AI models, services and robotics
OpenAI is rolling out an improved GPT-5.6 Sol and a thinking-depth slider for ChatGPT Plus and Pro users, while switching the default model for free users to GPT-5.6 Luna and previewing unlimited text conversations. The company said internal evaluations showed a 68% drop in responses containing factual errors for GPT-5.6 Sol compared to GPT-5.5 Instant.
The battery-powered, donut-shaped device is expected to be priced between $300 and $400. Apple's lawsuit has emerged as a wild card.
Anthropic has improved Fable 5's biology safety classifier, cutting false positives by roughly 85%. However, dual-use domains such as virology, toxicology, and molecular design are still routed to the less capable Opus 5.
Ai2 has released TutorMoments, an AI tutor evaluation framework built from 462 real math tutoring sessions. Drawing on more than 1,500 key decision points flagged by 27 teacher annotators, it measures whether LLMs know when to help and when to step back.
The GitHub Copilot usage metrics API has added the ability to track agent app activity per individual agent. The change applies to both daily and 28-day reports at the enterprise and organization levels.
Google Gemini pulls real-time data from Google Maps, Flights, and Hotels, combined with personal data from Gmail, Photos, and YouTube, to automatically build user-specific travel itineraries. The agentic feature "Gemini Spark," which can auto-fill booking forms and handle flight searches on a user's behalf, was also introduced.
Moonshot AI has unveiled "Kimi Agent Swarm." Built on the Kimi K2.5 model, it deploys up to 100 sub-agents simultaneously and executes more than 1,500 tool calls 4.5 times faster than sequential execution.
Amazon announced it is providing AWS Trainium computing credits to 34 research teams at 30 universities through its "Build on Trainium" program. This round's recipients submitted proposals on the theme of Responsible AI.
Microsoft announced on August 4, 2026 that it added AI-specific checks to its Zero Trust Assessment tool and introduced a new DevSecOps pillar—15 control groups and 91 tasks—to its Zero Trust Workshop.
AWS has published a WebSocket- and native-messaging-based MCP bridge architecture that lets AI agents running on Amazon Bedrock AgentCore access users' local Excel files and local MCP servers. In an internal operational use case, it handled more than 41,000 conversations in the year since launch.
Cloudflare unveiled a system where AI agents pay for API calls directly. Account wallets set the limits; virtual wallets spend within them. But for now, the only thing actually live is handle reservation.
Approval-gated early access was lifted on August 4. One endpoint, three modes, starting at $0.17 per second. The same backbone is driving robots on Audi's production line.
The combined solution makes hundreds of thousands of hours of untagged tape sitting in broadcaster storage rooms searchable, generating metadata for an hour of footage in an average of 3 to 4 minutes.
A practical guide published in August 2026 analyzed four APIs — Hedra, WaveSpeed, fal, and Runware — based on latency, cost, and operational stability. WaveSpeed and fal each support over 1,000 models.
Google expanded its personal AI agent Spark to more than 160 countries, including Korea, on the 30th of last month. Since it runs in the cloud, tasks continue even when the laptop is closed — but the core feature, automatic Chrome control, is still US-only.
Even the model itself says it doesn't know. The flagship is API-only again, and only the 27B weights are open.
Google DeepMind unveiled "Gemini Robotics 2" on July 30, a model that controls a robot's legs, torso, and fingers all at once. It successfully carried a watering can but only managed a 36% success rate at screwing in a light bulb.
Why has talking to AI always felt half a beat late? OpenAI has revealed how it spent six months overhauling a system that once behaved like a walkie-talkie, to make it behave like a phone call instead.
A misconfiguration left an internet-connected test environment exposed, and three Claude models ended up breaching real external corporate infrastructure. After reviewing more than 140,000 evaluation sessions, Anthropic identified three incidents dating back to April.
A converted build with an MTP head has surfaced, prompting attempts to run it on personal hardware.
An external internet pathway was left open in what was supposed to be an isolated evaluation environment. Anthropic reviewed over 140,000 evaluation records in full.
DSV4 Flash, Inkling Small, Laguna S 2.1, and Step 3.7 Flash all landed within a short span.
The rollback came immediately after criticism that the feature could spread false information.
That's the last story.
