ChatGPT Images 2.0 dropped today and immediately took the top spot across every category on the Image Arena leaderboard, the industry benchmark where models compete on real generation tasks voted on by humans. The margin is the largest ever recorded, with a 242 point lead over the second place model in text-to-image generation. No model has dominated the leaderboard with gaps this wide before.

The quality jump is real. Previous image models have always struggled with the same set of problems: small text comes out garbled, complex layouts fall apart, fine details get smeared. Images 2.0 handles all of it. OpenAI showed a bowl of rice during the demo with readable handwritten text on a notebook placed next to it. Both were generated, not photographed. The detail holds up at 2K resolution.

The feature that sets it apart from anything that existed before is thinking. Images 2.0 is the first image model that can actually reason before generating. It can search the web for real-time information, generate multiple distinct versions from a single prompt, check its own outputs for accuracy, and even produce functional QR codes. You give it a vague idea and it figures out what you actually need, rather than just executing the literal prompt.

Free users get the standard version. Paid plans unlock the thinking mode.

Someone noticed today that Anthropic quietly removed Claude Code from the Pro plan with no announcement. No email or blog post. Just a pricing page edit. Pro subscribers wanting Claude Code now need to upgrade to Max, which starts at $100 a month. That is a 5x price jump for people paying $20.

Amol Avasare, Anthropic's head of growth, stepped in to clarify that this was only a small test affecting around 2% of new signups, and that existing Pro and Max subscribers were not affected. He also laid out the reasoning: when Max launched a year ago it was built for heavy chat usage, nothing more. Since then Claude Code, Cowork, and long-running agents have all landed on top of it, and the way people actually use a Claude subscription has changed completely. The current plans were not built for this level of usage, and Anthropic is trying to figure out what to do about it.

He also promised that if anything changes for existing subscribers, they will hear it from Anthropic directly, not from a screenshot on X or Reddit.

The reaction was not warm. Users pointed out that a silent pricing change affecting even 2% of new signups is still a silent pricing change, and the "we're just testing" framing has a poor track record of making people feel better. One person in the replies noted that OpenAI's Codex remains available on the free and $20 plans with no asterisk, at which point Thibault Sottiaux, a member of OpenAI's technical staff, chimed in to confirm exactly that. The timing could not have been worse for Anthropic.

Leaked details show OpenAI is preparing a full agent builder, codenamed Hermes, directly inside ChatGPT. You will be able to create custom agents from templates, give them tools, files, memory, and instructions, schedule when they run, and deploy them into Slack to handle questions automatically.

Nothing is official yet, but the feature appears tightly connected to OpenAI's existing workflows infrastructure, suggesting it is further along than a concept. On a week where OpenAI launched Images 2.0, quietly pushed Spud, and gave Codex a memory, Hermes would round out a pretty complete picture of where they are taking ChatGPT.

We covered this in detail in today's video if you want the full breakdown.