• Everyday AI
  • Posts
  • Claude 3.7 Sonnet: World’s first hybrid AI model. How it works and when to use it

Claude 3.7 Sonnet: World’s first hybrid AI model. How it works and when to use it

Claude 3.7 Sonnet explained, Google and Salesforce expand partnership, DeepSeek eyes R2 launch, Google unveils free coding assistant for developers and more!

Author

Everyday AI
February 25, 2025

Outsmart The Future

Sup y’all 👋

OpenAI JUST released a pretty big update — Deep Research is available for all paid users. And, they dropped a bunch of new-ish features. 

In our Deep Research throw down, we found OpenAI’s variation heads and shoulders above everyone else’s, except on price. 

It was previously only available on the Pro plan at $200/mo. 

Now that it’s available to all paid users, should we do a (shorter) version on the updated version of OpenAI’s Deep Research?

Should we do a shorter update on OpenAI Deep Research?

Vote to see live results

Login or Subscribe to participate in polls.

✌️
Jordan 

Today in Everyday AI
7 minute read

🎙 Daily Podcast Episode: Claude’s newest AI model is finally here and it’s the world’s first hybrid model. We break down what that means for the AI world. Give it a listen.

🕵️‍♂️ Fresh Finds: OpenAI makes Deep Research available for paid users, Microsoft rolls out free access to Voice and Think Deeper and Chegg sues Google. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: Google and Salesforce expand partnership, DeepSeek eyes R2 launch and Google unveils free coding assistant for developers. For that and more, read on for Byte Sized News.

🚀 AI In 5: Here’s a sleeper feature inside Perplexity that’ll save you an insane amount of time. See it here

🧠 Learn & Leveraging AI: So what’s a hybrid model and why does it make Claude 3.7 Sonnet so special? We explain what you need to know. Keep reading for that!

↩️ Don’t miss out: Did you miss our last newsletter? We talked about Anthropic unveiling Claude 3.7 Sonnet, Apple looking to add Gemini to Apple Intelligence and Perplexity announcing its Comet browser. Check it here!

 Claude 3.7 Sonnet: World’s first hybrid AI model. How it works and when to use it 🧠

The world's first hybrid LLM is here.

We've been waiting since June for Anthropic's next heavyweight model. With Claude Sonnet 3.7, not only is that wait over, but we also have the world's first hybrid model.

What's it mean?

And how should you use it?

Join the conversation and ask Jordan questions on Claude here.

Also on the pod today:

• Performance Benchmarks 📊
• Potential Use Cases 🤔
• Discussion on Hybrid AI Models 💭

It’ll be worth your 1 hour and 14 minutes:

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 – Polymet is an AI product designer, Magic Inspector is an AI web test automation platform and Nuvio provides AI-powered financial management.

OpenAI – OpenAI is rolling out Deep Research to ChatGPT Plus and a free version of its Advanced Voice mode

Microsoft – Microsoft is rolling out free unlimited access to Voice and Think Deeper modes.

Perplexity – Perplexity has launched a $50M seed and pre-seed VC fund.

Trending in AI – Dow Jones has expanded its AI marketplace to nearly 5,000 publishers.

AI Governance – UK newspapers are protesting the loss of AI protections on their front pages with the ‘Make It Right’ campaign.

Business of AI - Chegg is suing Google for AI’s impact on Chegg’s traffic and revenue and its causing the company’s stock to sink.

AI in Media – More than 1,000 musicians have released a silent album protesting the UK’s changes to copyright law.

AI Models - Covergence has released proxy-lite-3b, a small open weights model.

AI Research – A new study by the Pew Research Center shows that U.S. workers are more worried than hopeful about AI use in the workplace.

1. Salesforce and Google Cloud Forge Powerful AI Alliance 🤝

Salesforce and Google Cloud have announced a significant expansion of their partnership, enabling Salesforce customers to leverage Google’s cutting-edge Gemini AI models. This collaboration aims to enhance the deployment of AI agents and provide access to real-time data analytics, a move likely to reshape how businesses approach AI integration in their operations.

With a burgeoning market for agentic AI projected to hit $2 trillion, both companies are positioning themselves at the forefront of this transformative technology.

2. DeepSeek Gears Up for R2 Launch 👀

Chinese startup DeepSeek is reportedly accelerating the release of its next-gen R2 model, now expected even sooner than its initial May launch. Following the success of the budget-friendly R1, which outperformed several high-cost competitors, R2 aims to enhance coding capabilities and broaden reasoning skills beyond English, potentially disrupting U.S. tech dominance.

Industry experts are already buzzing about the implications, with some voicing concerns about the geopolitical stakes involved. This rapid advancement could force competitors to rethink their strategies and pricing, ultimately influencing how companies leverage AI in their operations.

3. Google Unveils Free Gemini Code Assist for Developers 🚀

Google has launched Gemini Code Assist for Individuals, a free AI coding assistant that offers a staggering 180,000 code completions per month—90 times more than GitHub's free Copilot plan. This tool allows developers to interact with an advanced AI model using natural language, making it easier to debug, complete, and understand code.

Additionally, Gemini Code Assist for GitHub automatically checks pull requests for bugs, enhancing productivity within coding teams. With these offerings, Google aims to attract early-career developers and potentially upsell them to enterprise plans in the future.

4. Cisco and NVIDIA Team Up for AI Networking 🌐

Cisco and NVIDIA revealed plans to develop a unified architecture aimed at simplifying AI-ready data center networks. This collaboration will integrate NVIDIA's Spectrum-X Ethernet platform with Cisco's Silicon One, making Cisco the exclusive partner silicon in this innovative approach.

As enterprises grapple with the complexities of AI integration, this partnership promises to streamline AI workload deployment, potentially transforming how companies manage their data centers. With updates expected mid-2025, professionals looking to leverage AI technology will find new pathways to optimize their infrastructure investments, signaling a significant shift in the tech landscape.

5. DeepSeek's API Access Restored ✅

After a three-week hiatus, Chinese AI startup DeepSeek has reopened access to its API, allowing developers to resume building applications on its cloud-hosted AI models. This reopening comes at a crucial time, as DeepSeek's R1 "reasoning" model has gained traction, challenging the likes of OpenAI and prompting competitive responses in the tech landscape.

However, users should note that server resources are still strained during peak hours, indicating ongoing demand for these advanced AI capabilities. With rivals like Alibaba also pushing new models, the AI sector is heating up, promising exciting developments for developers and businesses alike.

STOP Ignoring This Perplexity Feature! Save Hours Daily

So many people are sleeping on this ONE feature inside Perplexity that’ll save you an insane amount of time.

We’re showing you Perplexity Collections and how to use it.

🦾How You Can Leverage:

Claude 3.7 Sonnet is bravely going where no other publicly available AI has gone before. 

It's deciding—all by itself—when to kick its brain into overdrive.

Anthropic recently released Claude 3.7 Sonnet, the world’s first ‘hybrid’ model. In other words, Claude will choose when it uses the ‘old school’ transformer type of fact-spitting, and when it’ll use the reasoning model that uses more compute. 

After testing it extensively on math problems, podcast analytics, and brain teasers, we discovered something fascinating: this model literally toggles between instant responses (23 seconds) and deep thinking sessions (3+ minutes) on identical prompts. 

Sometimes a brilliant thinker, sometimes head-scratching.

(It’s early, after all, so maybe we’ll have to check back in a few months?) 

The API pricing? A jaw-dropping $15 per million output tokens—20-50ish TIMES more expensive than small yet mighty models like GPT-4o mini and Gemini 2.0 Flash-Lite. 

Meanwhile, Anthropic quietly released "Claude Code," a terminal-based tool that edits your entire codebase without copy-pasting. 

Forget building chatbots—Anthropic is coming for GitHub Copilot and Cursor. The battle for your terminal has begun.

A lot to cover today, so you might wanna watch/listen/read the whole dang thang.

If you just want the Cliff Notes, let’s goooooooo. 

1 – Hybrid = Higher Floor, Lower Ceiling 🤷

The hybrid approach means one model now does what used to require two completely different systems. 

As an example, OpenAI has its GPT-4o (transformer) and its o3-mini-high (reasoner)

Google has its Gemini 2.0 Pro (transformer) and its Gemini 2.0 Flash Thinking (reasoner) 

Now, Claude has both in one package with its hybrid Claude 3.7 Sonnet. 

No more switching between transformer models (like GPT-4o) and reasoning models (like o1 Pro). Claude 3.7 Sonnet does both—and decides which approach to use without asking.

Free users? Sorry, no advanced thinking for you. That toggle is premium-only.

The hybrid approach introduces a frustrating paradox. Average users get more reliable baseline performance (higher floor), but power users lose the specificity and control that makes these tools truly exceptional (lower ceiling).

The model now decides when to think deeply versus answer directly – not you.

Try This:

Create thinking instruction prefixes for your prompts. Start important questions with "Please use extended thinking on this problem" or "This requires step-by-step analysis."

For simple queries, try "Please answer directly without extended analysis." Test which phrases consistently trigger the right thinking mode. Build a prompt library of these prefixes tailored to your specific needs.

The model may decide on its own whether to think deeply, but your prompt engineering can heavily influence that decision 

2 – Anthropic Competing as an IDE? 🤔

Claude Code isn't just another feature—it's Anthropic declaring war on developer tools.

Unlike Claude's chat interface, this command-line tool works with your actual files and codebase, directly in your terminal. Give it a folder containing JavaScript, HTML, CSS, and other files, and it will modify the entire codebase at once. 

No more copy-pasting snippets between windows. No more "here's how you might approach this." It just does the work.

Talk about eating your customers. Hom Nom. 

Claude Code is even available to free users as a "research preview," signaling Anthropic's hunger for developer mindshare. 

And it makes perfect sense when you consider the 70.3% score on SWE bench verified coding tests—demolishing every competitor by 20+ percentage points.

Try This:

Want to try Claude Code?

You don't need coding experience. Install it through GitHub, point it at any folder on your computer, and start with basic requests like "Create an interactive website that displays a diet and exercise plan for a 40-year-old man” or "Build me a tool that organizes my photos by date." 

Claude Code handles the technical implementation while you focus on describing what you want built.

Just talk to it like a human.

3 – Coders Rejoice.Everyone Else? Watch Your API Bills 💰

Let's talk cash. 

$3 per million tokens input, $15 per million tokens output.

Double Yikes. 

That's not a typo. That's how much Claude 3.7 Sonnet costs on the API if you’re building in the backend. 

(Frontend users… you’re not paying anything extra.) 

Compare that to GPT-4o Mini at 15¢ input/60¢ output, making Claude approximately 25 TIMES more expensive for everyday business applications.

Anthropic didn't cut prices like everyone expected. They doubled down on premium positioning.

And that’s the downside of a hybrid model. Yeah, there’s a bit of tuning for how much ‘thinking’ you wanna use via the API, but you’re prolly gonna end up gobbling up extra tokens you really don’t want. 

(And REALLY paying for it.) 

The market segmentation is crystal clear: Anthropic wants the coding and engineering market and ain’t gonna play the API price war games. 

For software companies that need to parse complex code repositories, formulate algorithms, or build sophisticated applications, the price tag makes sense. 

The quality gap in technical tasks is legitimately massive. TAU bench for agentic tool use? Claude scored 81% versus OpenAI's 73%.

But for customer service chatbots, content generation, or basic business analytics? 

The math doesn’t math for businesses to build on Claude 3.7 Sonnet. 

Try This:

If building AI applications, implement a "model router" that sends technical coding or software dev tasks to Claude 3.7 and everything else to cheaper models. 

For example, code generation, technical documentation, and complex mathematical analysis could go to Claude, while customer inquiries, content generation, and simple requests go to GPT-4o Mini, Gemini 2.0 Flash Thinking lite or even Claude 3.5 Haiku if you’re really in deep with Anthropic. 

Morale of the story? 

A hybrid model is nice in theory, but backend users need to watch their API bills. 

⌚

Numbers to watch

$15 Million

App building AI platform Lovable has raised $15M.

Now This …

Let us know your thoughts!

Vote to see live results

Do you attend our livestreams?

Every weekday, we bring you fresh AI insights, exclusive interviews, and breaking news with our Everyday AI livestream.

Login or Subscribe to participate in polls.

Reply

or to participate.