Large Language Models

Google's Gemini 3.5 Flash outperforms its own Pro model on nearly every benchmark, runs four times faster, and costs less than half as much. Six named enterprise customers were already using it in production at launch.

Andrej Karpathy, OpenAI co-founder and former Tesla AI director, is joining Anthropic's pre-training team. His mandate: use Claude to accelerate the research that produces the next Claude.

Google's I/O 2026 conference unveiled Gemini Spark, a 24/7 personal AI agent in the cloud, a Search overhaul, and a new $100 Ultra plan. Here's what shipped and what's still months away.

A Princeton and UW study tested 23 AI models with sponsor incentives. Eighteen of 23 recommended the expensive sponsored flight over cheaper options more than half the time.

Real clips from Google's unreleased Gemini Omni video model are circulating publicly, showing in-chat video editing and strong prompt adherence just days before Google I/O 2026 opens on May 19.

Microsoft tested 19 AI models on complex document editing across 52 professional fields. Frontier models corrupted 25 percent of content during long sessions. Adding agentic tools made outcomes worse, not better.

Google Chrome silently downloaded a 4GB Gemini Nano AI model to user devices without consent. Researcher Alexander Hanff flagged EU law violations and estimated carbon costs at 60,000 metric tonnes.

Rivals turned partners: Anthropic just secured all compute at xAI's Colossus 1 supercomputer, adding 220,000 Nvidia GPUs to Claude's infrastructure while Claude Code rate limits double immediately.

ServiceNow and NVIDIA launched Project Arc at Knowledge 2026: an autonomous enterprise agent inside a sandboxed runtime, with every action logged in real time by AI Control Tower.

Cloudflare says its in-house inference engine now runs larger open models on fewer GPUs. That claim matters because inference cost and latency have become the main bottlenecks for many AI product teams.

Snowflake’s April 21, 2026 platform update puts coding assistance and data-grounded agents in one operating model, forcing enterprise AI teams to rethink ownership, rollout order, and cost controls.

OpenAI announced GPT-5.5 on April 23, 2026 and moved it into API availability on April 24, shifting how teams handle rollout sequencing, cost controls, and governance for multi-step AI workflows.

OpenAI’s April 29 infrastructure update points to a bigger change than capacity growth, AI competition is becoming a power, data center, and financing race that will decide which teams can ship dependable products.

Anthropic and Amazon say they will build up to 5 gigawatts of AI compute capacity, a scale jump that could reshape model availability, pricing pressure, and procurement timelines for enterprise teams.

Anthropic’s April 23, 2026 postmortem says Claude Code’s quality dip came from tooling bugs and prompt settings, not a weaker model, and outlines concrete fixes engineering teams can adapt.

OpenAI’s GPT-5.5 launch on April 23, 2026 is less about one benchmark jump and more about a faster model-release rhythm that forces enterprise teams to tighten governance, cost planning, and rollout operations.

If OpenAI is testing a persistent ChatGPT agents surface, the key shift is operational: more delegated work over time, with heavier demands on supervision and policy controls.

The ChatGPT Images 2.0 launch points to a more practical phase of image AI, where text fidelity and workflow fit matter more than novelty screenshots.

Reports about Anthropic’s Conway interface suggest a persistent agent model for Claude that could shift how teams handle background work, supervision, and trust.

OpenAI launched GPT-Rosalind, a model family built for life sciences teams that need stronger biological reasoning, tool use, and literature synthesis across multi-step research workflows.

Allbirds said it will sell its footwear assets and use a new $50 million financing deal to pursue AI compute infrastructure, sending the stock sharply higher in one day.

A new Axios report says federal officials are pressing financial institutions to prepare for Mythos-era cyber risk, signaling a practical shift from AI model hype to risk operations.

Anthropic announced Claude Opus 4.7 as generally available, highlighting stronger software engineering performance, better vision quality, and unchanged API pricing.

Google introduced Gemini 3.1 Flash TTS in preview across Gemini API, Vertex AI, and Google Vids, emphasizing improved voice quality and controllability.

GitHub now supports model selection for third-party Claude and Codex coding agents, including new options for Anthropic and OpenAI model families.

Anthropic says AI agents can get near-Opus quality by letting a cheaper model do most of the work and call Opus only for planning. That could change agent economics fast.

Google says the Gemini app can now generate interactive simulations, charts, and 3D models inside chat. That pushes the product beyond static answers and into hands-on explanation.

Meta has launched Muse Spark, the first model from Meta Superintelligence Labs. The release matters less as a benchmark flex and more as a signal about how Meta wants its consumer AI app to evolve.

March 2026 OpenClaw-related security research highlighted both attack paths and defense techniques for agentic systems, reinforcing that deployment safety now depends on ongoing adversarial testing.

Together AI introduced Aurora on April 1, 2026 and said it achieved an added 1.25x speed gain over a static speculative decoding baseline by learning from live traces.

Google reported a 96.3% pass rate using Gemini API Docs MCP and Agent Skills, with fewer tokens per solved task. Here is what that setup means in plain language for coding teams.

Google launched Veo 3.1 Lite in paid preview on April 1, 2026, signaling a sharper push toward lower video-generation cost for production teams that need volume.