Two months ago, we called Anthropic the 2026 leader in AI.

Well, the race looks a bit different now.

Why?

Google just had its biggest product week in company history at I/O. New models, a video-to-anything engine, and a developer platform that tried to become five products at once.

More announcements than any single team could process in a week.

And somehow, Codex STILL won.

Not because OpenAI out-announced anyone. Because while the industry was parsing keynote slides, Codex shipped the ability to run autonomously on your Mac while it's closed, locked, and sitting in your bag.

If your AI strategy is still anchored to waiting for Google to get it right, this week should prolly make you very uncomfortable.

On today's Everyday AI, we broke down all seven, including the cost model that undercut Opus 4.7 by 90%, the PowerPoint upgrade most people missed, and why Codex winning quietly may be the most important story in enterprise AI right now.

Let's get after it.

Not Pro. Flash.

Google's fastest model tier just outperformed Gemini 3.1 Pro on coding and agentic benchmarks while running four times faster than comparable frontier models.

What does this mean for your team?

If your staff uses Google's AI Mode in Search, it silently upgraded this week with zero action required. For API and agentic workflow builders, this is the current model to build on.

One catch: the cheap Flash pricing from older generations is gone. Gemini 3.5 Pro arrives next month, which will give a clearer picture of where Google actually stands.

Try This

Rerun two or three of your team's most complex recurring research questions inside Google's AI Mode today.

If the output is sharper and more connected, your baseline just improved for free. Note where the biggest jump happened. That is your new default research surface.

This one slipped under the radar. Almost nobody covered it because Codex dominated the news cycle that day.

ChatGPT is now a native add-in inside PowerPoint, pulling directly from Gmail, Outlook, and SharePoint to build slides from your actual work. Outputs stay fully editable.

Here's the part worth caring about.

There's a reasoning layer that critiques your deck's narrative, flags weak storytelling, and anticipates executive pushback.

Available in beta for EVERY ChatGPT plan. Install via the Add-ins menu in PowerPoint.

Try This

Install the add-in and run a recurring deck through it alongside the emails and documents that normally inform it.

Ask it what a senior executive would push back on. If it surfaces a real gap, you just found a prep tool that pays for itself in the first meeting.

Everyone's comparing this to Seedance 2.0 on output quality. Wrong frame.

Gemini Omni Flash accepts text, images, audio, and existing footage as input. The unlock is describing changes to footage you already have conversationally.

Swap a background, change a scene, alter what's happening. Omni executes without touching the rest of the clip.

On raw quality, Seedance 2.0 still leads. Omni plays a genuinely different game.

Available for Google AI Plus, Pro, and Ultra subscribers via the Gemini app and Google Flow.

Try This

Find a piece of recorded content your team has wanted to update but can't justify reshooting.

Upload it to Gemini Omni Flash and describe one specific change conversationally. If it executes cleanly without disrupting the rest, you just found a post-production workflow that costs almost nothing.

This is the pricing story Anthropic should be worried about.

Composer 2.5 is built on Moonshot AI's Kimi K2.5 base with 25 times more synthetic training data. Pricing: $0.50 per million input tokens versus $5 for Opus 4.7. Output: $2.50 versus $25.

What do you get for one-tenth the cost?

On SWE-Bench multilingual, it runs neck and neck with Opus 4.7. Roughly 80 to 90% of the capability. For teams hitting Anthropic token budget ceilings this quarter, that's a Biiiig lever.

Live now for all paid Cursor plans.

Try This

Pull your three most token-intensive agentic coding tasks and run them through Composer 2.5 this week alongside your current setup.

If output quality holds at 80 to 90%, you have a real cost lever that doesn't require any replatforming conversation with leadership.

Amazon launched Alexa Podcasts inside Alexa+: two AI-generated hosts, any topic, on demand, no document uploads required.

That "no documents needed" framing is Amazon calling out NotebookLM directly. The problem is it's also removing the feature that makes the category work.

NotebookLM Audio Overviews are great because they're grounded in your documents. Alexa Podcasts pulls from the open web, and Alexa+ has well-documented accuracy problems.

Available with Amazon Prime at no extra cost. Hard to see it replacing NotebookLM for anything requiring accuracy.

Try This

Ask Alexa Podcasts to cover a topic your team is actively researching, then run the same topic through a NotebookLM Audio Overview using your own source documents.

The accuracy difference makes the right tool for your team immediately obvious.

Google's Antigravity 2.0 launched at I/O as a desktop app, CLI, SDK, and managed agents tier. Five surfaces. One week.

The actual launch video had a folder named "Codex" visible on screen. The app looks virtually identical to Codex. Multiple developers opened it and thought their Codex chats had migrated over.

Google also deprecated the Gemini CLI for consumer users in June, replacing it with Antigravity. The product got more keynote spotlight than it was ready for.

Available on macOS, Windows, and Linux. Watch and wait.

Try This

Install Antigravity 2.0 and run one non-critical workflow through it this week.

Don't benchmark it against Codex today. Evaluate against where Google's resources suggest it could be in six months, then decide if being early is worth the rough edges.

Three features. One Thursday. ZERO fanfare.

App shots: press both Command keys and Codex captures the frontmost app window, screenshot and full text context included. The copy-paste friction that slows down every agentic workflow just disappeared.

Goal mode is out of experimental and now live in the desktop app, IDE extension, and CLI. Set an objective and Codex works toward it for hours or even days.

Remote computer use: Codex operates desktop apps after your Mac locks, laptop closed, controllable from your phone. Plugin sharing across teams also shipped.

Try This

Enable goal mode tonight, write one tightly scoped task with specific success criteria, and let it run while you sleep.

If usable work is waiting for you by morning, y'all need a very different conversation about how your AI workday is actually structured.