• Everyday AI
  • Posts
  • Ep 775: Open Source AI 101: Why Local Models, Cheap APIs, and AI Agents Change Everything (Start Here Series Vol 24)

Ep 775: Open Source AI 101: Why Local Models, Cheap APIs, and AI Agents Change Everything (Start Here Series Vol 24)

Google is launching its Gemini Omni video agent, Google releases Gemini Intelligence and AI hardware, and OpenAI unveiled a new AI cyber initiative called Daybreak and more.

Author

Everyday AI
May 12, 2026

 In Partnership With Adobe  

Adobe just introduced an entirely new way to create with Firefly AI Assistant. 

Adobe just introduced an entirely new way to create — bringing the power and precision of its creative suite into one conversational experience. With Firefly AI Assistant — now live in the Adobe Firefly app, the all-in-one creative AI studio - just describe what you want to create, and the assistant orchestrates the workflows, drawing on 60+ pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premiere, Lightroom and more — to help bring your ideas to life.

Every step the assistant takes is visible, so you can refine, redirect, or take over at any time. You stay in control of the outcome as the creative director. Check it out today at firefly.adobe.com.

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: Open source AI models are now close enough to frontier systems that, in Start Here Series Vol. 24, we break down why companies are debating whether the cost savings outweigh the legal and security risks. Give today’s show a watch/read/listen.

🕵️‍♂️ Fresh Finds: Thinking Machines is previewing a real-time multimodal AI model, Gemini can now build travel itineraries from your personal Google data, and Isomorphic Labs just raised $2.1 billion for AI drug discovery, and more. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: Google is launching its Gemini Omni video agent, Google releases Gemini Intelligence and AI hardware, and OpenAI unveiled a new AI cyber initiative called Daybreak, and more. Read on for Byte Sized News.

💪 Leverage AI: Most companies built their AI strategy around always using the best model. But local models, cheap APIs, and always-on agents are changing the math faster than most leaders realize. Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: OpenAI just launched a $14 billion AI consulting division, Google confirmed hackers used AI to find a major software vulnerability, and the Musk vs. OpenAI trial is heating up again, and more.Check it here!

Ep 775: Open Source AI 101: Why Local Models, Cheap APIs, and AI Agents Change Everything (Start Here Series Vol 24)


Until a few months ago, open source AI was kinda a hobby project.

Now, it's tearing corporate boardrooms apart.

Why?

Over the past 6ish months, the gap between frontier closed AI and open sourced AI has shrunk to pretty much nothing. And with the surge of always on agents driving open models, their development and release schedule is on pace with the frontier labs.

So if your team isn't paying attention to -- and running test cases through -- open AI models, there's a good chance you'll either be overpaying or playing catch up soon.

We walk you through the 101 and what you need to know when it comes to open source AI in this Start Here Series special.

Also on the pod today:

• Open vs. closed AI showdown 🤖 
• Chinese model distillation exposed 🇨🇳 
• API prices crash to pennies 💸 

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight –  Display.dev lets you Publish, gate, comment, iterate — at one URL, Free AI SEO Auditor lets you See your page through the eyes of ChatGPT, Claude, and Perplexity, Knooth records your screen, highlights key moments, and exports a clean video quickly.

Thinking Machines Interaction Model — Thinking Machines is previewing a new AI model that responds and interacts in real time across audio, video, and text.

Google Personal Intelligence — Google Gemini can now build custom travel itineraries by pulling info from your Gmail, Google Photos, and more.

Ismorphic Lab Drug AI — Isomorphic Labs just raised $2.1 billion to boost its AI-powered drug discovery.

Claudes Consitution AudioBook — Anthropic released an audiobook of Claude's Constitution, read by its authors, with a Q&A on AI ethics and design.

Heremes in Sigma Browser — Sigma Browser is pushing private, local AI agents right inside your browser.

Google Finance in Europe — Google Finance just rolled out its new AI-powered features across Europe, offering smarter research, real-time market updates, and advanced charting.

Agent View — Claude Code just added an agent view that lets power users run and manage multiple AI coding sessions from one terminal.

Softbank Chips — SoftBank just pumped $457 million into Graphcore to boost its AI chip ambitions.

Hosting Qwen on Blackwell — NVIDIA’s new GB200 NVL72 racks let you run massive MoE models like Qwen3 235B faster and cheaper, thanks to smarter parallelism and high-speed GPU links.

Google Biology — Google and top universities are teaming up to use quantum tech and AI for breakthroughs in biology.

Claude and AWS — Claude is now fully available on AWS, letting developers use the latest models with native AWS tools, security, and billing.

1. Malicious Code Hits Popular AI Packages 🐛

Microsoft confirmed Monday it is investigating a new supply‑chain attack after hackers slipped malware into the mistralai PyPI package, a widely used AI developer tool, causing malicious code to run automatically when imported on Linux systems.

The attack downloaded a hidden second-stage payload and launched malware in the background, echoing a broader wave of package compromises hitting both Python and JavaScript ecosystems, including popular TanStack and Mistral npm libraries. According to Microsoft and security researchers, the goal appears to be stealing developer credentials, which can open the door to much larger breaches across cloud services, CI pipelines, and software distribution channels.

2. Gemini Omni Set for Official Debut as Google’s Video Agent 📸

Just hours ahead of the Android Show kickoff, Google preps to release Gemini Omni, a new video agent that lets users craft and refine videos by blending images, text, and clips—plus starring their own avatar.

Unlike its rivals, Omni isn’t trying to outshine the best generators; it’s designed to orchestrate creative flows across multiple media types. The upcoming Avatars feature, tied to a quick selfie scan, is flagged as “new” and appears ready for public launch.

3. OpenAI Unveils “Daybreak” to Reinvent Cyber Defense ⚒️

OpenAI has just announced “Daybreak,” a major push to embed AI-powered cyber defense directly into the software development process.

This initiative marks a shift from simply patching problems to making software resilient to attacks from the start, using the same advanced models that power Codex and GPT-5.5. The rollout comes as OpenAI partners with leading tech and security firms, aiming to accelerate how quickly defenders can spot, prioritize, and fix vulnerabilities.

4. Google Unveils Android 17 and Gemini Automation Ahead of I/O ✨

Days before Google I/O on May 19, Google used its Android Show livestream to reveal Android 17 and a major expansion of Gemini, signaling a push toward deeper automation across phones, cars, browsers, and wearables.

The updates focus on Gemini handling real-world tasks like bookings, form filling, and voice dictation, while Android 17 adds security, digital well-being, and sharing improvements that affect daily use. Google also previewed a new AI-driven widget system and a broader device strategy that stretches from smartphones to laptops and vehicles.

5. Palantir Deepens AI Ties with Ukraine in War Tech Push 🪖

Ukraine's President Zelenskyy has just met with Palantir CEO Alex Karp in Kyiv, ramping up the country's use of artificial intelligence to gain a battlefield advantage against Russia.

The partnership includes the BRAVE1 Dataroom, a platform giving Ukrainian and allied developers access to real combat data to train AI for detecting and intercepting enemy drones. Ukraine's Defense Minister says the tech has already delivered detailed air attack analysis, smart intelligence processing, and better planning for deep strikes.

6. Google Replaces Chromebooks With Gemini-Powered Googlebooks 💻

Google on Tuesday unveiled Googlebooks, a new line of AI-first laptops built around its Gemini models, marking the company’s biggest shift in personal computing since the launch of Chromebooks 15 years ago, according to TechCrunch.

Launching this fall with hardware from Acer, Dell, HP, Lenovo, and others, the devices introduce features like an AI-powered cursor, deep Android phone integration, and Gemini-built widgets baked into the system. Googlebooks also signal a gradual move away from ChromeOS toward an Android-based operating system designed around AI from the start.

The real AI edge in 2026 is knowing when NOT to use the best model.

Yeah, read that one again. 

That’s the uncomfortable shift. A lot of companies built their 2025 AI strategy around one frontier vendor, one premium default model, and one very expensive assumption: best model equals best business decision.

Yes. But also, not always. 

That’s because early 2026 has exploded with local models, cheap APIs, and always-on agents that have changed the math. 

The downside? The legal tradeoff of open source models doesn’t politely vanish because the invoice got smaller.

On today’s episode of Everyday AI, we break down open source AI 101: why the open-versus-closed default flipped, why Gemma 4 puts serious capability on consumer hardware, why agent swarms are suddenly less ridiculous, and why your AI routing strategy now needs finance, IT, legal, and operations at the same table.

Because yeah, your competitors may not win by having better AI.

They may win by sending the right model to the right job while you’re still torching premium budget on PDF parsing.

1. Stop defaulting to premium models ⚡

Should the best model be used for everything? Not always. 

The ceiling capabilities gap between frontier closed models and open models has collapsed hard over the past year. We’re talking about a gap that moved from about 250 ELO points on Arena to about 30, and sometimes less. 

If your company burns through API tokens like a kid burns through cereal on a Saturday morning, it might be worth looking at triaging some of the lower-hanging fruit to an open model. 

Summarization, extraction, classification, PDF parsing, basic research, first-draft content, and other high-volume low-risk work don’t automatically deserve the AI model with the high price tag. 

Try This

Monday morning, pull your last 30 days of API usage and tag every workflow by task type, volume, risk, privacy sensitivity, and customer visibility.

Anything high-volume, repetitive, internal, and low-stakes should be tested on an open model. Not production. Test lane.

That one audit can show where you’re paying frontier prices for work that no longer needs 2026 frontier intelligence.

2. Put agents on cheaper rails 🚀

Local models change the agent math, whether you’re an OpenClaw aficionado using something like Ollama. 

Google’s open source Gemma 4, at its release, was about 20X more efficient than its closest open source competitors. And throw in the fact it can run 24/7 for free on a consumer laptop with 2025-level frontier intelligence, and it changes the agent equation. This changes what employees can do locally from occasional usage to around-the-clock agentic operations.

That matters because agents used to feel like special projects with special budgets and special permission slips.

Now some of that work can run locally, self-hosted, or through cheaper open APIs.

A 100-agent swarm moving from 1,200-plus dollars on Opus to 60-some dollars on DeepSeek is a sizable jump. 

Try This

Pick one agent workflow that runs every day: sales-call summaries, support-ticket classification, research briefs, document parsing, or internal content drafts.

Run it three ways: cheap open API, local or self-hosted open model, and premium closed model as reviewer. Compare output quality, cost, latency, privacy posture, and failure risk.

The winner probably won’t be one model.

It’ll be a routing rule.

3. Price the legal tradeoff first 🔥

This is where cheap AI can get weird fast.

DeepSeek v4 Pro pricing at $0.43 per million input tokens and $0.87 per million output tokens looks extremely tempting when your API bill is acting rude. Or, if you have beefy hardware, running a model like GLM-5.1 locally is even a possibility. 

That tradeoff matters most when the work is regulated, customer-facing, legally sensitive, or used for final review.

For those workflows, the premium model may be the cheaper decision.

Annoying? Yes.

Also probably true.

Try This

Build an AI triage map before shifting workloads.

Cheap open APIs get bulk internal tasks. Local open models get private workflows and local agents. Premium closed models get final review, regulated work, high-value reasoning, and customer-facing output.

Then make legal, finance, IT, and business owners approve the map together.

That’s how you stop treating AI like a tool preference and start treating it like operating infrastructure.

Reply

or to participate.