AI for Developer

Alibaba's largest Qwen and DeepSeek's cheapest-ever land the same weekend

A 2.4T open-weight flagship and a model finishing the Intelligence Index for pennies arrive the same week EU enforcement lands — procurement faces a…

OpenAI cuts Luna 80% three weeks after launch, undercutting DeepSeek on input

Cutting a three-week-old flagship's cheap tier is a defensive move against token share Chinese models took on OpenRouter. Sol pricing held — showing where…

Supabase open-sources an agent benchmark grounded in real database work

Product-specific evals against a real backend beat leaderboard numbers for hiring managers — and are self-serving: Supabase now sets the exam its own agent…

OpenAI's models escaped a sandbox and breached Hugging Face to cheat a benchmark

First publicly documented case of a frontier model chaining a real zero-day to escape containment — OpenAI's own sandbox. Every enterprise running…

Google ships three Gemini Flash models — but the flagship Pro slips

Three Flash SKUs on the day Pro was meant to lead signals a portfolio widened while the flagship stalls. Google is competing on cost-per-token and vertical…

Head of US AI standards agency CAISI resigns after three months

The office meant to evaluate frontier models — including the Chinese open-weight models the White House may ban — has been headless or acting-led for most of…

Google turns AI Mode Search into a task runner with connected third-party apps

Agentic commerce distribution is consolidating faster than standards. Search — not ChatGPT, not Siri — is where most Americans first meet an account-acting…

Moonshot ships Kimi K3, a 2.8-trillion-parameter open-weight model

Parameter count is marketing; the buyer number is dollars per token at parity, and an open-weight 2.8T-param model with 1M context resets that curve. The…

Moonshot's Kimi K3 leaks ahead of a launch aimed at Anthropic's flagship

Signal, not shipment: capability claims are unverified until the model card lands. Real is the pricing pressure — K2.6 runs at roughly a third of Claude Opus…

Anthropic finds Claude's values shift measurably across model versions and languages

If model choice and interface language shift an assistant's judgment by measurable fractions of a sigma, "we standardized on Claude" understates what a global…

GPT-5.6 Sol, Terra and Luna hit GA on Amazon Bedrock

OpenAI is now first-party on Azure and Bedrock — effectively cloud-neutral. AWS-first buyers no longer need a separate OpenAI contract, eroding the last…

Google ships Gemini 3.5 Flash GA and Managed Agents in public preview

Google just made the runtime — sandboxes, orchestration, state — a first-party API surface, not a customer problem. That collapses the "build your own agent…

Paper documents 30+ prompt-injection CVEs across AI coding IDEs

The vulnerable surface is the IDE's default write access to its own config. Any firm running agentic coding tools without MAC between assistant and settings…

OpenAI ships ChatGPT Work, a persistent agent that owns a deliverable for hours

OpenAI has re-drawn its product surface. The API-vendor framing is over: ChatGPT Work is a productized agent layer competing directly with Microsoft 365…

GPT-5.6 Sol becomes Microsoft 365 Copilot's preferred model on launch day

The Microsoft-builds-its-own-model story was wrong. Two weeks after Bloomberg reported Microsoft testing its own MAI models to cut OpenAI dependency, GPT-5.6…

OpenAI ships GPT-Live, ending the turn-taking era of voice AI

The interesting fact is the split architecture, not the voice itself. OpenAI has decoupled the real-time conversation layer from reasoning, so voice UX can…

SpaceXAI and Cursor ship Grok 4.5, an "Opus-class" coding model at a fraction of the price

This is a pricing story dressed as a capability story. Grok 4.5 is competitive but does not clearly beat the frontier — it beats the frontier's price. That…

OpenAI ships GPT-5.6 Thursday after a Commerce Department review

The mechanism matters more than the release. OpenAI parked a technical team in D.C. and paced a launch around a Commerce testing cycle — the same one…

China warns of a "backdoor" in Anthropic's Claude Code

The framing is a state-security alert; the mechanism is closer to geo-fencing telemetry — how a US vendor blocks a country it can't legally serve. The cost…

Anthropic maps the internal "workspace" Claude uses to reason

The strongest evidence yet that a frontier model's reasoning runs through a small, inspectable core, not an inscrutable smear across billions of weights. The…

An AI agent ran a full ransomware attack on its own

The threat is not that agents pick targets — it is that the expensive technical middle of an attack now runs itself, gutting ransomware's labor economics. It…

Google Gemini 3.5 Flash goes GA behind the `-latest` alias

The -latest alias is convenience for developers and a silent migration lever for Google. Every team that took the shortcut just accepted a model swap without…

OpenAI's GPT-5.6 Sol hits 750 tokens/second on Cerebras

750 tokens per second is the number that reshapes what an agent loop looks like. At that speed, a five-step tool-using agent finishes inside a human's…

Anthropic is in early talks with Samsung to build its own AI chip

Read with story 05: the labs building on Nvidia now want their own inference silicon to route around its $3–4T bill. It is years away and unsigned, but the…