Anthropic just reset the price-to-performance curve for frontier models. The new Claude 3.5 Sonnet is not an incremental update; it delivers intelligence exceeding the previous top-tier Claude 3 Opus, but at twice the speed and a fraction of the cost. This isn't just a new model—it's a new baseline for what you should expect from a workhorse AI, especially for complex coding and agentic tasks.
what changed: flagship intelligence at mid-tier cost
The key takeaway is the compression of the intelligence-speed-cost tradeoff. Claude 3.5 Sonnet outperforms Claude 3 Opus on multiple graduate-level reasoning and coding proficiency benchmarks, including GPQA and HumanEval. But it's priced at the original Sonnet's rate: $3 per million input tokens and $15 per million output tokens.
For builders, the most significant metric comes from an internal agentic coding evaluation. Given a natural language description of a bug or feature, Claude 3.5 Sonnet solved 64% of the problems. Claude 3 Opus solved 38% on the same test. This isn't just a benchmark win; it's a step-change in reliability for autonomous code manipulation tasks like updating legacy applications or migrating codebases.
It also operates at twice the speed of Claude 3 Opus, making it viable for more latency-sensitive applications like context-aware customer support and orchestrating multi-step workflows.
how to use it: api access and the new artifacts ui
Accessing the model is straightforward. It's available through the Anthropic API, as well as on Amazon Bedrock and Google Cloud's Vertex AI. The integration is a simple model string update.
import anthropic
client = anthropic.Anthropic(
# defaults to os.environ.get("ANTHROPIC_API_KEY")
)
message = client.messages.create(
model="claude-3-5-sonnet-20240620",
max_tokens=4096,
messages=[
{
"role": "user",
"content": "Write a Python script to analyze a git repository and identify the top 5 contributors based on commit count."
}
]
)
print(message.content)
More interestingly, Anthropic also launched a new feature called Artifacts on Claude.ai. When you ask the model to generate content like a code snippet, a document, or a website design, it appears in a dedicated window next to the conversation. You can see, edit, and build upon the generated content in real-time. This transforms the interaction from a simple chat to a collaborative workspace, integrating the AI's output directly into your workflow without constant copy-pasting.
the vision and agentic coding leap
Beyond raw intelligence, Claude 3.5 Sonnet is now Anthropic's strongest vision model. It surpasses Opus on standard vision benchmarks, showing marked improvement in interpreting charts, graphs, and transcribing text from imperfect images. This has direct implications for applications in logistics, finance, and retail that need to extract structured data from visual inputs.
The jump in agentic coding performance is the real story for many of us. The ability to independently write, edit, and execute code with sophisticated reasoning is what we've been chasing. The 64% solve rate on Anthropic's internal eval suggests a higher degree of reliability for tasks that require understanding an existing codebase, reasoning about changes, and implementing them correctly. This makes it a more viable candidate for building agents that can genuinely offload development tasks, not just generate isolated snippets.
so what this week
The release of a model that is simultaneously better, faster, and cheaper than the previous flagship is a significant event. For builders, it's an immediate signal to re-evaluate your model stack. Workflows that were too expensive or slow with Opus-level models may now be practical with Claude 3.5 Sonnet. The improvements in coding and vision open up new possibilities for more complex, autonomous agents. The tradeoff curve has shifted, and your default model choice for hard problems should probably shift with it.