What's New in Cursor in 2026

Cursor's current release wave adds Origin code hosting in early beta and an always-on cloud-agent harness: Subscriptions, Custom modes, subagent VMs and /goal. Earlier in the wave: Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition., Router modes, Cursor StartAn India-only individual plan centered on Cursor Models, with cloud and mobile access but without Other Models, Auto, Bugbot, Automations or the Agent SDK. Press Enter for the full definition. in India, iPad, Google Workspace pluginsSeparate Cursor plugins for Gmail, Google Drive and Google Calendar that let an authorized agent read or change workspace data. Press Enter for the full definition., Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition., Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition. and Cloud Agent BuildsPrepared Cloud Agent environments built in the background, with the last successful build kept active when a newer build fails. Press Enter for the full definition.. Use this page for the dated release sequence, then follow the canonical guide for setup, pricing, privacy, or model selection instead of treating a launch note as permanent documentation.

Knowing what shipped is the easy part. Turning latest Cursor releases into something a team relies on takes a few deliberate steps, mapped below.

1. 1Start with the job: code hosting, always-on cloud agents, persistent cross-app work, model choice, cloud setup, mobile review, integrations, plan access or admin policy.
2. 2Use the dated launch entry to learn what changed, then open the canonical guide for the current behavior and boundaries.
3. 3Read the primary release page and current documentation before quoting mechanics.
4. 4Update prompts, policies, training and pricing examples together so learners see one current story.
5. 5Re-check dates and price claims before using the content in sales, security or procurement conversations.

New capability comes with new ways to get it wrong. Keep these boundaries in place as you adopt it.

This release wave touched a lot of surfaces at once. Each one has its own guide with the detail and the sources, so jump to whichever you actually need.

Two ways, both built to dodge a stuffed context window once you pass roughly 80%. In a new chat, @past chats pulls compressed context from a prior conversation. It smushes the earlier context as concise as it can and drops the completed external and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. tool calls, so you spin a fresh thread off the gist instead of the full history. The second is fork a chat, which takes a prior chat's context and starts a new chat from it.

Because the team expects review to be the next bottleneck. Every agent change is already inspectable through a review button at the bottom of the chat that shows all edits the agent made, including the ones that built a chart or touched a config. The bet: the more code AI generates, the more it has to be reviewed, so review becomes where the work piles up.

A few directions were flagged as forward-looking, not released. Treat them as signals, not features you can use today.

• Simulators in cloud agents for iOS and Android, a deeper integration than today's MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. cross-app bridge. Because Android is Linux-based, someone on the team already got a cloud agent to run an Android simulator with a lot of manual instructions, but a no-hand-holding version is not rolled out and has no date.
• Customer-customizable PR security loops. Cursor runs security bug-finding internally with a security BugbotCursor's automated PR reviewer that posts inline findings and can push fix commits from isolated VMs. Press Enter for the full definition., and teased unannounced features to let customers build a validation loop on top of PRs.
• More sandbox controls, actively being built to give users finer say over what the agent can and cannot run.
• A very exciting update to the Supabase and Postgres integrations was mentioned without detail. Verify the specifics against the changelog before relying on it.

Start with the changes that act on you whether or not you act on them. Cursor Router shipped on July 22 and is on by default for Teams plans. Auto is routed per request, and only Cost keeps the bundled Auto Cost treatment. Balance and Intelligence bill the model that served the request, with the Cursor Token Rate when that model is third party. Nobody on the team has to change a setting for the billing basis to change underneath them.

On Enterprise the same feature inverts. The router is off until an admin turns it on, so Enterprise has a decision sitting unmade while Teams has a default to go and check. Enterprise model access controls interact with it too, since the router routes around blocked models and Cursor says blocking too many reduces routing quality and can disable routing altogether.

The August wave needs a different sort. Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. and Origin are early betas with plan gates (Origin adds a cloud-storage gate too, since legacy privacy mode excludes it), so a team should decide its data and access boundary before trying either. Google Workspace pluginsSeparate Cursor plugins for Gmail, Google Drive and Google Calendar that let an authorized agent read or change workspace data. Press Enter for the full definition. can act in Gmail, Drive and Calendar, which makes permission design part of setup. Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition., Cloud Agent BuildsPrepared Cloud Agent environments built in the background, with the last successful build kept active when a newer build fails. Press Enter for the full definition., iPad review, Cursor StartAn India-only individual plan centered on Cursor Models, with cloud and mobile access but without Other Models, Auto, Bugbot, Automations or the Agent SDK. Press Enter for the full definition., Subscriptions, Custom modes and /goal are opt-in capabilities you evaluate for a concrete job. The larger from-scratch model remains an announcement rather than a tool to plan around.

That four-way sort is most of the value in a release page: default-on, gated beta, opt-in and announced. The opt-in bucket is where people spend their reading time. The default and gated buckets are where billing, privacy and approval surprises usually hide. The August 19 harness changes sit mostly in opt-in, with one default worth knowing: cloud agents now subscribe to the pull requests they create and keep driving them, fixing CI and answering bot comments, without being asked again.

Because Cursor publishes three, against three different baselines, and the feature is a saving or a spend depending on which one describes your team. Measured against bundled Auto Cost pricing, Cursor states that Balance and Intelligence average about twice the cost, and up to two to four times depending on the mode selected. Measured against pinning every request to Opus 4.8, Cursor reports early-access customers paying 30 to 50% less on Auto-routed requests, with Balance about 36% lower.

The widely repeated 60% figure belongs to a third baseline again. It is measured against Fable 5, not Opus 4.8.

A team that has been living on Auto's bundled pricing should therefore read the router as an increase, and a team that pinned an expensive model for everything should read it as a decrease. Same feature, opposite direction, and what decides it is only what the last invoice was built on.

Intelligence is the mode where this gets misread most often. Cursor's own numbers put it at about 15% higher satisfaction than Opus 4.8 at close to the same cost, $6.76 against $7.34 per commit, so the saving on pinned Opus 4.8 traffic sits with Balance rather than with Intelligence. Budget a 60% cut on that migration and you will be wrong by a wide margin.

Never quote one of these figures without naming its baseline. It is the easiest mistake to make with this release, and it is the kind that gets caught in a procurement conversation rather than in an editing pass.

Separate the claims that expire from the ones that do not. A mechanic (what /in-cloud does, what a side chat inherits) holds until Cursor changes the feature. A number (review time, cost multiple, price per seat) expires quietly, and a status word like beta or waitlist expires the moment somebody ships. This tracker carries its own checked date rather than the site-wide one for that reason, since a release page's only real claim is currency.

Status words deserve the most suspicion, in both directions. Origin went from waitlist to early beta on August 17 with no standalone pricing, no self-hosting and the agent-native review features still marked as coming, so a slide that treats it as a finished GitHub replacement is a commitment nobody can keep, and one that still calls it a waitlist is a month stale. The larger model is training work. Cursor now has a paid-plan iPad experience for full pull-request review, while Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. is still an early beta with macOS, Windows and iPhone clients rather than Linux, Android or iPad apps.

The small releases are where the detail actually bites, and side chats are a good example. They arrived in 3.11, they go one level deep because a side chat cannot spawn another, and they are local-only today with Cloud AgentsAgents that run in a Cursor-managed virtual machine, check out the repo, do the work and open a pull request, then shut down, with no load on your laptop. Press Enter for the full definition. support flagged as coming soon. Any one of those can invalidate a workflow you were about to build on the feature.

So the habit is smaller than the page. Read the release note for the surface you own rather than a summary of all of them, then update the prompt, the policy and the pricing example together. A training deck that teaches the new mechanic on last quarter's pricing confuses a new hire more than one that is out of date all the way through.