Cursor attacks agent costs with an India plan at ₹649 and a model router

Cursor's latest release notes read like an answer to one question: what should agentic coding cost? The two biggest items are an India plan at ₹649 a month and an Auto mode that now routes every request through a model router to the cheapest model that can handle it. Everything else, from iPad PR reviews to agents in Slack, is the layer you use to keep an eye on the workflow.

The release is unusually focused on spend: local pricing for India, a router that favors cheaper models, and a tier built around daily agent use. It is not the only coding tool chasing cheaper agent runs; OpenCode's lite tier cuts coding costs by 80%.

The ₹649 bet on India

Cursor Start, at ₹649 a month including taxes, opens to Indian developers on July 28, 2026. Billing is monthly with auto-renewal, charged in INR through UPI or a bank card. Existing Indian users on the free tier upgrade from the dashboard; new users pick Start during onboarding at cursor.com/signup.

The package bundles Grok 4.5 at a fixed medium effort level and no fast mode. Composer also ships without fast mode. You also get a usage volume Cursor says is enough to build with agents daily, always-on cloud agents that create, test and deploy code, Cursor for iOS with remote control, and the usual extension surface: plugins, MCP servers, hooks and skills.

Auto mode now thinks about the bill

The more consequential change sits under the hood. Auto mode now relies on Cursor Router, which classifies every request by task type and complexity, then sends it to the model best suited for the job. Frontier models take the tasks that demand them; the cheaper ones get the rest. Per-call routing is usually an afterthought in agentic systems, which is exactly the blind spot TRACE-Router was built to fix.

Three optimization modes place you at different points on what Cursor calls the cost-intelligence Pareto frontier. Intelligence targets quality comparable to the most expensive models, the kind usually too costly for daily use. Balance targets the quality most users pick day to day. Cost keeps quality reasonable while trimming token spend. Balance and Intelligence are billed at whatever rate the routed model carries.

Administrators can enable the router per team or group, restrict which modes members can use, set a default, and allow or block underlying models. The routed model is hidden by default. Auto can be enforced softly or strictly. One detail stands out: Grok 4.5 is required as the cost-optimized routing option. That tracks with xAI's rollout of the model across coding tools. The router ships on desktop, web, iOS, CLI and SDK, defaults on for Teams, and Enterprise admins turn it on from the dashboard.

Your phone is now a merge tool

On mobile, Cursor for iPad is now included in all paid plans. The iPad layout uses the extra space: pinned sidebar chats let you track several agents running at once, split screen keeps a review open next to a chat, and file diffs display in full. Annotations get room too. Attach a screenshot, tap a precise point to comment, or draw straight onto the image with Apple Pencil.

On iPhone and iPad, the review screen now covers the entire PR: comments, checks and approvals. You can add or change reviewers, read comments and ask the agent to resolve them. The full flow, from the agent's output to merging the PR, stays on the device. An inbox sorts what's in progress, what needs attention and which PRs are under review. The update also adds Bitbucket and Azure DevOps SCM support, multi-PR sessions that open every PR a chat created rather than just the last one, and in-app team switching. The mobile push extends the iOS beta that turned a phone into an agent dispatcher.

Slack agents present a plan before they act

Cursor in Slack now shares its plan before it starts, so you can redirect it early, and updates its status while it works. It runs in named multi-repo environments. If your front end, back end and shared code live in different repos, Cursor picks an environment with access to all of them. When it needs a repo outside that setup, a Change repo button gives you an escape hatch. It can read and send messages in other Slack channels and threads, pull context from across the workspace and post updates back to the original thread. The old inline buttons are gone, replaced by compact footer links.

Back in the editor, parallel chats let you ask questions without interrupting the main agent. They start with /side, /btw or the + button, inherit the main chat's context, and stay open as persistent conversations you can mention later with @. By default they focus on reading, searching and answering, which is a diplomatic way of saying: don't derail the plan. Agent transcripts are searchable with Cmd+K through a local index, and Cmd+F searches within one conversation. Cloud agents also gain new hooks, including beforeSubmitPrompt, afterAgentResponse, afterAgentThought, stop and subagentStart. The hooks are aimed at observing reasoning and building self-correction loops.

The repo and project selectors got rebuilt too. A No Repo option replaces the old Home state, an explicit Run on picker chooses Cloud, This Computer or Remote Machines, and branch selection opens on default and recently used branches.

What it adds up to

No single feature here is the story. What they add up to is a tool that stopped being a local editor. The model has become a cost to route, the agent no longer needs your laptop, and the interface follows you to your phone and your Slack workspace. The direction has been visible since Cursor 2.0 rebuilt the IDE around agents. The ₹649 plan is how Cursor makes that pitch work in India, with local pricing and UPI keeping payments simple. The router is how it keeps the bill down for everyone else.