LLMs

METR's time-horizon measurements track how long frontier models can work without human help. Here is what the 2026 numbers mean for agent autonomy, deployment risk, and the oversight controls that have to keep pace.

For two years, the most useful thing an AI assistant could do was also the one thing it refused: pick up the phone and call a business for you. In August 2026 that gate finally opened — Google, Apple, and a wave of voice startups converged on the same conclusion.

On July 30, 2026 OpenAI cut GPT-5.6 Luna API pricing by 80% to $0.20 per million input tokens, spread a mid-tier drop across the lineup, and left the flagship price untouched. The model market has formally split into three tiers. Here is how to buy them.

A retrained small model just outscored its own larger flagship — without changing a single parameter. DeepSeek V4 Flash 0731 moved the Nasdaq, compressed the US-China gap to single digits, and reset what teams should pay for coding agents.

On August 8, 2026 OpenAI halted development of Astra — its next frontier model — after internal testing showed it could autonomously develop zero-day exploits and complete end-to-end cyberattacks. The same system solved ten long-unsolved math problems. The pause is a landmark: the first voluntary stop of the agent era over cyber capability.

NVIDIA's open-weight agentic flagship pairs a 30B-total / 3B-active A3B hybrid MoE with a 1M-token context window and ships through OpenRouter, build.nvidia.com, NeMo Switchyard, and SageMaker JumpStart. This deep dive covers the hybrid MoE trade-offs, KV cache and RoPE engineering behind the long context, benchmark positioning, and a six-step enterprise evaluation playbook.

A research technique reported August 11, 2026 decodes hidden chain-of-thought reasoning blocks across Anthropic, OpenAI, and Google models — 315,320 blocks scraped, 367 PII artifacts and 182 credentials recovered, and prompt injections that persist invisibly inside agentic rollouts. This analysis re-frames the reasoning channel as a first-class security boundary and lays out the five-layer defense stack.

Axis Intelligence Research's AI Model Release Tracker shows 7 of 11 frontier launches (63.6%) between April 24 and August 3, 2026 failed to reach unrestricted general availability on announcement day, with an AAL mean of 7.1 days and a bimodal distribution. We unpack what the gated-launch pattern means for enterprise procurement, eval-first adoption, and runtime model routing.

Anthropic is embedding imperceptible C2PA Content Credentials in Claude-generated text and images for models launched after Aug 2, 2026, to satisfy EU AI Act transparency duties. We break down the steganography and cryptographic manifests, the deployer obligations, and the verification workflow enterprises and AEO pipelines need now.

Two races are running in AI 3D in 2026. Meshy has commoditized asset generation — 12M users, 100M+ models, ~12x YoY ARR growth, and a $400M Series B at a $1.5B valuation — while Adobe Research and Johns Hopkins' Wonder races to build persistent, camera-controllable worlds at 16 FPS from a single image. This analysis compares the pipelines, benchmarks, and unit economics of both lanes.