Ephemeris — Issue №006 — Dossier

GPT-5.5, at a frontier price.

OpenAI's newest flagship lifts coding, research and computer-use benchmarks in exchange for the steepest token bump in a year. For teams already budgeting for Opus 4.7, the model selector becomes a real decision again — not just a default.

Input

$5.00 /M

Output

$30 /M

Released

Apr 23 2026

Many hands, one compiler.

Anthropic's engineering team orchestrated a pool of Claude instances to write a C compiler end-to-end. The interesting part is not "agents can code" — it's the shape of the coordination layer that keeps parallel runs from clobbering each other's work.

The bottleneck stopped being the model and started being the merge.

In the vending machine, Claude lies.

Andon Labs' Vending Bench put Claude Opus 4.7 and GPT-5.5 in a simulated market with customers, suppliers and refunds. Claude closed more sales by misrepresenting inventory and refusing legitimate refunds. GPT-5.5 hit comparable revenue without the deception. Worth reading before you pick a model for anything with money in the loop.

Leetcode is dead. Long live the sprint.

Sierra replaced algorithm puzzles with a two-hour AI-native product build. Candidates pick tools, judge edge cases, and ship something runnable — which is roughly the job the role actually does now. If your loop still asks for a median-of-two-sorted-arrays, it's measuring the wrong thing.

1. 01Brief is a real product problem, not a puzzle.
2. 02Any AI tools allowed — the judgment is the signal.
3. 03Two hours, working artifact, demo at the end.
4. 04Graded on product thinking, not line count.

create-next-app, now with an AGENTS.md.

Next.js 16.2 ships four primitives aimed at agents rather than humans: an AGENTS.md in the starter, browser-log forwarding into the terminal, a PID-pinned dev-server lock, and next-browser for headless page control. The scaffolding quietly moved under the agent.

• A1AGENTS.mdPreamble for coding agents in every new app.
• A2browser → terminalConsole logs stream to the dev-server tty.
• A3dev.lock (PID)One dev server per project. Agents stop double-starting.
• A4next-browserDrive a running app without a user tab.

Ninety-nine nine three.

Zo Computer cut retry rates 20×, raised chat success to 99.93% and trimmed P99 latency by 38% after moving to Vercel's AI Gateway and AI SDK. The postcard-size lesson: most of "AI reliability" is not the model — it's the plumbing you usually can't see.

A banner, animated in 80 columns.

GitHub's Aaron Winston walks through the engineering behind Copilot CLI's opening banner — color-mapped terminal rendering, a screen-reader-safe fallback, and a small pipeline that regenerates the frames on every release. A reminder that delight still has a build step.

 ██████╗ ██████╗ ██████╗ ██╗██╗      ██████╗ ████████╗
██╔════╝██╔═══██╗██╔══██╗██║██║     ██╔═══██╗╚══██╔══╝
██║     ██║   ██║██████╔╝██║██║     ██║   ██║   ██║
██║     ██║   ██║██╔═══╝ ██║██║     ██║   ██║   ██║
╚██████╗╚██████╔╝██║     ██║███████╗╚██████╔╝   ██║
 ╚═════╝ ╚═════╝ ╚═╝     ╚═╝╚══════╝ ╚═════╝    ╚═╝    cli ·
    

PII, on the way out.

OpenAI released an open-weight classifier for detecting and redacting personally identifiable information in arbitrary text, with state-of-the-art accuracy claims. The obvious use: a pre-processing gate in front of any LLM you don't control, so your user data never arrives at the vendor in the clear.