Recent stories
Each story's summary, with the places it was discussed.
Summaries are written by AI from the linked sources. Open the source for full context.
Google introduced Gemini 4 Hydrogen, a fast, efficient model for everyday coding and agentic workflows, available across developer, enterprise, consumer, and Search products.
It scores 51 on the Artificial Analysis Intelligence Index at default high thinking, reaches 54 at xhigh, and costs $1.50 per million input tokens plus $7.50 per million output tokens across its 1-million-token context. In Google Antigravity, Orchestrator achieved near-Carbon success while using 24% less quota and running 19% faster.
Gemini 4 Carbon is now available
Artificial Intelligence · Large Language Models · Gemini · AI Agents
Google has made Gemini 4 Carbon available to developers, enterprises, scientists, and consumers as its strongest model for cybersecurity, long-horizon agentic workflows, complex software engineering, and scientific reasoning.
It ranks first on CursorBench 4.0, WANDR, and AutomationBench, with over 2× lower cost per task than the next external model. Carbon supports Google Antigravity’s /swarm and costs $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory price. It also reports stronger safety, vulnerability discovery, and scientific results.
Gemini 4 Argon: Built to solve your hardest problems
Artificial Intelligence · Large Language Models · Gemini · AI Agents
Gemini 4 Argon is now globally available through Google’s API, AI Studio, Gemini platform, Gemini app, and Search.
It targets complex coding, business, cybersecurity, science, and robotics tasks, with a 1M unified input-output token limit and introductory pricing of $2 per million input and $10 per million output tokens. Google reports #1 on the Vals Index, 35% more agentic code submissions, major optimization gains, and an 87% ERQA score. It includes safeguards against misuse, agent misalignment, and prompt injection.
Google introduces Gemini 4 Argon, a physical reasoning model for robotics spanning control, Real2Sim, physical understanding, and research.
It scores 87% on the Embodied Reasoning (ERQA) benchmark and supports MuJoCo practice, zero-shot control, simulation reconstruction from robot or human video and sketches, hindsight labeling, value estimation, and self-improvement. In simulation-data testing, 4,500+ episodes raised real-world Success Rate from 56% to 75% and cut completion time from 35 seconds to 24 seconds. Argon is available via Google AI Studio, Gemini API, and Google Antigravity.
Google has dramatically increased its Chrome security patching rate using LLMs like Gemini, fixing over 1,000 bugs in one month.
This trend of AI-driven vulnerability discovery is transforming cybersecurity, as evidenced by similar initiatives at Microsoft. This marks a shift toward industrial-scale automated security, outpacing manual efforts and allowing companies to address risks preemptively.
At AGNTCon Europe, the author demonstrated WebMCP, an experimental browser API that lets AI agents call website-exposed tools without rebuilding existing products.
Their WebMCP Local Agent Chrome extension supports Google's Prompt API, Ollama, and cloud Gemini, including local inference and configurable loops. Tool annotations require confirmation for consequential actions, highlighting a practical, privacy-conscious path to agent-friendly websites.
How Google Duplex evolved to also call businesses for you in Gemini
Artificial Intelligence · Large Language Models · History · Gemini · AI Agents
Google integrated Duplex into Tasks in Gemini, enabling the AI to call businesses and coordinate real-world tasks such as reservations, appointments, and service requests.
Over nearly a decade, Duplex evolved from closed-domain recurrent neural networks to hybrid systems and now a unified Gemini stack using free-text task prompts, multimodal planning, and dialogue generation. The system addresses noisy phone lines and sub-second latency, discloses that it is automated, and lets businesses opt out without installing software.
Google announced the winners of the I/O 2026 Writing Challenge, honoring creators who provided insightful analyses of new tech like AppFunctions, Gemini 3.5 Flash, and the Managed Agents API.
The winning entries focused on practical experimentation, architectural shifts, and deep-dive technical explorations of Google's latest developer tools and local AI capabilities.
Google introduced Gemini Robotics 2.5, a vision-language-action (VLA) model that expands robots’ dexterity, reliability, speed, and safety across more than 200 tasks.
Compared with previous baselines, it raises success rates by 30%–100% and completes tasks 20%–30% faster. In a Mercedes-Benz proof of concept, Apptronik’s Apollo 2 achieved 96% success sorting automotive parts. Gemini 4 Argon accelerated development through simulation, while new safety guidance and ASIMOV-Jailbreak testing improve operation around hazards, people, and adversarial prompts.
Google launched Gemini Omni 1.2 Flash across the Gemini API and Gemini for business, with general availability and a 30% price reduction to $0.07 per second for 720p video.
It generates and edits clips up to 15 seconds, supports multi-turn conversational changes, better reference adherence and frame-to-frame consistency, more natural multi-speaker speech, and smoother cinematic camera motion. Workflows can draft at 360p/720p, then upscale to 1080p/4K; outputs include SynthID watermarking.
Google has updated the Gemini API File Search tool to support multimodal RAG, enabling native processing of text and images.
With new features including custom metadata filtering for efficient data retrieval and page citations for source verifiability, developers can build more accurate, context-aware AI applications with improved transparency.
Avian Visitors is a fork of BirdNET-Pi that tracks local bird populations using a Raspberry Pi, a USB microphone, and AI-powered acoustic classification.
It features an automated collage layout of species sightings styled as kachō-e woodblock prints using Gemini API. The project supports integrations with Home Assistant, Cloudflare, and MQTT for real-time monitoring.
Modern technology products often suffer from 'over-improvement' by prioritizing corporate metrics over user utility.
The user criticizes Google Home's evolution into Gemini, noting it has become slow, overly cautious with medical queries, and less effective at simple tasks, ultimately leading them to switch to Alexa out of frustration with declining product quality.
DEV’s Dog Days Edition recognized five projects: analysis of 99,916 shelter records challenged black dog syndrome; Snowflake detected canine limping patterns; Solana secured lost-pet rewards; ElevenLabs turned shelter records into voices; and Gemini supported dog-adoption matching.
Winners receive $200, DEV++ Membership, and a badge, while future challenges with Sanity and Kaggle are announced.
DevGraveyard is a creative platform where developers bury abandoned projects using GitHub data.
It combines personality-driven epitaphs and AI-generated eulogies via Gemini with a 3D Three.js environment. Users can mourn, leave messages, or vote to resurrect projects, turning forgotten code into a poignant, interactive memorial.
Letters to Tomorrow is a narrative web game created for the June Solstice Game Jam.
Players act as a postmaster helping townspeople finish poignant letters reflecting on life, kindness, and personal growth. Developed as a single-page app, the game emphasizes privacy and accessibility, using Gemini as a development partner to refine its reflective, storytelling-driven experience.
An inexpensive RAG system in Go uses Gemini File Search as a hosted, API-key-scoped corpus, avoiding a vector database and custom embedding pipeline.
Two model calls abstract a proposal into pattern-based retrieval queries, then produce a cited review. Markdown preprocessing, SHA-256 incremental ingestion, SQLite manifests, key-aware failover, deterministic validation, and law-level prompt traceability improve reliability.