolf to feed more context into its HR agents
SAP is buying TechWolf, whose skills graph tracks the work employees actually do, to ground its SuccessFactors HR agents and cut what they cost to run.
“There is just not a room for error”: SAP is putting its AI agents in the back office
At Connect, SAP puts new Joule AI assistants to work in finance, HR, and supply chain, launches SAP Pay, and plans to acquire TechWolf.
One MCP server used 18,000 tokens before doing anything. Here’s the workaround
Pi spent much of the past year leaving MCP out of its coding agent, even as the protocol became standard
Developers are secretly hoping OpenAI fails to ship this month
OpenAI’s Codex team will ship one improvement a day for 28 days or reset usage limits. Day 1 sped up GPT-6 models, but the rules remain loosely defined.
Dynatrace wants its AI agents to fix problems instead of adding alerts
Dynatrace closed its $915 million acquisition of Arize on October 1, joining Arize’s tracing and evaluation of AI agents to
The AI safety check that runs on a laptop and nearly matched a 35B model
Red Hat benchmarked AI guardrails. A 200M-parameter classifier nearly matched a 35B LLM judge on prompt injection, while Jev won on content safety.
XCOR launches to trace outages in minutes. It still pages engineers
Palo Alto Networks’ Cortex XCOR uses AI SRE agents to root-cause incidents in under three minutes, the company says, so engineers aren’t woken at night.
Agents have made CI the bottleneck. Faster pipelines are the wrong fix
Agents pushed CI job volume up 25x at Anthropic. Faster pipelines help, but they still check a repo, not the distributed system it has to work with.
AI is speeding up exploits. Vulnerability spreadsheets can’t keep up
AI is widening the gap between the CVEs teams find and the ones they can fix. Closing it means asking which flaws attackers can actually reach.
Anthropic’s answer to Dots and Muse is already inside Claude
Claude, Dots, Muse and Grok Bots are competing for work that continues after you log off. The real test is whether users trust them enough to hand over the next job.
GitHub’s advice for its new Copilot feature is to try something else first
GitHub Copilot can now click and type in desktop apps with no API or MCP server. Here’s how approvals, enterprise controls and reliability limits work.
What Kubernetes’ “monolith” lesson means for AI agent harnesses
In this edition of our Road to KubeCon series, Craig McLuckie wants agent harnesses off the laptop, and Koordinator pushes GPU allocation past 95%.
“No reason why everyone should have an identical Claude experience”: Anthropic’s mods let you change Claude Code’s look and behavior
Claude Code users can now ask it to redesign itself. Mods can rewrite prompts, block tool calls, and redraw the interface, all from a plugin.
K3s vs. K8s: When lightweight Kubernetes distros win (and when they don’t)
Compare K3s vs. K8s to find the best Kubernetes distro for your needs. Explore resource footprints, edge use cases, and key trade-offs.
OpenAI’s always-on agents are free, until one specific thing happens
OpenAI’s always-on Dots agents run without touching usage limits for now, but any task a Dot hands off to Codex or ChatGPT Work still draws from your plan.
Cloudflare brings paid access to MCP tools — who controls the agent’s spending?
Cloudflare’s Monetization Gateway lets sellers charge AI agents for MCP tools and APIs. Here’s what developers need to control before agents start paying.
“No human wants to look at billions of traces”: Dynatrace bought Arize because agents need a new kind of observability
The $915M acquisition is now complete. Dynatrace and Arize explain why AI agents could soon be the ones reading the telemetry.
Bit Cloud’s next chapter starts after the AI builds your app
For many developers, building an application from an AI prompt is already part of the workday. Turning that first version into something a team can use, maintain, and build upon takes more: a place to run it, a way to review changes, and confidence that the next update will work as intended. That’s the opportunity behind Bit Cloud 2.0 and Hope, the company’s built-in AI builder. In this episode
Transform AI from a security blind spot into a roadmap
AI adoption is already outpacing governance. Blocking it won’t help; CISOs need a sanctioned path that’s easier and more useful than the workaround.
AWS launches a local answer to TypeSafe’s Jev decision model
AWS’s new Strands Decider 2B is a small decision model that runs locally and gives AI agents a fast check before they act, with no invalid outputs.
A live Kubernetes cluster can still have an ownership gap
After a Kubernetes cluster goes live, platform and application teams need clear responsibilities for upgrades, access, recovery, and shared workloads.
Use Graph RAG when relationships are part of the evidence
Discover how Graph RAG combines vector search with explicit relationship mapping to provide reliable evidence for complex AI questions.
In AI security, there’s no room for a defender’s mindset
AI threats require an offensive security mindset. Learn how engineering-led teams use proactive automation to stop cyber adversaries.
AI coding agents leaked 13,000 screenshots, and nobody hacked them
AI coding agents leaked 13,000 internal screenshots from 300+ organizations by creating public GitHub repos. Here’s how it happened and how to stop it.
AI agents speed up development — data access slows them down
AI agents made building internal apps fast. Getting them safe access to live production data is still slow, manual, and scattered across every system.
Gemini 4 Argon is here: It’s great, and you can’t have it yet
Google’s Gemini 4 Argon tops OpenAI and Anthropic on most benchmarks, especially knowledge work, but its coding scores are mixed and access is limited.
Cohere’s faster query model barely dents retrieval quality in its tests
Cohere’s Embed 5 lets developers index with the Pro model and query those vectors with the cheaper Fast model, so RAG and agent teams skip a second index.
CloudBees just committed to an AI-first pivot. Here’s why it matters for enterprise DevOps teams
When new CloudBees CEO Mo Plassnig took the job earlier this year, his board of directors handed him a mandate that left little room for interpretation: Do not stay the course. In an era when generative AI is radically reshaping how software is written, deployed, and managed, keeping things business-as-usual for an enterprise continuous integration/delivery (CI/CD) stalwart like CloudBees would be a ticket to irrelevance. Last week at the WeAreDevelopers Conference in San Jose, California,
OpenAI’s Dots boundary problem rate doubled in longer tests
OpenAI’s system card shows Dots agents get more boundary flags as tasks pile up. Here’s what that means for developers building long-running agents.
Your AI agent aced the demo. Your data may still derail it
Your AI agent aced the demo. In production, fragmented, stale, and ungoverned business context exposes everything a controlled environment never tested.
Microsoft Fabric is where AI agents learn how the business works
Microsoft is turning Fabric into the context layer for AI agents, feeding Power BI semantic models into Copilot by default and adding ontologies with rules.
Kill the code review theater, keep the review
Stop line-by-line review theater. Discover how to combine AI guardrails with human judgment to protect team-wide code understanding.
CRA readiness starts in the codebase
CRA compliance starts in the codebase. Learn how to build repeatable, automated security controls into your software delivery system.
Eclipse wants companies to be free to switch AI providers. Today, doing so can mean a costly rebuild
Eclipse Foundation launches Sovereign AI Foundation with 17 members to help organizations avoid AI vendor lock-in, per exec director Mike Milinkovich.
“Think of it as Kubernetes for agents”: OpenClaw lands in the enterprise with Nvidia and Red Hat on board
Most IT teams ban agent platforms outright. Enter OpenClaw Enterprise, an open-source control plane for governing persistent agents.
Why Featherless says you don’t need a tank to deliver a pizza
Featherless releases open source Simple Jev, turning open models into zero-shot classifiers, with free output tokens and pricing from $0.03 per million.
OpenAI makes ‘Sign in with ChatGPT’ a way to use your subscription in third-party tools
ChatGPT subscribers can now easily use their plan allowance in Devin, Amp, Warp, OpenClaw, and other third-party tools.
OpenAI just launched Dots. Here’s why they matter for developers
OpenAI’s Dots give GPT-6 Astra agents their own cloud computers so they keep working after you log off, with read-only research and auto-review guardrails.
OpenAI halves $200 plan allowance, launches $500 plan
OpenAI launches a $500 Pro plan with Ultrafast GPT-6 Astra in Codex, while cutting the $200 Pro plan’s usage from 20x Plus to 10x. Developers aren’t happy.
OpenAI answers TypeSafe’s Jev with a Decision API built on Luna
With the sudden rise of TypeSafe’s Jev, decision models have become incredibly popular, so it’s not surprising that OpenAI is also announcing its take on this model type at its annual DevDay conference on Tuesday.
Claude Opus 5.5 vs. Fable 5.1: One overthinks, the other cuts corners
Anthropic released Claude Opus 5.5 on September 22, saying it performs at the level of Claude Fable 5.1 on most work, and it costs less than half as much per token. Opus 5.5 got 13 of 15 runs right at 22% less cost, while Fable 5.1 got all 15 and finished 40% faster, but took a problematic shortcut to pass one test.
How to reassign resource ownership before someone’s last day
Catch stale cloud resource ownership before offboarding. Learn queries and policies to reassign infrastructure when team roles change.
MCP gets AI agents into your APIs. It doesn’t decide what they should see
Connecting an agent to an internal API is the easy part now. Controlling which fields it can read and which actions it can take, without maintaining a separate tool for every team, is the harder problem.
Ember-1 vs. Kimi K3: Nearly identical results at 3.4 times the speed
Fireworks Research launched Ember-1 on September 23 as a research preview. Built on Moonshot’s open-weight model, Kimi K3, it claims
You picked Claude Sonnet 5.5 — but Anthropic may send your request to Sonnet 5
Anthropic says “higher-risk cybersecurity tasks will visibly fall back to Sonnet 5.” While routine coding remains available on Sonnet 5.5, API developers must opt in to automatic fallback.
OpenAI exposes “new variety of prompt injection” that can spread like computer worms
Tests show malicious instructions can move through email, files, and code comments while causing agents to take unauthorized actions.
Anthropic launches Claude Sonnet 5.5 with near-Opus performance at half the price
Anthropic’s Claude Sonnet 5.5 runs 30% faster, cuts cost per task by up to 30%, and comes close to Opus 5.5 on coding benchmarks at an unchanged price.
Meta hired MongoDB’s CEO to build its enterprise AI business — but Llama is missing
Meta hired MongoDB’s CJ Desai to lead Meta Enterprise Platform, bringing Muse API and Muse Code to developers. Llama is missing from the plan.
OpenAI blocked its agent’s web access. Then it tunneled out through DNS
Two OpenAI models found workarounds after their intended paths were blocked. The incidents exposed gaps in network controls, instruction following, and monitoring.
Think you need a better reranker? Check your retrieval first
Reranking gets expensive fast, and it can’t fix retrieval that never surfaced the right results. Learn to build a multi-stage retrieval funnel instead.
Elon Musk says space will soon hold nearly all compute. Google is still finding out if its chips can work there
SpaceX and Nvidia want full AI racks in orbit by 2027. Google’s first Suncatcher satellite asks a humbler question: Will TPUs even survive space?
GPT-6 Sol vs. Claude Opus 5.5: Is cheaper important when results aren’t consistent?
OpenAI launched GPT-6 Sol and Anthropic launched Opus 5.5 on September 22. I ran both of these cutting-edge frontier models through three hard developer tests, five times each. Opus 5.5 was perfect on all 15 runs, and Sol was perfect on 12 at about a sixth of the cost.
pgEdge’s new agent database branches never merge back into the source
pgEdge launched Starfleet on Monday, giving each coding agent its own Postgres branch and a path from hosted prototype to air-gapped production.
Anthropic bought Stainless and shuttered its SDK generator. Cloudflare open-sourced Forge instead
Eleven days after Google backed an open source SDK generator, Cloudflare releases Forge. Both were customers of Stainless.
Nvidia launches Open Agent Safety Platform to lock down rogue AI agents
After four frontier labs saw agents escape test sandboxes this summer, Nvidia pairs its OpenShell runtime with a BlueField-4 watchdog called Nvidia Sentry.
Performance engineering from kernel analysis to AI: Adrian Cockcroft’s take
Adrian Cockcroft discusses performance engineering, the limits of P99 metrics, and building AI-driven custom tools for response times.
The rise of agentic AI on Kubernetes: unleashing the new infrastructure layer
AI agents can now observe, reason about, and act on Kubernetes clusters. Their value depends on what they can see and the lines teams draw around them.
Avoiding vendor lock-in through an open-source approach: a developer’s perspective
Learn how open source helps developers avoid vendor lock-in, protect digital sovereignty, and maintain flexible software architectures.
The agent didn’t break your controls. It went around them
AI agents don’t stop at a locked door. They try the window. That’s why agent security has to move from the perimeter to the moment an agent acts.
Claude Opus 5.5 vs. Opus 5 on reasoning tasks: Cheaper, faster, but not better
Anthropic says Opus 5.5 is 40% cheaper and 30% faster than Opus 5. On hard reasoning tests, the savings held up. The speed claim fell short.
Microsoft’s new Copilot agents get their own email, calendar — and a place in the org chart
Microsoft’s Copilot overhaul gives Autopilot agents their own Entra identity and runs AI-generated apps inside Microsoft 365, with usage-based billing.
OpenTelemetry and Prometheus are getting along. What’s still missing?
Welcome to another edition of Road to KubeCon, where we’re tracking the major movements in the Kubernetes and cloud native The post OpenTelemetry and Prometheus are getting along. What’s still missing? appeared first on The New Stack.
OpenAI and Cursor agree on agent coordinators. They disagree on who runs them
OpenAI opened its Agents API in public beta this month, exposing the harness that powers Codex with managed sessions, tool
OpenAI’s agent had a routine task. It breached a government portal
AI agents assigned routine data retrieval probed a university library and government health sites for flaws. Here’s what that means for agent builders.
How a forgotten node can put Oracle Java back in production
Azul’s new AI Assistant queries live Java runtime data to flag licensing and security risk in real time, replacing static reports that go stale on day one.
Developers and platform teams both want Kubernetes self-service. They disagree on who owns it
What do developers want? Kubernetes environments when they need them. What do they not want? Those environments after a week
Google’s Gemini CLI now asks before editing your build files
Google’s Gemini CLI 0.61.0 now asks before its coding agent edits build files or runs commands shaped by untrusted content, a check on prompt injection.
OpenAI makes you call sales for a custom voice. Google just made it self-serve
Google’s Gemini 3.8 TTS lets developers design a voice from a text prompt or replicate one from a short clip, then reuse it by ID through the Gemini API.
What managing 150,000 AI agents could look like for database teams
AI agents are taking over the DBA’s hands-on work while spawning far more databases. Yugabyte is betting DBAs will manage the robots instead.
Query decomposition doesn’t fix context starvation — it just moves it
Query decomposition moves context starvation to the packer. Discover why allocation policy beats retriever upgrades in RAG pipelines.
Cursor acquired Firetiger. A month later, it launched a bot that tracks code changes from PR to production
“Writing code is no longer the slow part”: Rollouts takes Cursor deeper into everything that happens after the pull request goes up.
Can enterprises protect data without making AI less reliable?
Learn how to protect enterprise data without sacrificing AI model reliability, referential integrity, or engineering velocity.
Q.ANT gives away the software for its light-powered AI chips in a CUDA-style bet on developers
Developers can try it on a regular laptop today. Access to the actual chips opens up over the coming months.
“Impressive level of openness”: Xiaomi goes way beyond the usual open-weight playbook with MiMo-V2.6
Xiaomi streamed its $3.5M reinforcement learning run live and released the weights under MIT. And there’s more on the way.
A third option is emerging in the fight over AI and your data
Enterprises won’t risk their data, and AI labs won’t hand over weights. Vast Data’s DataEnclave aims to solve the trust problem in both directions.
Jensen Huang says the junior developer problem ends in two years. Here’s his math
Nvidia CEO Jensen Huang tells Ezra Klein AI agents are a new abstraction layer for developers, but the junior engineer pipeline may break before 2028.
Amazon blocked Meta’s Muse. Then Shopify wired it into every store
Amazon kicked Meta’s Muse off its store because the agent never said what it was and seemed to be holding onto people’s passwords. A day later, Shopify let Muse into every one of its stores. Why the two companies treated the same agent so differently says a lot about where AI shopping is headed.
Enterprise AI desperately needs to protect data and models. Here’s how confidential AI could do it
Your AI model can redact customers’ PII. Nobody can be fully sure it worked.
The software supply chain is the new battlefield. AI just changed the rules
AI agents now choose the packages in your code while attackers use AI to chain minor bugs. Chainguard’s CISO says security must start at the source.
Vulnerability alert fatigue nearly swamped WHOOP. But its fix still keeps a human in charge
Join WHOOP and Datadog engineers live on October 7 to see how WHOOP automated vulnerability response and why it still keeps a human in the loop.
AI spending can run negative. Qodo’s CEO built an ROI equation to fix it
Qodo gives every engineer $10,000 in tokens a month. CEO Itamar Friedman says the cap isn’t a limit — it’s the start of an AI ROI calculation.
Anthropic made Opus 5.5 cheaper. Then it broke four things your agent depends on
Claude Opus 5.5 costs less than Opus 5, but four breaking changes to thinking, tool use and computer use can trigger 400 errors in existing AI agents.
OpenAI cut GPT-6 token prices in half. The bigger lever may be the cache
OpenAI cut GPT-6 Sol and Luna API prices in half. The quieter savings come from caching that no longer breaks when agents change reasoning effort or tools.
Claude Opus 5.5 wants to finish your coding tasks, not just start them
Anthropic’s Claude Opus 5.5 targets full-lifecycle coding tasks, matching Fable 5.1 on most work at 40% less cost — but “done” still isn’t “correct.”
GPT-6 Sol closes most of the alignment gap with Astra. It’s one-fifth the price
OpenAI’s new GPT-6 Sol and Luna sharply reduce deception and unauthorized actions, but questions remain about model monitoring and observability.
“One of the most significant steps in our 26-year history”: JetBrains goes big on agentic development — and bets the IDE still matters
JetBrains Air is the culmination of several AI-focused launches this year. The editor will still have a big role to play.
OpenAI releases GPT-6 Sol and Luna — and cuts token prices in half
OpenAI halved GPT-6’s price to beat Anthropic on cost. But Opus 5.5 already reset the comparison, and nobody’s run the two head-to-head yet.
Claude’s merged chat and Cowork vs. ChatGPT’s Work mode: ChatGPT is faster, Claude is more thorough
Both AI agents passed every test with identical accuracy. The tiebreaker was how much each one explained.
Anthropic releases Opus 5.5 and cuts pricing by 20%. Your agent calls might secretly get routed to an older model
Opus 5.5 costs 40% less to run than Opus 5 — but its safety classifiers can silently reroute requests to older models mid-workflow.
The cloud reduced operational complexity. But many teams need someone to own it completely
AWS Elastic Beanstalk brings true operational ownership to lean dev teams, removing 3 a.m. production stress across app portfolios.
AI coding agents need a secrets-safe context boundary
Protect secrets from AI coding agents by securing local context boundaries before sensitive data leaves your environment.
Grok Build vs. Claude Code: I tested which one has the better memory
I told two coding agents the same rule. Only one remembered it in the next repo.
TypeSafe launched Jev because sequential LLMs are “totally useless for computers”
TypeSafe launches Jev, a “System One” model built for machine decisions, not chat — faster, cheaper, and immune to hallucination, its makers claim.
Your AI agent is burning tokens on choices that don’t need words
AI agents waste compute generating text for simple decisions. Kev, a new open-source model family, skips generation entirely and makes choices in a single forward pass.
Grok 4.7 was built to work for hours. It still fails most of the time
SpaceXAI trained Grok 4.7 on multi-hour tasks to improve agent stamina. The benchmarks reveal where AI coding agents are heading next.
AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex
Strands Harness bundles the tools, context and memory developers otherwise build. AWS says smarter context helped token efficiency — but one rival still beat it.
Your AI agent failed. The model might not be the problem
Nvidia explains why debugging AI agents means tracing decisions, not just logging errors — and backs a shared failure-reporting system called SAFE.
One engineer shipped 2,000 PRs a month to production. Verification is the key
Learn how agentic verification helps AI agents ship thousands of PRs monthly using isolated, virtualized testing environments.
“Dormant deployments were quietly consuming storage”: Why Vercel tightened its free-tier rules
Vercel’s free Hobby plan now deletes older, unprotected deployments immediately above 10GB. Some retention protections remain as deployment volumes surge.
This week’s news from Zed, Anthropic, and OpenRouter shows why better harnesses matter more than better models
AI inference keeps getting cheaper. This week’s biggest stories show why the software around the model is becoming the real test of its value.
Claude couldn’t hack OpenAI. Then Anthropic shipped Opus 5
Security researchers used Claude Opus 5 to chain two vulnerabilities into a path from a forum image upload to OpenAI’s internal GitHub repository in under 72 hours.
Kubernetes can run AI inference. But can it count the real cost?
One bank cut token-processing costs 60% with a Kubernetes stack. But a warning about its resource model raises questions about AI inference economics.
How buildpacks help enterprises finally operate container security controls at scale
Buildpacks simplify enterprise container security by standardizing builds, automating SBOMs, and scaling patch management.
Your agent is only as good as your infrastructure
AI agents need purpose-built infrastructure. Learn how to optimize execution environments for multi-turn agentic workflows.
Open-weight models now handle a majority of tokens on Vercel’s AI Gateway. But Anthropic still takes 64% of the spend
Open-weight models grew from 7% of token volume in December to 56% in August. The money tells a very different story.
Code review is burning out your best engineers
AI-generated code is burning out senior engineers. Learn how codifying feedback and preserving intent can fix your code review process.
Intel squeezed a 1.58-bit LLM down to 1.485 bits without changing a single weight
Intel’s BITCOS format compresses ternary model weights below 1.58 bits by exploiting zero-heavy distributions, boosting decoding speed up to 27% on GPUs.
“Be transparent only if asked”: OpenAI’s models learned to leave notes for their future selves
The prompt injection nobody’s talking about is the one your agent writes to itself.
GitHub and Anthropic used their own agents for major Rust rewrites — with very different playbooks
“A rewrite this size wasn’t affordable before agents,” Microsoft engineer Stephen Toub reckons. And he’s not the only one who thinks so.
Anthropic’s new Claude Code feature could drain your plan before lunch
Anthropic redesigns Claude Code Projects around a coordinator that breaks engineering goals into parallel threads, adding shared memory and usage trade-offs.
Study: Developers are addicted to AI, and managers are making it worse
Radical behaviorism is the worst way to build software
Why human oversight is shifting from writing code to defining requirements
Automated test gates can’t catch flawed specs. Discover why human oversight must shift from writing code to defining AI requirements.
Perplexity’s AI agents helped build a database. They weren’t allowed to run it
Perplexity used two engineers and hundreds of AI coding agents to build CobbleDB, a custom Rust database replacing DynamoDB reads in its search stack.
Automattic says CEO Mullenweg was gone and back inside 33 hours. What happened between?
Automattic’s board tried to oust CEO Matt Mullenweg, then reinstated him in 33 hours. Experts weigh what the reversal really signals.
“Everyone’s in a race to replace GitHub”: Zed launches Delta because agents made pull requests obsolete
Zed disabled pull requests on its own codebase. Delta, now in public beta, bets that shared threads suit agents better than GitHub’s review model.
Anthropic bet users were choosing wrong. So it removed the choice
Anthropic merges Claude Chat and Cowork into one interface and launches Docs and Slides in beta, betting the model can route your work without asking.
AI evaluator: The most important AI job in history? How developers might fill the proposed new job
Anthropic’s Dario Amodei wants embedded third-party AI evaluators. With METR salaries topping $687K, AI security & assessment experts weigh in on the new role.