29 May 2026

The leaderboard broke. Pins got cheaper.

DeepSWE exposed a wider gap between coding models, while my Pinterest automation experiment came in at less than one cent per pin.

---

Hello,

Two posts this week: one big-picture sweep of a turbulent few days in AI, and one very practical deep-dive into what it actually costs to automate Pinterest with Claude.

The first is about what happens when AI benchmarks stop being polite. The second is about the small but satisfying joy of proving that an automation idea will not quietly eat your budget while you make tea.

Grab a cuppa.

---

The AI Coding Leaderboard Just Broke — Here’s What It Means https://daehnhardt.com/blog/2026/05/29/the-ai-coding-leaderboard-just-broke-here-s-what-it-means/

Think of AI benchmarks like standardised school exams: useful in theory, but a little too easy to teach to the test.

This week, Datacurve’s DeepSWE arrived to spoil the curve. Its tasks are larger, its prompts are shorter, and the setup is closer to how developers actually delegate work to AI coding agents. GPT-5.5 scored 70%. Claude Opus 4.7 followed at 54%. Claude Haiku 4.5, which scored 39% on SWE-Bench Pro, collapsed to 0%.

That would be interesting enough on its own. But the same week also brought a larger backdrop: OpenAI moved toward a possible trillion-dollar IPO, Pope Leo XIV published 42,300 words on AI and human dignity, Illinois passed a frontier AI audit bill, and Anthropic’s Project Glasswing found more than 10,000 high- or critical-severity vulnerabilities in critical software.

The signal I took from all of this the benchmark race is still real, but the world around the benchmarks is changing just as quickly.

---

I Used Claude to Generate Pinterest Pins. Here Is the Actual API Cost. https://daehnhardt.com/blog/2026/05/28/pinterest-pin-generator-api-cost/

There is a particular kind of programmer’s dread that appears before looping a script over your entire post archive — not “will it work?” but “will it silently charge me forty pounds while I am making tea?”

I built a Python script that reads Jekyll blog posts, sends them to Claude for SEO-optimised Pinterest metadata, and renders a 1000×1500 JPEG ready for upload. Before letting it loose on hundreds of posts, I calculated the exact API cost.

The answer: roughly 0.7 cents per pin.

For longer posts, the smarter move is not to send the whole article. I extract the first 2,000 and last 2,000 characters — the introduction and conclusion — so Claude sees both the promise and the payoff of the post at a fixed, predictable cost.

Switching to Anthropic’s Message Batches API halves the bill again: 1,000 pins go from about $7.04 to $3.52. The surprise bottleneck is not Claude at all, but waiting for Pinterest to approve API access.

---

Until next time, Elena

---

← All issues