Vibe Check: Anthropic Cooked on Claude Haiku 4.5

Was this newsletter forwarded to you? Sign up to get it in your inbox.

Anthropic just dropped Haiku 4.5—Anthropic’s newest “small” (aka least powerful, and cheapest—Claude model, and it’s great. We got our hands on it early and put it through its paces.

This is the story of the new Haiku, which just jumped version numbers from 3.5 to 4.5: It’s almost as powerful as the new Sonnet 4.5, faster, and much cheaper. And when I say almost as powerful, I mean it—I had a hard time telling the difference when testing it on complex queries like “given this P&L, analyze our Q3 performance.”

Here’s your day zero vibe check:

Everything you love about Sonnet 4.5 priced like Haiku

It’s priced at $1 per $5 per million input and output tokens. For comparison, GPT-5-mini is about $0.25 per $2.50 per million input and output tokens—and Gemini 2.5 Flash is around the same. So Haiku 4.5 still costs about four times more than GPT-5 mini or Flash.

But it’s about three times cheaper than Sonnet 4.5, and it performs surprisingly close to as well. That makes it a gift for developers.

Your agentic apps just got an upgrade

If you’re a developer building an agent, Claude models are the premium option. They’re the best at tool calling and running for long periods of time without going off the rails. But Sonnet 4.5 is really expensive.

When it first came out, we used Sonnet 4 in Cora’s email assistant, and it rocked. Then we got our Anthropic bill that month, and COO Brandon Gell told us that Every would go out of business if we didn’t figure out how to make it work with GPT-5-mini.

But as of today, we’ve switched back to Haiku because it works incredibly well inside of Cora, and it’s not priced like beluga caviar.

Cora general manager Kieran Klaassan tested Haiku head-to-head against GPT-5-mini on agentic search queries inside of Cora. Haiku was 44 percent faster than GPT-5-mini (on the priority tier)—returning responses to questions like, “How much did I spend on Ubers in Guadalajara?” and, “What was my first ever email?” in an average of 19.7 seconds versus 28.3 seconds for GPT-5-mini (Priority tier).

Chat built for how teams actually work

Team chat hasn’t changed much in a decade. It’s still cluttered, disconnected, and hard to focus in.

Glue is rethinking what chat can be if AI is built right into the conversation. Glue AI connects directly to your tools through MCP, so your team can plan, collaborate, and take action without switching context or losing focus.

Most chat tools were built for a pre-AI world. Glue is built for what’s next. Try it free.

Get a free trial of Glue

Want to sponsor Every? Click here.

Duking it out in AI Diplomacy

Claude Haiku 4.5 had an interesting performance in AI Diplomacy, where we pit AI models against each other in a reimagined version of the classic strategy game. Overall, it ended up performing about as well as Grok 4 fast (non-reasoning) and GPT-5 with minimal reasoning—not one of the top performers but able to hold its own.

In the prompt impact leaderboard for Diplomacy, Haiku displayed tendencies that other small models (like GPT-5 Nano and GPT-OSS 20B) share, where the optimized prompts that lead to dramatic improvements for some of the bigger models led to degraded performance for Haiku.

That’s consistent with our overall take that it's a capable, cheap, fast model that can be a useful workhorse.

Where Haiku struggles: Math and sycophancy

We found at least one notable instance where Haiku struggles. In the earlier example, where Kieran asked Haiku to add up his bill on Ubers in Guadalajara, Haiku did a great job of finding all of the relevant emails. However, it failed at the math:

When Kieran pointed out the error, Haiku told him he was right—and then proceeded to make the same mistake again. We recommend sticking with GPT-5 mini for workloads where mathematical reasoning or lack of sycophancy is important—or equipping Haiku with a tool for math problems it encounters.

The final verdict

If you’re a developer or founder building complex agentic apps with Sonnet 4.5, you should switch to Haiku immediately. You’ll save a lot of money and sacrifice very little performance.

If you’re building on Gemini 2.5 Flash or GPT-5-mini, we highly recommend trying out Haiku. While it’ll cost more, it will return better responses for use cases that require tool calling and autonomy.

With the release of Sonnet 4.5—which beat Opus at coding tasks—and now Haiku, Anthropic is perfecting packing its frontier intelligence into smaller and smaller form factors. Developers are reaping the benefits.

Want to learn Claude Code hands-on? Join us for a one-day Claude Code for Beginners course on November 19. Dan Shipper will walk you through setup, basic commands, and practical workflows you can use immediately—no coding experience required. The course launches tomorrow, and early bird pricing of $1,000 per person (which includes the price of an annual Every subscription) is now live. Existing Every paid subscribers get our lowest-ever price of $712.

Dan Shipper is the cofounder and CEO of Every, where he writes the Chain of Thought column and hosts the podcast AI & I. You can follow him on X at @danshipper and on LinkedIn, and Every on X at @every and on LinkedIn.

Kieran Klaassen is the general manager of Cora. Follow him on X at @kieranklaassen or on LinkedIn.

Alex Duffy is the cofounder of Good Start Labs.

We build AI tools for readers like you. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.