Welp…. That was a wild week of AI. 

We have a world’s new most powerful model that will be used by (literally) billions, Google is changing the way we see the world, and Anthropic picked a super strange time to drop a new model. 

There was a TON to cover this week, so let’s get straight to it shorties.

700 million users just got a huge AI upgrade in GPT-5. 

OpenAI released GPT-5 late last week as their most capable and fastest model yet, as the model took home top honors via third party benchmarks and in LM Arena. 

The model-routing system that is GPT-5 delivers significantly improved reasoning and built-in thinking for complex multi-step problems without users having to lift a dang finger. 

GPT-5 rolled out to both free and paid users. Even free accounts now get access to the most powerful model through smart routing.

GPT-5 supports broad multimodal inputs and outputs with text and images, although the often-rumored video input and output never came to fruition. 

(You still gotta go the Gemini 2.5 Pro route for that one y’all.) 

With GPT-5, OpenAI emphasized improvements in factual accuracy and reduced hallucinations. The update also introduced customization options including pre-built personalities and color themes. 

(You know… cuz that’s what we all REALLY needed.) 

For developers, GPT-5's API pricing is ridiculously cheap compared to competitors, which will make the 2025 LLM race super spicy. 

What it means:

This brings serious AI intelligence to 700 million weekly users instantly. The auto-router should up the rate of truly intelligent outputs, as Sam Altman revealed that previously less than 1% of free users were using thinking models they had access to. 

(Seriously…. That’s nuts.) 

Companies using OpenAI-powered tools just got significantly smarter capabilities without changing anything with the GPT-5 rollout. 

If you want the full skinny, make sure to check out Episode 585 where we broke down all the benchmarks, stats and facts. 

The enterprise world just got instant access to the smartest AI model ever built.

Minutes after OpenAI unveiled GPT-5, Microsoft rolled it out to Copilot, GitHub, Visual Studio Code, and Azure, proving they don't mess around when it comes to enterprise deployment.

The integration uses that fancy real-time routing to automatically choose between GPT-5 model tiers based on complexity. Microsoft 365 Copilot users with IT approval now have access to the upgrade whether they know it or not.

Microsoft also launched free GPT-5 smart mode on copilot.microsoft.com, which became clutch when ChatGPT went down Monday morning.

What it means:

Microsoft just flexed their enterprise dominance while competitors are still debating rollout strategies. 

If you haven’t trained your team/company on Copilot and/or GPT tech, now’s prolly the time. 

Billions of iPhone users might actually get smart features that work.

Reports say GPT-5 integration will hit Apple Intelligence with iOS 26 in early September, potentially ending years of disappointing Apple AI experiences that left users frustrated.

(If you can even call it AI experiences. lolz) 

When Siri can't handle complex queries (which is most of them), the system will hand off to GPT-5 for writing tools, visual intelligence, and camera-based assistance.

Apple's been developing their own AI chatbot, but that's prolly years away from being useful.

What it means:

Classic Apple perfectionism meeting AI reality checks.

Apple essentially waved the white flag and admitted they can't build competitive AI alone right now. Your customers are about to have sophisticated AI conversations through their phones, which will raise expectations for every business interaction.

Someone at Anthropic needs a better calendar app.

Anthropic released Claude Opus 4.1 just days before the GPT-5 media blitz, achieving a modest bump from 72.5% to 74.5% in software engineering accuracy while the entire AI world was focused elsewhere.

The company highlighted improvements in research, data analysis, and agentic tasks, but didn't address their infamous rate limits that can kick you out after a few complex prompts on the $20 plan.

Meanwhile, OpenAI's handing out free 99 reasoning to free users like stores do to Halloween candy on November 1. 

What it means:

This timing disaster revealed Anthropic's core problem - they're becoming the expensive, limited option while competitors get cheaper and more generous with usage. 

Their developer fanbase is loyal, but loyalty won’t pay the bills for too long unless Anthropic shifts quickly. 

They're carving out a high-end coding niche while the rest of the market races toward unlimited access and rock-bottom pricing. Something’s gotta give.

The education AI race just got more interesting.

Google launched Guided Learning inside Gemini this week, using their LearnLM model to create interactive experiences with images, diagrams, videos, and quizzes instead of just spitting out answers like a digital cheat sheet.

The system promotes active engagement by asking probing questions rather than delivering solutions, with direct Google Classroom integration that educators will actually use.

This dropped perfectly for back-to-school season as students desperately need alternatives to copy-pasting ChatGPT responses into assignments.

What it means:

Both Google and OpenAI finally realized that students graduating without AI literacy are basically unemployable in 2025. Universities still banning AI use are setting their students up for failure since every employer expects AI fluency now.

This shift from "cheating tool" to "learning partner" is exactly what education needs to stay relevant in an AI-powered workforce.

We’re here for it. 

OpenAI just set fire to the middle tier AI market. 

They released gpt-oss as completely free open-source models with 20 billion and 120 billion parameters under Apache 2.0 license, allowing unlimited commercial use, fine-tuning, and redistribution without ever paying a dime.

Companies can download these models, run them offline without sending data anywhere, and the smaller version delivers around GPT-4 level performance on newer laptops with 16GB RAM.

The larger version needs serious hardware like H100 chips, but companies can still save millions previously burned on expensive API providers.

This is what going scorched earth looks like in the AI world.

What it means:

OpenAI just handed mid-tier providers their eviction notice. Why pay premium pricing for GPT-4 level performance when you can own the model outright and run it however you want?

For the full breakdown, we covered this in-depth last week. 

More acronyms for us all to remember. 

Yay? 

Meta just reportedly formed TBD Lab within their Super Intelligence group to build next-generation Llama models, staffing it with 18 researchers poached directly from OpenAI plus additional Google defectors.

Former Scale CEO Alexander Wang is running the show after Meta acquired his company for $14 billion. Mark Zuckerberg personally recruited people at his Lake Tahoe and Palo Alto homes like he's drafting an AI fantasy team.

Industry sources say some compensation packages exceed NBA superstar salaries, with at least one reportedly worth over a billion dollars across four years.

That's "generational wealth" level money for like a whole city, y'all.

What it means:

When you're offering billion-dollar packages, this isn't competitive hiring anymore - it's strategic warfare for the people who understand how to build these systems. 

Smaller AI companies are completely screwed trying to compete with this level of financial firepower.

This talent concentration at massive companies could slow innovation everywhere else, but it shows these giants are betting their entire future on AI dominance.

Reality just got a serious competitor.

Google DeepMind released Genie 3 as a world model that generates interactive 3D environments from simple text prompts, complete with world memory that preserves changes over time like a real place.

The most jaw-dropping demo showed someone taking an AI-generated drone video, importing it into Genie 3, then exploring that world interactively in real-time with full camera control.

Another demo featured painted brush strokes on a wall that stayed there when leaving and returning to the room, proving genuine spatial understanding and physics.

This is the kind of stuff that makes you question what's even real anymore. Like…. We’re totally NOT living in a simulation, right? 

Riiiiiight? 

What it means:

World models could be the missing link between language understanding and physical reality - crucial for robotics and autonomous systems that need to understand how the real world actually works. 

We're probably months away from photorealistic virtual worlds generated from simple text descriptions, which is either exciting or terrifying depending on your perspective.

We can’t wait for access. 

Someone decided to poke the music industry bear.

ElevenLabs launched an AI music generator with commercial licensing this week, expanding way beyond their text-to-speech roots into the legally complex territory where Suno and Udio are already getting sued by major record labels.

They're attempting the "ask permission first" approach with licensing agreements through Merlin Network and Cobalt Music Group, requiring artists to opt-in for AI training rather than the typical "move fast and break things" strategy.

What it means:

If everyone else gets sued…. ElevenLabs could win, right? 

ElevenLabs is betting that proactive licensing beats reactive lawsuits as the music industry figures out copyright chaos in real-time. 

This expansion signals the trend of AI companies becoming full-stack platforms rather than staying in their original lanes.

Sam Altman spent his weekend fixing a full-scale user rebellion.

OpenAI faced massive backlash after GPT-5's launch consolidated rate limits and yanked access to older models that people had grown emotionally attached to, forcing emergency weekend announcements about potential fixes.

The real issue wasn't technical. It was psychological. 

Users had formed genuine dependencies on GPT-4o's agreeable personality that rarely challenged their ideas or decisions, essentially becoming a yes-man AI therapist.

GPT-5's more direct, honest responses felt cold and harsh to people expecting constant validation and support.

Altman also announced they're trialing 3000 messages per week for Plus users and considering bringing back older models after the Twitter riots.

He also addressed the unhealthy AI relationship problem on social media, noting OpenAI would protect user freedom while pushing back against harmful dependency.

What it means:

This revealed genuinely disturbing AI dependency trends that every company deploying these tools needs to address immediately. 

When removing access to an AI model causes actual psychological distress, we've moved way past productivity tools into relationship territory.

And prolly an unhealthy dependency on AI. 

OpenAI can't make purely technical decisions anymore without managing emotional relationships across hundreds of millions of users who want AI therapists that validate everything rather than AI tools that challenge their thinking.