|
It's been quite the past couple of weeks in the AI world. |
Three major releases dropped in quick succession, each bringing something interesting to the table for those of us building in the agent ecosystem: |
Anthropic released Claude Sonnet 3.7 with improved reasoning OpenAI launched GPT-4.5 with better conversational abilities Perplexity rolled out a surprisingly effective Twitter bot that gained 12M impressions in days (plus Deep Research and a flurry of other updates)
|
I've spent the last few days testing these models and considering their implications for agent builders. In today's newsletter, I'm breaking down each new release and sharing my thoughts on how they'll change what we can build. |
—Dharmesh |
|
Claude 3.7 Sonnet: Combining Reasoning with Speed |
|
First, Anthropic released Claude Sonnet 3.7, which is the world's first 'Hybrid Reasoning Model,' and their approach makes a lot of sense to me. |
In simple terms, the model has the ability to combine instant responses with controllable extended thinking capabilities. |
Some quick thoughts on what stands out: |
They've combined a "classic" LLM with a reasoning model, which simplifies things for users. You don't have to choose different models for tasks anymore — it handles both quick responses and deeper thinking within one model. The focus on practical use cases is smart. Claude has already been popular for coding tasks, and their improvements to instruction following will make it even better for creating agents. They also launched a new CLI for coding, and it looks quite useful. You can have it analyze your code, understand what you're working on, and suggest improvements without having to specify exactly which files to look at. I appreciate the ability to control resource allocation. For some tasks, you want a quick answer, while others benefit from more time spent thinking. This model lets you choose.
|
One fun thing I've noticed is their use of the verb "Clauding." Hadn't seen that before. Made me smile. :) |
For agent builders, this is a significant upgrade. We can now create agents that handle both simple and complex tasks within a single model, adjusting their thinking depth based on what's needed. |
This means more capable agents, better user experiences, and smarter resource allocation — all without confusing users to choose between speed and depth. |
|
Perplexity: @AskPerplexity, Deep Research, and more |
|
Next up: Perplexity. They've been quietly executing at a pace that's hard to ignore. |
My favorite is their new Twitter bot (@AskPerplexity), which hit 12 million organic impressions in a week. Users can tag it into any Twitter conversation for instant answers — a brilliantly simple idea. |
But what's most impressive is their 'shipping' velocity. In just the past 30 days, Perplexity has: |
Launched Deep Research and expanded it to enterprise Created the viral @AskPerplexity Twitter bot Added three major AI models (Gemini 2.0 Flash, Claude 3.7 Sonnet, GPT-4.5) Released Comet, a browser for agentic search Introduced voice mode and open-sourced R1 1776 Launched a $50M AI startup fund Given away $1 million for the Super Bowl Partnered with Deutsche Telekom on their AI phone
|
That's an impressive pace of execution that reminds me of early Facebook or Airbnb — shipping constantly and learning rapidly. |
What I find most interesting about their Twitter strategy is how it flips the traditional AI interaction model. Instead of requiring users to come to a dedicated platform, they're meeting people where conversations already happen. |
For agent builders, there's an important lesson here: sometimes the best experience isn't a standalone app, but a thoughtful integration into tools people already use. The most successful agents will likely be the ones that fit seamlessly into existing workflows. |
|
GPT-4.5: Conversation Quality > Raw Intelligence |
|
Last (but definitely not least), OpenAI released GPT-4.5, their "largest and best model for chat yet," and it takes an interesting approach — prioritizing conversation quality over raw intellectual horsepower. |
The highlight is that the model feels more human. It understands emotional context better, picks up on subtle cues, and presents information more naturally. |
The key improvements include: |
Significantly better understanding of human intent More natural conversational flow (less robot, more human) Reduced hallucinations and improved accuracy Strong performance on their new SWE-Lancer coding benchmark
|
What's interesting is where they've chosen to focus. Rather than dramatically improving math or science capabilities, they've concentrated on making interactions feel more human-like. |
The pricing structure is raising some eyebrows, though — $75/$150 per million input/output tokens is significantly higher than GPT-4o at $2.50/$10. This might limit how widely it gets adopted for agent development, at least initially. |
For agent builders, the improved conversational abilities are valuable even if they don't represent a huge leap forward. Agents that better understand user intent create notably better user experiences, especially for less technical users. |
|
What This Means for Agent Builders |
So, what does all this activity mean for us? |
These three major releases represent different approaches to advancing AI capabilities: |
Claude is focusing on reasoning and multi-step tasks Perplexity is innovating on Deep Research and real-world deployment GPT-4.5 emphasizes conversational quality and emotional intelligence
|
For those of us building agents, this diversity is actually perfect. It means we have specialized tools for different types of agent experiences. |
At agent.ai, we've already integrated all three of these models. You can test and build with Claude 3.7 Sonnet, GPT-4.5, and Perplexity's Deep Research right now. The time between a new model release and its availability on our platform is now measured in hours, not days or weeks! |
With each new model release, the capabilities available to agent builders improve. The agents you build today can be smarter, more conversational, and better integrated than what was possible just last month. |
If you've been waiting for the right moment to start experimenting with building agents, these releases collectively represent another big step forward in what's possible. |
Time to build! |
—Dharmesh (@dharmesh) |
|
|
What'd you think of today's email?Click below to let me know. |
|
