Explore what's new in AI

OpenAI adds ChatGPT add-on to Excel and Google Sheets

ChatGPT now lives inside spreadsheets, so analysts can clean data, write formulas, and explain changes without leaving the workbook.

Gemini Agent leak shows schedules, tasks, skills, and browser actions

The leak suggests Google is moving Gemini toward a full agent runtime, with connected apps and web interaction built in from the start.

Anthropic Orbit leak points to a proactive assistant across Gmail, Slack, Drive, and Figma

The leak suggests Anthropic wants a background assistant that briefs you before you ask, which is the same shift ChatGPT Pulse is chasing.

Perplexity Computer adds premium health sources like NEJM and BMJ

Perplexity is widening Computer beyond finance, letting users run health research on the same cited-source stack used for markets.

Gemini 3.2 Flash appears in the model selector before Google I/O

Select users briefly saw Gemini 3.2 Flash and 3.1 Lite in the app, which suggests a Flash refresh is already close to shipping.

GPT-5.5 xHigh tops the Artificial Analysis intelligence index

The benchmark chart puts GPT-5.5 xHigh ahead of Claude Opus 4.7 and Gemini 3.1 Pro, a useful snapshot of the frontier race.

Codex shows task progress as a parseable checklist inside chat

The new UI makes it easier to see what the agent has done, what is blocked, and where the next human action sits.

Codex Auto Review keeps tasks moving by auto-approving routine changes

OpenAI is trying to reduce approval bottlenecks, which matters if you want long-running tasks to keep moving without constant human taps.

Cursor debug mode reproduces a Swift bug, reads logs, and fixes it on the first pass

The demo is a good workflow pattern: instrument, reproduce, inspect logs, patch, and verify, all inside the agent loop.

Microsoft Research says long-horizon agent failures are a horizon problem

The fix is not just more compute, the paper argues, but compressing action space so agents can train stably and still generalize later.

Warp open-sources oz-skills for SEO, accessibility, docs, and DevOps workflows

Warp is turning repeatable team practices into shareable skills, which is a clean example of how agent workflows are getting productized.

Meta's Agentic Code Reasoning cuts patch errors by forcing explicit proof

The paper shows that a structured reasoning template can make code review much more reliable, without running expensive test loops.

Pookie is an open-source Slack helper that can search, code, and connect tools

Pookie collapses a lot of glue work into one Slack-native bot, which is useful if your team lives in chat and wants fast self-hosting.

Codex++ patches the Codex desktop app with a startup tweak system

The project turns Codex into something users can extend locally, with a tweak tab, runtime loader, and update-safe install flow.

ComfyUI adds Luma Uni-1.1 as a partner node for prompt to reasoning to pixels

The integration exposes a different image pipeline, where the model reasons about constraints before it draws, instead of going prompt to pixels.

A trace and span guide shows how to debug LLM systems component by component

The explainer makes a strong case that end-to-end logs are not enough, because cost, latency, and failures live inside each span.

Anthropic says it uses public models internally plus a Mythos descendant

Boris Cherny says Anthropic mostly runs the same models everyone else can access, which undercuts some mystery around internal advantage.

Pi.dev pitches a minimal terminal harness for agents and humans

Pi emphasizes sessions, skills, and extensibility over a heavy bundled workflow, which is a direct reaction to tool sprawl in agent IDEs.

HeavySkill trains parallel reasoning into the model instead of the harness

The paper argues the real skill is internal parallel thinking plus deliberation, and reports strong gains on code and instruction benchmarks.

Matt Pocock says AGENTS.md should use context pointers, not giant docs

The practical tip is to keep agent instructions small and use links as navigation, so models can find the right details without overload.

Hermes Agent adds browser-harness skill with parallel cloud browsers

The new skill gives Hermes fuller browser control, including self-improving tools and multiple cloud browsers for harder web tasks.

Manus recommends connectors in context and asks for approval before enabling them

Manus is moving connector setup into the task loop, which should cut friction while keeping user approval in the path.

gog 0.16 turns Google Workspace into a safer terminal API

The new release adds sanitized Gmail reads, Drive inventories, Sheets tables, and safer profiles, making Workspace more scriptable for agents.

xAI rolls out Grok connectors for GitHub, Notion, Linear, Google, and Microsoft

Grok can now pull context from the tools teams already use, which makes it more useful inside real workflows instead of a standalone chat box.