The cost of AI just changed. 

Big time. 

OpenAI just handed its billion free users unlimited GPT-5.6 Luna, a model hanging right around Claude Sonnet 5, a tier that caps paid users elsewhere at about 20 prompts every five hours.

"Free plans are always a tier behind, right?"

Yeahhh, that rule just died this week.

And that was only the loudest drop in the wildest free and open AI week ever: Adobe's creative arsenal inside ChatGPT, frontier coding at 95% off, open models crashing the top five.

We broke em all down on today's episode of Everyday AI, no technical chops required, plus which ones y'all gotta get reps on immediately. 

Let's get into it.

Adobe just collapsed its entire creative empire into one single ChatGPT plugin that taps more than 70 tools across Photoshop, Firefly, Acrobat, Express, Premiere, Lightroom, Illustrator, InDesign, and Adobe Stock.

The unlock? You never have to guess which app does what again.

Describe the outcome, dump your files in, and it routes everything automatically.

Guest mode even works without an Adobe account, though signing in unlocks the heavier generative firepower plus Creative Cloud sync.

Try This

Dump a folder of product photos and a messy spreadsheet into ChatGPT, then ask for a polished photo set plus a one-page PDF.

If it nails the routing on the first pass, you just deleted three app tabs from your creative workflow for good.

Meta released Muse Code, a terminal coding agent powered by its brand new Muse Spark 1.2 model and a million-token context window, with standard pricing at $1.25 per million tokens in and $4.25 out.

The catch? A new contributor tier drops that to 10 cents in and 20 cents out, roughly 95% off, if you let Meta TRAIN its future models on your data.

Enterprises should sprint the other direction. But for hobbyists starting from scratch, honestly, not a terrible trade.

The bigger story: Anthropic reportedly makes about 80% of its revenue on coding, and Meta just aimed its cheapest weapon straight at it.

Try This

Got a personal project with zero proprietary code in it? Install Muse Code on the contributor tier and let a background agent grind through your weekend backlog.

Just keep anything client-facing or remotely proprietary far, far away from that tier.

Riding that same open wave, Alibaba dropped Qwen 3.8 Max, a 2.4 trillion parameter open weight model that reads text, images, and video.

Per Artificial Analysis, it sits in the same tier as GPT-5.6, Claude Fable 5, Claude Opus 5, and Kimi K3, meaning two of the world's top five models are now Chinese open weights.

The weights drop within days, though a fresh Reuters report says Alibaba plans to start charging major enterprise users a share of their revenue next week.

You won't be self-hosting this one without a spare data center in the basement. The real win is pricing pressure that keeps the big US labs honest.

Try This

There's nothing for most of us to install this week unless you have serious compute collecting dust.

Instead, forward this one to whoever negotiates your AI contracts, because frontier-level open models are your best leverage at renewal time.

The tool formerly known as NotebookLM got its glow up: it's now Gemini Notebook, and for paid users, it's agentic by default.

Every notebook gets a secure cloud computer that writes and runs real code against your sources, then hands back charts, spreadsheets, slide decks, and PDFs.

It runs on Gemini 3.5 with the Antigravity harness plus 100 built-in skills, and it can even query the web now, asking permission first so your grounding stays intact.

One catch: under the hood sits Gemini 3.5 Flash, scoring a 52 on Artificial Analysis, roughly what free ChatGPT users now get unlimited.

Access: Gemini Pro accounts, web only.

Try This

Load last quarter's reports into a notebook and ask for the charts plus a slide deck in one prompt.

If it saves your analyst a full afternoon, we did the complete walkthrough in episode 834.

There's a new open, vendor-neutral standard for packaging AI agent extensions: one folder with a plugin.json file bundling your agent skills and MCP servers together.

Build a plugin once and it runs across ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and more, with Vercel, Amazon, Anysphere, GitHub, and Microsoft all on board.

That matters because loading these into every tool separately burns hours and quietly locks your whole AI stack into one vendor.

Who's missing? Anthropic.

Almost every other big name jumped on within hours, which makes the lone holdout look like a lock-in play.

Try This

Point your coding agent at agent-plugins.org and repackage your team's existing skills and MCP servers into one portable plugin.

If part of your company ever needs a different AI stack, future you says thanks.

Speaking of open, Cloudflare open sourced Cloudflare OS, the internal AI workspace it built to actually run its own company, and thousands of its employees already use it daily, so this ain't demo software.

Every employee gets a browser-based agent workspace: governed access to internal systems, research tied to live company data, automated workflows, and shareable internal apps with zero developers needed.

It competes head-on with ChatGPT Work, Claude Cowork, and Gemini Enterprise, except it's free, open source, and runs in your own Cloudflare account with your own model and spend caps through Cloudflare AI Gateway.

Try This

Send the repo to your IT lead with one question: what would a two-week pilot actually cost us?

If the answer comes back as mostly staff time, you're looking at a small fraction of what per-seat enterprise AI pricing runs.

MiniMax pushed the open wave into video, releasing the weights for H3, the first open model to ever top a major AI video ranking: number one in video editing on Artificial Analysis, number two in text to video.

It reads text, images, video, and audio as one context, then spits out four to 15 second clips at up to 2K with native stereo audio.

Seen suspiciously good AI video flooding your feed this week? This was the engine.

The gnarly part? The Hugging Face weights exclude the US, EU, UK, and South Korea from local use, so those regions are API only.

And video this good before a US election should worry everyone.

Try This

Run one 15-second product or explainer clip through the hosted API before your next campaign brief.

If the quality holds up, your video production budget just changed for good.

OpenAI capped the week by giving its billion users unlimited GPT-5.6 Luna on the free plan, replacing the old GPT-5.5 and GPT-5.4 combo entirely.

Luna hangs right around Claude Sonnet 5 territory, and free users now get it with NO usage cap.

Paid users get a cleaner setup too: instant and thinking merged into one model, with a slider controlling how much reasoning goes into each answer.

OpenAI also claims 68% fewer factual errors on financial, medical, and legal prompts versus GPT-5.5 Instant.

The subtext, fam: between cheap open models and free frontier-adjacent chat, it has been a rough stretch for Anthropic ahead of a public listing expected in as soon as two months.

Try This

Paying for seats your team barely touches? Have your light users re-test free ChatGPT this week and count who actually needs the paid tier.

If half your seats can downgrade, that budget just moved to tools doing real work.