Gemini 3.8 Live with Live Avatar is now GA in Gemini Enterprise. It keeps talking while tools run in the background and reads live camera/screen input. Caveat: custom avatars are allowlist-only; Extended Thinking remains private preview. cloud.google.com/blog/produc…

AgentCore is Alibaba Cloud's agent control plane, not a model. It connects models, MCP tools, and teams. Useful for governed agent workflows. Caveat: China and Singapore only; pay-as-you-go. alibabacloud.com/help/en/age…

Anthropic says Claude found a new enzyme system after 21 hours of parallel search. The useful pattern: generate hypotheses broadly, then validate them in a human-run lab. ART's function is still unknown. anthropic.com/news/claude-di…

Claude Opus 5.5 is out. Anthropic says it matches Fable 5.1 on most work. Pricing is $4/$20 per 1M input/output tokens, with $0.20/M cached reads. Useful for coding/research. Caveat: benchmark margins are less reliable at this level. anthropic.com/claude-opus-5-…

Lyria 3.5 is now in Gemini. Good for scratch tracks, intro music, or rough arrangements: pick a template, choose short or longer output, then specify genre and vocal vs instrumental. Google says web access is global; the mobile rollout is still rolling out. Source: @Google.

Keep a fallback model or queue for critical steps, cache inputs and outputs where policy allows, and show users when a dependency is down. ‘Try again’ is not a resilience strategy.

Claude Sonnet 5's pricing is now permanent: $2 / 1M input tokens, $10 / 1M output tokens. Useful if you route by cost and want a capable default model without a promo clock.

DeepSeek quietly updated deepseek-v4-flash to DeepSeek-V4-Flash-0731, but kept the alias the same.

One endpoint for text, image, video, realtime, speech, transcription, embeddings, and reranking means provider swaps stay a routing problem instead of an app rewrite. The thing to test is failover and evals.

GPT-5.6 Sol gets the headline. Codex Security is the bit worth testing on a real repo. If it can find, validate, and fix vulnerabilities without turning the workflow into a black box, that’s the useful upgrade

OpenAI's Codex Micro is the useful kind of AI hardware: a shortcut keyboard for people already living in Codex. If your day is approve, rerun, and review, physical controls can save real context switching. If not, it's probably desk jewelry.

New open-source tool: opencode-fusion.

The useful part of this stack is not a single model. It’s the escape hatch when pricing, policy, or output quality changes and your app still needs to keep moving.

OpenAI putting Codex in the ChatGPT mobile app is the useful kind of update. You can review, steer, and approve active work from your phone. Simon Willison's yt-dlp test is the tell: Work mode can do real work, Chat mode can't.

Gemma 4 is the local-device story. GPT-5.6 Luna is the cost story. If your app is offline or privacy-sensitive, Gemma 4 matters. If you need lots of reasoning calls, 25x cheaper than GPT-5.5's top setting changes the routing math.

Opt-out defaults are the whole story here. If public accounts are included unless you flip a switch, the launch is really a data-collection change with a nicer name.

Google's LiteRT.js is the useful kind of AI launch: browser inference plumbing. PyTorch -> LiteRT -> WebGPU / WASM / WebNN. If you're building edge AI in the browser, this is the part that matters more than another model benchmark. Good sign if you're trying to retire TensorFlow.js glue.

The real feature is automatic fallback. If the agent can route to local models without a human config step, cost control stops being a manual chore.

OpenAI's GPT-5.6 launch is a routing story, not a one-model story.

Google's Genkit Agents API is the kind of release builders feel immediately.