ChatGPT Just Changed What an Answer Can Look Like

Morning, {{first name | folks}}! AI is getting cheaper to run, easier to prototype with, and more capable of handling work on its own. This week’s biggest stories show where those changes are starting to show up in real products and workflows.

OpenAI is rolling out GPT-6 to ChatGPT users, bringing a new feature called Intelligent UI. Instead of returning only text, ChatGPT can now build interactive responses with charts, forms, visual explanations, and even simple tools or games inside the conversation. GPT-6 also starts answering while it continues to think, rather than making users wait for the full response.

The bigger change is how ChatGPT handles tasks. A question can now produce the interface needed to work through it, whether that’s a calculator, an interactive explanation or a planning tool. OpenAI says GPT-6 is rolling out globally to paid ChatGPT plans first, with free and go users getting access from October 8.

Anthropic has launched Claude Haiku 5.5, its fastest and most capable Haiku model yet. It is designed for coding, computer use, customer support, classification, and other high-volume workloads, with pricing starting at $0.10 per million input tokens and $0.50 per million output tokens.

The bigger shift is the cost of running capable models at scale. Anthropic says Haiku 5.5 delivers roughly twice the coding performance of Haiku 4.5 while costing 75% less on average. That makes it much more practical for workloads where running a larger model on every request would be too expensive.

Google is testing Playground, a new experiment that turns text descriptions into playable browser games. Users can describe a game idea, set the rules and mechanics, and then make changes to the result without writing the underlying code.

Playgrounds can generate different types of games, including trivia, racing, and tower defense, and let users create and share their own. The experiment gives Google another way to test what happens when building a working prototype becomes as simple as describing what you want.

Arena tested Jev Router from Typesafe AI across more than 4,700 agent sessions. It generally picked strong models, but the trade-off was cost and speed. Matching DeepSeek V4.1 Flash (Max) cost 38% more, while median request latency was 1.7x higher.

Where Jev did stand out was how well it responded to user feedback. It scored +10% on steerability, close to Claude Opus 5.5 at +10.48%. So the router isn’t beating the frontier on efficiency yet, but it is getting better at changing its model choices based on what users actually want.

A three-month NBER study gave 133 patent lawyers across 11 U.S. firms access to an AI drafting assistant. With AI, their patent work scored 0.34 standard deviations higher after 10 days and 0.38 after 90 days, with larger gains among junior lawyers.

The researchers then gave everyone a patent review task without AI. Lawyers who had used the tool scored 0.32 standard deviations higher overall, but the improvement came entirely from senior lawyers, who scored 0.45 standard deviations higher. Junior lawyers showed no average gain, although their results split between more very good and very poor scores.