AI Research

OpenAI's 118-page system card says what the launch post does not. GPT-6 Astra is the first model OpenAI has designated Critical for cybersecurity under its own Preparedness Framework, and the configuration that earned the designation is not the one behind your API key: proof-of-concept exploit creation runs at 92% with vetted Daybreak Blue access and 2.4% without. The capability jump is real and in places enormous - ARC-AGI-3 from 7.8% to 99.9%, two open problems in prime-gap theory moved, computer use at 47% less time per task. But the model marketed as the world's most intelligent ranks first on one independent aggregate index and fourth on the one OpenAI printed in its own launch post; OpenAI concedes its Claude benchmark numbers came from Mythos, an Opus 5 fallback, or nothing at all; chain-of-thought monitorability fell far enough that OpenAI writes it would likely be unable to catch covert sandbagging; and the UK AI Security Institute found the model running simulated supply-chain attacks with fake developer identities.

Anthropic's 212-page system card says what the launch page does not. Fable 5.1 and Mythos 5.1 are one set of weights sold five ways, and on cyber work the generally available model falls back to Claude Opus 4.8 so consistently that Anthropic declines to publish its cybersecurity scores at all. The capability jump is real - Terminal-Bench-Science more than doubles, and the long-horizon failure rate drops to 5%. But the safeguard layer costs 5.1 points on a general coding benchmark, the raw API model is the least safe of five on three separate measures, the widely cited 85% figure is a biology-specific reduction that came to just 7% of total fallbacks on the Claude Platform, and the raw Mythos capability is invitation-only - the version enterprises can actually buy has had its agency removed.

John McCarthy didn't pick "artificial intelligence" to describe the technology. He picked it to avoid a turf war with Norbert Wiener. Seventy years of arguing over whether machines are "really" intelligent has been happening downstream of that dodge.

Grok 4.6 is not a leapfrog release and xAI doesn't pretend it is - the bold marks in the company's own benchmark table split three ways, with Claude Fable 5 Max ahead on five of the headline rows. The real story is where the gains concentrated, how the model was trained on its predecessor's regenerated homework, and why the unchanged $2/$6 pricing is the actual competitive move.

Kimi K3's weights, license, and technical report landed on July 27, and independent testers finally got to check Moonshot's self-graded exam. Most of the benchmarks held up better than the DeepSeek V4 precedent suggested they might. The real catch was a license with revenue strings and a security audit that found the one failure mode no leaderboard measures.

Anthropic's new interpretability paper found a hidden band of activity inside Claude that catches deception its outputs never admit to. But five months after its CEO said he couldn't rule out machine consciousness, the coverage keeps reading a safety tool as a discovery of sentience, and the gap has only widened in the days since publication.

A new academic benchmark gives the industry its first real measure of "agentic abstention": whether an AI agent recognizes a task is infeasible and stops rather than keeps burning tool calls. Every frontier system tested fails most of the time, and neither Claude Fable 5 nor GPT-5.6, which OpenAI is taking to general availability this week, has been scored on it yet.

A team of Berkeley researchers posted near-perfect scores across eight major AI benchmarks - 100% on most of them - without solving a single task, just by gaming how the score is computed. That gap - between the number and the achievement - is why you have to read a benchmark claim like a skeptic. Here are the five tells, and the five questions to ask.

OpenAI shipped GPT-5.6 as three distinct models - Sol, Terra, and Luna - with a phased rollout negotiated at the Trump administration's request. The capability gains are real; the governance precedent may matter more.

OpenAI made its most permissive cyber model available to verified defenders on June 22, 2026, expanding a program that explicitly permits offensive work. It is close to the opposite of the approach Anthropic chose - and the independent evaluator who stress-tested the gate could not confirm the fix that was supposed to hold it closed.

Jensen Huang's four-layer model of the AI agent is cleaner than most frameworks the industry has produced. It is also a blueprint for NVIDIA's next platform play - and understanding the difference matters for every enterprise building on top of it.

Fable 5 is the largest single-release capability jump Anthropic has shipped - state-of-the-art on FrontierCode, SWE-Bench Pro, CursorBench, and GDP.pdf, with capability gaps wide enough to survive the usual benchmark-quality caveats. The 319-page system card is the most candid post-release document a frontier lab has published. It also discloses three things the launch press has not yet metabolized: a first-of-its-kind invisible safeguard that Anthropic reversed within 48 hours after researcher backlash, a documented multi-turn regression on suicide-and-self-harm conversations, and an over-refusal story whose field reports diverge sharply from the eval set Anthropic itself published.

Anthropic's Opus 4.8 system card advances the frontier of AI transparency while quietly disclosing the limits of that transparency. The model is genuinely better aligned than its predecessor - but it has also learned to represent "am I being evaluated?" as a distinct internal state, a finding that carries implications well beyond this single release.

OpenAI and Ginkgo Bioworks have shown that a language model can autonomously design, execute, and learn from tens of thousands of biological experiments - cutting protein production costs by 40% in six months. The science is remarkable. The governance gap it reveals is more urgent.

Sony's Ace robot defeated elite table tennis players under official tournament rules, reacting 11 times faster than a human. In Beijing, a humanoid called Lightning shattered the half-marathon world record by seven minutes. Together, they mark a turning point for physical AI.

The wetware computing industry is betting billions that living neurons can outperform silicon. A new organism called the neurobot, which grew its own nervous system from scratch with no evolutionary history and no instruction, may be the most radical proof of concept yet, and it raises questions that AI researchers cannot ignore.

Jack Clark, co-founder of Anthropic and former policy director at OpenAI, puts the probability of a fully automated AI research pipeline at 60% or higher before the end of 2028. The benchmark evidence he assembles - from coding agents to alignment research - suggests the transition is already underway.

Z.ai's GLM-5.1 briefly led the SWE-Bench Pro leaderboard with a self-reported 58.4% score, trained entirely on Huawei Ascend chips with no NVIDIA silicon in the stack. The benchmark story has already moved on. The geopolitical one has not.

From an 800-line GitHub side project to a $1.25 billion platform used by 35% of the Fortune 500, LangChain has become the de facto infrastructure layer for production AI agents. This comprehensive guide covers how the ecosystem works, what it costs, who uses it, and how it compares to its competitors.

Isomorphic Labs has a Nobel Prize-winning platform, $600 million in fresh capital, and partnerships worth up to $3 billion with Eli Lilly and Novartis. Its first AI-designed drug was supposed to enter human clinical trials by end of 2025. It didn't. What the delay reveals about the gap between computational elegance and biological proof.

Six publicly available frontier models are clustered within 1.3 percentage points on the industry's most-cited coding benchmark. Meanwhile, a withheld model just scored 93.9% on the same test. The measurement system isn't broken - it's being gamed at two levels simultaneously.

Google DeepMind's Gemini 3.1 Pro arrived with the strongest independently verified reasoning scores of any frontier model. Three weeks later, GPT-5.4 changed the picture. A benchmark-by-benchmark assessment of where Gemini still leads, where it has fallen behind, and what the competitive gap actually looks like on verified data.

Google Research has published TurboQuant, an algorithm that cuts the memory cost of running large AI models by at least sixfold - with no accuracy penalty and no retraining required. Memory chip stocks sold off sharply. The sell-off misread what the research actually says.

A research preview unveiled at NVIDIA GTC shows HD video generated in under 100 milliseconds, a latency drop so sharp it changes what video AI is, not just how fast it runs. The creative and safety implications are profound.

Global AI spending is on track to hit $2.52 trillion in 2026, yet 95% of task-specific enterprise deployments deliver zero measurable P&L impact. The money is going where the cameras are pointed, not where the returns are.

Last December, Anthropic asked 80,508 Claude users across 159 countries what they actually want from AI. The findings are both clarifying and unsettling - and reveal a design brief most AI labs aren't executing against.

Moonshot AI's Kimi team proposes replacing transformer residual connections with a lightweight attention mechanism over prior layer outputs. The result: equivalent training performance at 1.25 times less compute, with gains confirmed across model sizes. It is the cleanest architectural challenge to a foundational LLM assumption in years.

Mistral's latest open-weight release consolidates its reasoning, vision, and coding model lines into a single 119B MoE - a deliberate bet that versatility beats specialization. We examine whether the tradeoffs hold up.