OpenAI just upgraded Deep Research with GPT-5.2 and added website-specific search. You can now lock research to trusted sources instead of letting it crawl the entire web. The interface got an overhaul too for full-screen reports, real-time progress tracking, and the ability to interrupt mid-research to add sources or redirect. Meanwhile, OpenAI's celebrating 1 million Codex downloads in week one while quietly warning that free tier limits are coming. Classic move: build demand, then meter access.

Today in AI:
→ Opus v Codex: Why benchmarks stopped mattering
→ OpenAI's free Codex tier won't last (1M downloads in week 1)
→ How To Use Claude Code Insights To Fix Your Workflow

Let's dive in..

Opus 4.6 and Codex 5.3 dropped within days of each other. Both companies hyped their numbers. Both hit new highs on SWE-Bench and coding evals. And when researchers used them, the benchmarks meant nothing.

Codex 5.3 wins on paper. It's better at finding bugs. Faster. Uses fewer tokens. Beats Opus on most evals.

Opus 4.6 wins in practice. More reliable on real tasks with better UX. Doesn't ignore instructions when juggling multiple requests. Claude Code's interface makes it easier to actually get work done.

The gap between "scores well" and "works well" is now massive. Gemini 3 topped benchmarks and basically nobody uses it. Codex technically outperforms Opus on evals but loses on adoption. Why? Claude's product is just more usable.

Three things replace the benchmarks:

Both models regressed on instruction following when handling queued tasks. That's a step backward disguised as a capability upgrade. The companies chasing benchmark points missed what users need.

We're in a world where the "worse" model wins if it ships with better UX. Product execution beats model capability. And the launch playbooks haven't caught up yet.

For your team: Stop asking which model scores better? Start asking which one do people actually use? If you're building on frontier models, test with real workflows, not evals.

Kajabi was built for people with earned expertise. Coaches, educators, practitioners, and creators who developed their wisdom through real work and real outcomes.

In a world drowning in AI-generated noise, trust is the new currency. Trust requires proof, credibility, and a system that amplifies your impact.

Kajabi Heroes have generated more than $10 billion in revenue. Not through gimmicks or hype, but through a unified platform designed to scale human expertise.

One place for your products, brand, audience, payments, and marketing. One system that helps you know what to do next.

Turn your experience into real income. Build a business with clarity and confidence.

Kajabi is where real experts grow.

Start Your Free Trial

Codex hit 1 million downloads in its first week. Now OpenAI's warning: the free tier won't last.

Sam Altman posted the update himself. "We'll keep Codex available to Free/Go users after this promotion; we may have to reduce limits there but we want everyone to be able to try Codex and start building." Translation: if you're not paying, get ready for throttling.

Here's what's happening:

The rush: 60% week-over-week growth in Codex users after the app launched. Developers downloaded it faster than any OpenAI tool in history. The desktop app, parallel agents, and scheduled tasks hit exactly what people wanted.

The reality check: Running agents for 30 minutes per task at scale costs real money. Free and Go tier users are about to hit stricter limits. Paid subscribers already get doubled rate limits and will stay prioritized.

The bigger picture: This is OpenAI testing demand before locking down pricing. They did the same dance with GPT-4 and DALL-E. Give it away, build dependency, then meter access. Classic SaaS playbook.

Meanwhile, ChatGPT Deep Research got a quiet upgrade worth knowing about:

For your team: If you're using Codex heavily on the free tier, budget some cash before the limits hit. Don't get caught mid-project when throttling kicks in.

Claude Code just shipped a command that reads your last month of work and tells you what you're doing wrong. It's called /insights, and it's basically a performance review for how you use AI.

Here's how it works and why it's useful:

The report has four sections:

Why this is useful:

Most people use 20% of Claude Code's capabilities. The /insights report shows you the 80% you're missing, with copy-paste examples specific to your work.

📄 ChatGPT Deep Research got a proper document viewer. Full-screen mode with table of contents on the left, citations on the right. Export to Markdown, Word, or PDF. Plus: GPT-5.2 model upgrade, website-specific search, and real-time research controls.

🇨🇳 Five Chinese AI labs racing to ship before Lunar New Year. Zhipu, MiniMax, ByteDance, Alibaba, and Moonshot are all debuting next-gen models in February. GLM-5, Doubao 2.0, Qwen 3.5, Kimi K2.5 all dropping within weeks of each other. The most concentrated period of Chinese AI releases ever.

❄️ Snowflake launched Cortex Code. An AI coding agent that understands your Snowflake data context. Ask in plain English, get SQL, dbt pipelines, or admin actions. Unlike generic coding assistants, it knows your databases, roles, and governance.

For a full list of 1500+ AI tools, visit our Supertool Directory