Maya: Welcome to Cursor HQ. Hi, I'm Maya. Today: multiplayer Cursor Projects, Claude Fable topping CursorBench, a Gemini 4 Carbon discount, the Model X video plugin, and routing subagents to different models.
James: And I'm James. Let's get into it.
Maya: So Projects, Cursor's coordinator-led fleet of agents for big work, now goes multiplayer. You create one from the left-hand nav, invite teammates, then plan, build and ship alongside everyone's agents. Cursor says it's rolling out in beta to all users starting November tenth.
James: That's a real shift, Maya. Each teammate keeps their own one-to-one chat with their coordinator and worker agents, each with local context. On top sits a shared system holding requirements docs, diagrams, specs and mockups. A decision in one chat reaches every agent.
Maya: Exactly right, James. A shared task board shows who's doing what and the PR behind each task. It also changes onboarding: an invited teammate gets context pre-configured and up to date, so they skip building their own before putting agents to work.
James: And the coordination piece is agent-to-agent across teammates. If you add reviews to a product page while a designer redesigns it, Cursor says the coordinators spot the overlap early and sync, so neither person hits a merge conflict at the end.
Maya: Onto models. Claude Fable five point two, an Anthropic model, is now in Cursor. Cursor says at max effort it's the top model on CursorBench, their own coding benchmark, at sixty percent.
James: Good to flag the effort levels, Maya. That sixty is the max effort figure. At default effort Cursor says it scores fifty-six point nine, up from forty-nine point two for Claude Fable five point one.
Maya: Right, so that's a gain of seven point seven percentage points between versions at default effort. These scores come from Cursor's announcement on X, reported at those two effort levels, max and default.
James: Now a price move. Gemini 4 Carbon is fifty percent off in Cursor through November thirtieth. The promotional rate is two dollars per million input tokens and ten dollars per million output tokens, charged separately on each side.
Maya: And there's a boundary to watch, James. On December first those rates rise to four dollars per million input and twenty per million output. Work that continues past November crosses that line even if the model choice stays the same.
James: Exactly. And Cursor separates two decisions here: the promotional price is one thing, but for teams choosing an effort level it points to its evaluations page to compare every level for the work.
Maya: The Model X plugin for Cursor can carry an animation workflow from model comparisons into repeated video requests. In the demo, Cursor submits jobs through the Model X media API, waits for completion, and downloads results, extending a caveman story into two connected thirty-second scenes.
James: And the setup starts in the console, Maya. A reference image made with GPT image two becomes the starting point for comparisons between SeaDance two point five and Kling V three, same prompt, ten-second animation at seven twenty p.
Maya: Right, then you create an API key in the console, install the plugin, and run a diagnostic check. Generation still runs through the media API; the plugin isn't a separate video service. What it supplies is Model X-specific guidance for Cursor about the models and parameters.
James: And the demo used SeaDance for the longer scenes. The first extension continues from the caveman picking up a spear, preserving character and hand-drawn style; the second brings in an animal, then a confrontation and escape, so the direction has to carry through.
Maya: Last one. When you use your own API keys with a custom model, the helper agents the main agent spawns through the Task tool inherit the parent's model. Picking a different model per subagent doesn't work as expected.
James: And the forum workaround moves that choice to Magpie, a gateway you control, Maya. You configure one custom model pointing to a routing group, then a small Skill and always-applied Rule prepend a routing marker to the Task prompt, which regex rules match.
Maya: One honest limit, James. Asking a subagent which model it is isn't reliable, since Cursor may still report the logical name while the gateway sends it elsewhere. You read Magpie's request logs instead. And a missed marker can silently route to the default.
Maya: That's your Cursor HQ briefing.