Gemini 2.5 Flash ‘thinks’ on a budget

Good morning, AI enthusiasts. The AI reasoning revolution just got so much cheaper — with Google launching its recent Gemini 2.5 Flash in preview with performance rivaling top models at significantly lower costs.

With a toggle to regulate when pondering kicks in and a “budget” for balancing quality, speed, and price, could this be the model that finally scales reasoning to the masses?

P.S. — our next workshop is TODAY at 4 PM EST. Learn how you can unlock the complete potential of o3, o4-mini, and GPT-4.1 and construct with OpenAI’s newest models. RSVP here.

In today’s AI rundown:

• Google’s Gemini 2.5 Flash with ‘pondering budget’

• Profluent finds scaling laws for protein‑design AI

• Transform your spreadsheets with AI in Google Sheets

• Meta’s FAIR shares recent AI perception research

• 4 recent AI tools & 4 job opportunities

LATEST DEVELOPMENTS

GOOGLE

🤔 Google’s Gemini 2.5 Flash with ‘pondering budget’

The Rundown: Google just launched Gemini 2.5 Flash — a hybrid reasoning AI in preview that matches o4-mini, outperforms Claude 3.5 Sonnet on reasoning/STEM benchmarks, and introduces a brand new ‘pondering budget’ to optimize cost vs. quality.

The main points:

• 2.5 Flash shows significant reasoning boosts over its predecessor (2.0 Flash), with a controllable pondering process to toggle the feature on or off.

• The model shows strong performance across reasoning, STEM, and visual reasoning benchmarks, despite coming in at a fraction of the associated fee of rivals.

• Developers may also set a “pondering budget” (as much as 24k tokens), which fine-tunes the balance between response quality, cost, and speed.

• It is accessible via API through Google AI Studio and Vertex AI, and can also be appearing as an experimental option inside the Gemini app.

Why it matters: OpenAI could have dominated the conversation this week, but Google is shipping right alongside them. The controllable, budgeted reasoning is an interesting customization, with users in a position to hit the feature only when a task needs it — unlocking reasonably priced, high‑volume use cases and saving pondering for more complex jobs.

TOGETHER WITH INNOVATING WITH AI

🚀 Launch a six‑figure AI consultancy in six months

The Rundown: Innovating With AI’s “The AI Consultancy Project” gives you the frameworks, playbooks, and client‑ready templates needed to show “interesting AI ideas” right into a revenue‑generating business – helping you ride the wave of an AI consulting boom expected to grow by 8x this decade.

On this program, you’ll:

• Gain the tools and frameworks to seek out clients and deliver top-notch services

• Follow a 6-month plan to construct a 6-figure AI consulting business

• Join a 700‑strong cohort where some members landed their first AI client in 72 hours

Click here to request access to The AI Consultancy Project.

COMPANY

🧬 Profluent finds scaling laws for protein‑design AI

The Rundown: Profluent announced ProGen3, a brand new family of AI models that may design complex proteins from scratch — with the outcomes marking the primary evidence of AI scaling laws in biology, proving larger models and more data create stronger results.

The main points:

• The Biotech company’s 46B model was trained on 3.4B protein sequences, surpassing previous datasets and showing improved protein generation.

• It successfully designed recent antibodies matching approved therapeutics in performance, yet distinct enough to avoid patent conflicts.

• The platform also created gene editing proteins lower than half the dimensions of CRISPR-Cas9, potentially enabling recent delivery methods for gene therapy.

• Profluent is making 20 “OpenAntibodies” available through royalty-free or upfront licensing, targeting diseases that affect 7M patients.

Why it matters: If scaling trends hold, Profluent’s approach could turn drug and gene‑editor design from years‑long lab work right into a faster, more predictable engineering problem — rewriting how recent therapies are discovered. These trends also suggest we’re just firstly of AI’s impact on drug discovery and medicine.

AI TRAINING

🔢 Transform your spreadsheets with AI in Google Sheets

The Rundown: On this tutorial, you’ll learn how you can use Google Sheets’ recent AI formula to generate content, analyze data, and create custom outputs directly in your spreadsheet—all with a straightforward command.

Step-by-step:

1. Open Google Sheets through your Google Workspace account (it’s slowly being rolled out).

2. In any cell, type =AI(“your prompt”, [optional cell reference]) with specific prompts like “Summarize this customer feedback in three bullet points.”

3. Apply your formula to multiple cells by dragging the corner handle down a complete column for batch processing.

4. Mix with standard functions like IF() and CONCATENATE() to create powerful workflows, and use “Refresh and insert” anytime you would like updated content.

Pro tip: You can too include formatting instructions directly in your prompt, equivalent to “in table format” or “as a numbered list,” to regulate how your output appears within the cell.

PRESENTED BY IMAGINE AI LIVE

⏰ AI’s future won’t wait — will you retain up?

The Rundown: IMAGINE AI LIVE ’25 isn’t just one other event— it’s your ticket to leapfrog competitors who’re still scrambling to grasp AI’s potential. Join industry pioneers on the Fontainebleau Las Vegas and get practical strategies to implement AI that drives immediate results.

Attend on May 28-30 on the Fontainebleau Las Vegas and you’ll be able to:

• Speed up your AI journey by years, not months

• Connect with cross-industry innovators who’ve already solved your challenges

• Transform theoretical AI knowledge into actionable implementation roadmaps

Secure Your Competitive Edge today and save $150 when you enroll by April 18th, 2025!

META

🔬 Meta’s FAIR shares recent AI perception research

The Rundown: Meta’s FAIR research arm just published five recent open-source AI research projects focused on perception and reasoning, showcasing advances in computer vision, 3D understanding, and collaborative AI capabilities.

The main points:

• Perception Encoder shows SOTA performance in visual understanding, excelling at tasks like ID’ing camouflaged animals or tracking movements.

• Meta also introduced the open-source Meta Perception Language Model (PLM) and a PLM-VideoBench benchmark, specializing in video understanding.

• Locate 3D enables precise object understanding for AI, with Meta publishing a dataset of 130,000 spatial language annotations for training.

• Finally, a brand new Collaborative Reasoner framework tests how well AI systems work together, showing nearly 30% higher performance vs. working alone.

Why it matters: This research batch focuses on AI constructing blocks like perception, 3D understanding, and reasoning — key steps toward more capable embodied agents and machine intelligence. We’re officially crossing into recent territory, with systems that may finally understand and interact with the physical world in advanced ways.

QUICK HITS

🛠️ Trending AI Tools

• 🎯 Gamma 2.0 – Easily craft stunning AI presentations, interactive web sites, social carousels and more from easy text prompts*

• 🧠 o3 and o4-mini – OpenAI’s recent models with visual reasoning and power use

• ⚙️ Codex CLI – OpenAI’s open-source coding agent for users’ terminals

• 🖥️ Copilot Computer Use – Construct agents that may use and navigate GUIs

*Sponsored listing

💼 AI Job Opportunities

• 🧠 Deepmind – Strategy & Operations Principal

• 📝 Meta – Content Strategist

• 📱 OpenAI – iOS Engineer

• 💰 Harvey – Sr. Compensation Analyst

📰 All the pieces else in AI today

OpenAI’s recent o3 model scored a 136 (116 offline) on the Mensa Norway IQ test, surpassing Gemini 2.5 Pro for the best rating recorded.

UC Berkeley’s Chatbot Arena AI model testing platform is officially breaking out from its research project status into its own company called LMArena.

Perplexity reached a cope with Motorola and is reportedly in talks with Samsung to integrate its AI search platform into their phones because the default assistant or an app.

xAI’s Grok rolled out memory capabilities for remembering past conversations, also introducing a brand new Workspaces tab for organizing files and conversations.

Alibaba released Wan 2.1-FLF2V-14B, an open-source model that enables users to upload the primary and last frame image inputs for a coherent, high-quality output.

Music streaming service Deezer reported that over 20K AI-generated songs are being published every day, with the corporate using AI to filter out the content.

OpenAI reportedly explored acquiring Cursor creator Anysphere before entering the present $3B discussions with rival Windsurf for its agentic coding platform.

COMMUNITY

🎥 Join our next live workshop

Join our next workshop today at 4 PM EST with Dr. Alvaro Cintas, The Rundown’s AI professor. By the tip of the workshop, you’ll be ready to begin constructing with OpenAI’s latest models — o3, o4-mini, and GPT-4.1.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

🤝 Share The Rundown, get rewards

We’ll all the time keep this article 100% free. To support our work, consider sharing The Rundown with your pals, and we’ll send you more free goodies.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team