The 2-minute brief for AI builders.

Presented by No Code MBA

👋 Welcome to The 2-minute brief for AI builders.

In today's edition:

• OpenAI says Codex reached 7M+ weekly users and shipped 150+ updates in two months
• Perplexity open sourced WANDR, a 500-task benchmark for wide-and-deep research agents
• Vercel opened AI Gateway leaderboard data under CC BY 4.0 with daily updates and API export
• OpenRouter shipped a bigger MCP update so agents can discover, rank, test, and report on models from your editor
• Supabase shared how a 220-person company runs 20 to 25 projects powering 200+ internal apps with Claude Code and Vercel

🛠️ Codex keeps adding features — and usage is already big

OpenAI’s developer team says Codex now has 7M+ weekly users and has shipped 150+ updates in two months.

• The update list includes GPT‑5.6 and Ultra, parallel work with /goal, faster computer use, AppShots, inline edits, Sites, and mobile and SSH workflows.
• OpenAI also highlighted a real build example: command-line tools for evaluating new models on Notion’s architecture were built from start to finish with GPT‑5.6.
• This is a signal that coding agents are moving beyond demos and into day-to-day product work.

What it means: If you build products, you should treat agentic coding as a live workflow now, not a future bet. The takeaway is simple: test Codex for parallel tasks, quick edits, and remote workflows, then decide where it fits in your stack before your competitors do.

Read more on OpenAI Developers →

🔎 Perplexity open sourced a benchmark for research agents

Perplexity says it has open sourced WANDR, an internal benchmark used for deep and wide research inside Perplexity Computer.

• WANDR includes 500 research tasks that require 170,495 source-backed records across three tiers of difficulty.
• It focuses on two agent skills: searching broadly enough to find all qualifying entities and going deep enough to support every claim with evidence.
• Perplexity says WANDR is built from de-identified production use cases like competitive research, due diligence, literature review, market analysis, product comparison, and talent sourcing.
• Instead of checking against a fixed gold answer, WANDR re-fetches every cited page and checks each claim against the underlying evidence, so it can handle time-varying facts.
• Perplexity also says the pipeline doubles as a semi-automated factory for training data.

What it means: If you're building research agents, this is useful in two ways: you can benchmark quality on work that looks like actual business research, and you can copy the eval pattern itself. Evidence-backed evals are becoming a product advantage, not just a research exercise.

Read more on Perplexity →

📊 Vercel opened real model usage data

Vercel says AI Gateway leaderboard data is now open, covering production usage across models, labs, apps, and providers.

• The data is updated daily.
• You can export it as CSV or JSON from the leaderboards or an API endpoint.
• Vercel says the dataset is available under CC BY 4.0, so you can download, query, cite, and share charts.
• This gives builders a more grounded view of what people actually use in production instead of relying only on vibes or static benchmarks.

What it means: If you advise clients, route traffic across models, or build AI products, this is a new input for model selection and content. You can use the data for pricing pages, routing logic, market analysis, or public dashboards without starting from scratch.

Read more on Vercel →

🧩 OpenRouter made its MCP more useful inside your editor

OpenRouter says it shipped a batch of MCP updates so your agent can discover, rank, test, and report on models without leaving your editor.

• There are now 30 task categories in its new task type insights, showing what people actually use AI for based on real OpenRouter traffic, updated weekly.
• Model search can now filter by output price, model age, intelligence and coding scores, tool-calling success rate, and sort by Design Arena ELO.
• Send-message can now pin a provider, which OpenRouter frames as the difference between an eval and a dice roll.
• It also added generate-image and send-feedback tools, moved docs search to Mintlify, and removed permission prompts for read-only tools.

What it means: If you build agents or multi-model products, this is workflow glue. You can compare providers, keep evals stable, and test routing choices from inside your existing dev environment instead of bouncing across dashboards.

Read more on OpenRouter →

🏢 A 220-person company is letting every team ship its own AI apps

Supabase shared a customer story showing how delight.ai uses Supabase, Vercel, and Claude Code so each team can build production apps without pulling engineers off the core product.

• delight.ai is a 220-person company.
• It runs one Supabase project per team, or about 20 to 25 projects total.
• Those projects power more than 200 internal apps.
• The operating model matters as much as the tools: non-engineers can ship production tools while the main product team stays focused.

What it means: This is a practical blueprint for internal AI app sprawl that doesn’t turn into chaos. If you're a founder or operator, the lesson is to give teams their own lanes, shared infra, and AI coding help so internal tools stop waiting in the engineering backlog.

Read more on Supabase →

👋 Thanks for reading. If you liked this roundup, forward it to a friend!

Want to advertise on this newsletter? Click here.

This newsletter was created with the assistance of artificial intelligence. While we strive for accuracy, please note that AI-generated content may occasionally contain errors or biases.