Cloudflare’s new model reads messy data faster than normal LLMs

Welcome back. Cloudflare is back with another model for your workflow. Just weeks after shipping Clef, its decision model, it’s added a sibling that claims to handle your messy inputs even better. Meanwhile, a new personal agent wants to make it easier for you to collaborate.

Also: Vercel’s CEO on how dev workweeks are changing, a senior dev’s guide to fixing your prompting mistakes, and a viral story that suggests big tech interviews mostly come down to luck.

Cloudflare’s new model reads messy data faster than normal LLMs: The web infrastructure giant just previewed Clairic, an extraction model that skips text generation. Where its sibling Clef picks a decision from options you define, Clairic pulls typed values straight from the input in one forward pass and returns null when the value isn’t there. When you run both models before an LLM, the LLM receives clean data directly and never has to parse the messy original input. Cloudflare says this setup was 3.2x faster and 49% cheaper than using an LLM alone.

Your coding agent just got a creative studio: Two releases are giving agents more to do. The Sunra MCP server puts leading image, video and audio models inside Claude Code, Cursor, and Codex with one npm install. You can generate app assets, demo videos, and voice-overs without leaving your agent. Meanwhile, Replika’s founder just launched Wabi 2.0. It is a multiplayer personal agent that builds mini-apps inside your group chats.

Inception bets on diffusion to speed up agents: The AI startup just dropped Mercury 3, a new LLM that drafts answers in parallel at 1,068 tokens per second. That means less waiting on your agents, plus a 1M-token context window to work with. Inception claims it beats GPT-6 Luna and Claude Haiku 5.5 on agentic coding, though Haiku still wins on longer coding tasks. It’s live on OpenRouter, Vercel, and Baseten, with launch prices starting at $0.10 per million input tokens.

Your dev stack got an AI upgrade everywhere except the input layer. You're still typing every prompt, every ticket, every review comment by hand.

Wispr Flow closes that gap. Dictate into Cursor, VS Code, Slack, Linear, or anywhere else you work. It's syntax-aware: camelCase, snake_case, acronyms, and file names all come through clean. Mention a file in Cursor or Windsurf, and it auto-tags.

It's the voice layer for an AI-native workflow. Speak your intent. Your tools do the rest.

Available on Mac, Windows, iPhone, and Android. Used by millions of developers, including teams at OpenAI and Mercury.

Try free

Cheap, but unreliable. Smaller AI models are tempting for coding tasks because they cost less, and teams rarely choose a model on raw intelligence alone. But they often struggle with consistency. Specmatic founder Naresh Jain wanted to know if a behavioral harness (a set of specs, examples, and tests that flag mistakes as an agent codes) could close that gap.

40 builds. To test this, he decided to run an experiment. He had Codex build the same Java backend with OpenAI’s flagship GPT-6 Sol and cheaper GPT-6 Luna. Each model got 10 attempts on its own and 10 more backed by the Specmatic harness. Then each build faced 178 hidden acceptance tests.

Luna needed a lift. Left to itself, Luna passed just 3 of 10. With the harness, it passed all 10, matching Sol’s perfect record for $0.273 per run, about 33% less than unaided Sol’s $0.406. Giving Sol the same help changed nothing except its bill, which jumped about 7x. The extra guidance paid off only for the model that needed it.

Check smarter. Jain’s takeaway is that you should tailor your testing strategy to your model’s capabilities. Every model needs an independent check, but some need more guidance along the way. Here’s his advice:

Starting Point. Jain shared the full application brief, so you can rerun the comparison with your own models and decide what works for you.

AI tools already handle critical work across your company.

Most of them run quietly on personal accounts, connecting to sensitive production systems without anyone noticing until something breaks.

Major instantly turns vibecode built by anyone on the team into production-ready software with SSO, secrets management and auth that you can centrally govern. 

Your team gets the freedom to build with AI, and you keep full visibility and control over what goes live.

Code readers, take a demo to claim $200 in tokens at major.build.

How to build always-on agents with Pi Durable: This tutorial from Hugging Face’s Alejandro AO shows how you can use Pi Durable to help agents survive crashes and pick up where they left off. You’ll learn how checkpoints, conversations, and storage fit together, then build a local setup with llama.cpp and Qwen that can resume interrupted work from SQLite.

Pele: Obsidian creator kepano’s tiny open-source alternative to Mermaid.js for rendering diagrams in your app. Feed it Mermaid syntax, and it returns themeable SVG, with zero dependencies, a much smaller bundle, and claimed render speeds up to 500x faster.

How to review your agents for security issues (by Anthropic): Learn how to audit finished agent sessions for things like prompt injection, unauthorized hosts, leaked credentials, blocked calls that get retried another way, and data sent somewhere it shouldn’t go. It turns the logs into a ranked report with the exact evidence worth a human looking at.

Your first users hit bugs you never saw, then leave without telling you why. A dev shared a Claude Code prompt that spins up five subagents acting as your worst first-time users before launch.

Each subagent clicks from landing page to first success, screenshots every point of confusion, and merges findings into one ranked bug list. Claude fixes everything in one pass, then reruns all five to verify.

A dev used Claude Code to strip his smart TV of ads, bloatware, and lag, and the demo is phenomenal. Another dev tried it too and posted the prompt.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team