Cursor

Key Terms

• Grok 4.5 - Cursor's new flagship model, jointly trained by Cursor and SpaceXAI. A mixture-of-experts model trained on trillions of tokens of Cursor data plus broader STEM/knowledge-work content. Priced at $2/$6 per MTok (base) or $4/$18 (fast). Available across desktop, web, iOS, CLI, and SDK. Part of the first-party Cursor Models pool. Source: Cursor – Grok 4 5
• Composer 2.5 - Cursor's price-efficient coding model, launched May 18, built on Moonshot's Kimi K2.5 checkpoint with additional RL training. Remains offered alongside Grok 4.5 as a smaller model weight class. Part of the first-party Cursor Models pool. Source: Cursor – Grok 4 5 , Cursor – Composer 2 5
• Cursor Models pool - One of two usage pools on individual plans (renamed from "Auto + Composer pool"). Includes Grok 4.5 and Composer 2.5. Described as "significantly more included usage" than the Other Models pool. Source: Cursor – Pricing
• Other Models pool - The pool for third-party models, charged at each model's per-MTok API rate. Pro includes $20/mo, Pro+ $70/mo, Ultra $400/mo of Other Models usage. Start plan does not include this pool. Source: Cursor – Pricing
• Cursor Router - Cursor's intelligent model router, launched July 22. Replaces the old single "Auto" option with three modes: Intelligence (frontier quality), Balance (daily-driver quality), and Cost (fixed per-MTok pricing). On by default for Teams plans. Source: Cursor – Router
• Auto Cost - Router mode with fixed per-million-token pricing regardless of which model is used. Exempt from the Cursor Token Rate. This is the closest to the old "Auto" behavior. Source: Cursor – Pricing
• Auto Balance / Intelligence - Router modes that bill at the routed model's API rate, including the Cursor Token Rate for third-party models on Teams/Enterprise. Users can be charged for frontier models (e.g. GPT-5.6 Sol) without explicitly selecting them. Source: Cursor – Pricing
• Cursor Token Rate - A $0.25/MTok surcharge on Teams and Enterprise plans for third-party model requests. Applies when selecting a third-party model directly or when Auto Balance/Intelligence routes to one. Auto Cost and all first-party Cursor models (Grok 4.5, Composer 2.5) are exempt. Source: Cursor – Pricing
• CursorBench 3.2 - Cursor's internal eval suite for multi-file agent tasks from real Cursor sessions. Updated July 8 with instruction-following and advanced tool-use problems. Source: Cursor – Cursorbench
• Cursor Start - A new India-only plan at INR 649/mo (tax inclusive), billed in INR with UPI. Includes Grok 4.5 (fixed medium effort, non-fast) and Composer 2.5 (non-fast), cloud agents, iOS app, and plugins. No Other Models pool, no on-demand usage, no Bugbot, no Auto, no SDK. Source: Cursor – Cursor Start India
• Privacy Mode (Legacy) - The strict "do not store my code" setting. June's iOS app launch removed the ability to return to this mode once the softer current Privacy Mode is enabled. No July resolution. Source: News – Item

Latest Changes

Changes since the 2026-06 report.

June watch-item verification

• SpaceX acquisition close timing: still-pending. No close announcement was found in July. The definitive $60B all-stock agreement signed June 16 remains expected to close in Q3 2026. Source: TechCrunch – Spacex To Acquire Cursor For 60B In Stock Days After Blockbuster Ipo
• Composer 2.6 launch: failed (superseded). No Composer 2.6 shipped in July. Instead, Cursor launched Grok 4.5 as the new flagship model on July 8. The Grok 4.5 blog states "Composer 2.5 will remain offered, and we will release new models of this size going forward," confirming Composer is now the smaller model weight class, not the flagship. Source: Cursor – Grok 4 5
• Privacy Mode Legacy downgrade: still-pending. No resolution, fix, or documentation was found in July. The issue remains as reported in June. Source: News – Item
• New model additions: confirmed. Grok 4.5 launched July 8 (first-party). Claude Opus 5, GPT-5.6 Sol/Terra/Luna, and Kimi K3 are now in the Other Models pricing table. Source: Cursor – Grok 4 5 , Cursor – Pricing

New July items

• New first-party model: Grok 4.5 (July 8). Jointly trained by Cursor and SpaceXAI. Mixture-of-experts, trained on trillions of tokens of Cursor data plus broader STEM content. Described as "our most intelligent model and the first we've built for more than software engineering" (data science, finance, legal work). Base pricing: $2/$6 per MTok. Fast variant: $4/$18 per MTok. Available on all plans including Start. Double usage for the first week. Source: Cursor – Grok 4 5
• Training data contamination disclosed. Cursor disclosed in a footnote that "an earlier snapshot of the Cursor codebase was accidentally included in training" for Grok 4.5, giving it an advantage on CursorBench. "The exact impact is unclear. That data has been removed for future models." CursorBench results for Grok 4.5 carry an asterisk. Source: Cursor – Grok 4 5 , Cursor – Cursorbench
• Grok 4.5 Model Card published (July 14). Documents capability benchmarks and safety evaluations, including new safeguards for cybersecurity capabilities. Source: Cursor – Grok 4 5 Model Card
• Cursor Router launched (July 22). Replaces single "Auto" with three modes. Auto Intelligence aims to match Fable 5 quality at ~60% lower cost. Auto Balance aims to match Opus 4.8 quality at ~36% lower cost. Auto Cost uses fixed per-MTok pricing. Trained on 600k+ live requests. On by default for Teams plans. Available across desktop, web, iOS, CLI, and SDK. Source: Cursor – Router
• Router cost data published. In online A/B tests, Auto Intelligence cost $6.76/commit vs Fable 5 at $12.69/commit and Opus 4.8 at $7.34/commit. Auto Balance cost $4.63/commit. Three early-access enterprise accounts saved 30-50% vs routing everything to Opus 4.8. Source: Cursor – Router
• Cursor Start launched (July 28). India-only plan at INR 649/mo (tax inclusive). Cursor reports 3M+ developers in India (tripled in the past year), its third largest market globally and home to the most power users per developer. Includes Grok 4.5 (fixed medium effort, non-fast), Composer 2.5 (non-fast), cloud agents, iOS app, and plugins. Excludes Other Models pool, on-demand usage, Bugbot, Auto, Automations, and SDK. Source: Cursor – Cursor Start India
• iPad app launched (July 29). Extends the June iOS launch. Adds split-screen sidebar chats, full PR review surface (comments, checks, approvals), Apple Pencil markup on screenshots, and an Inbox for tracking agent progress. Bitbucket and Azure DevOps SCM support added to both iPhone and iPad. Source: Cursor – Ipad
• Side Chats and Conversation Search (v3.11, July 10). /side or /btw creates a side chat with main-chat context for research without interrupting the main agent. Conversation search across agent transcripts via Cmd+K (local search index scaling to thousands of conversations). Redesigned project and repo pickers with combined Remote Machines menu. Source: Cursor – Side Chat
• Slack improvements (July 17). Cursor in Slack now shares a plan before starting, supports multi-repo environments, and can read from and post to other Slack channels and threads. Source: Cursor – Slack Improvements
• New cloud agent hooks (v3.11). Hooks added for conversation-level events: beforeSubmitPrompt, afterAgentResponse, afterAgentThought, stop, subagentStart. Enables self-correcting loops. Source: Cursor – Side Chat
• Pricing language change on plans page. Pro+ is now described as "3x Pro limits on Agent" (was "more included usage" in June). Ultra is now "20x Pro limits on Agent" (was "maximum included usage"). Hobby now lists "Access to Composer" (was "Limited agent requests + limited tab completions"). Actual dollar prices are unchanged. Source: Cursor – Pricing
• Model table expanded. The Other Models pricing table now lists cache write prices for all models. New entries since June: Claude Opus 5 ($5/$25), Claude 4.6 Opus/Sonnet, Claude 4.7 Opus, GPT-5.6 Sol ($5/$30), GPT-5.6 Terra ($2/$12), GPT-5.6 Luna ($0.20/$1.20), Kimi K2.7 Code ($0.95/$4), Kimi K3 ($3/$15), Gemini 3.6 Flash ($1.50/$7.50). Removed: Grok 4.3, Grok 4.20, Grok Build 0.1 (replaced by first-party Grok 4.5). Source: Cursor – Pricing
• "Cursor 3.5" verification: not found. No product, version, or changelog entry named "Cursor 3.5" exists. The current editor version is 3.11 (July 10). This appears to be a search artifact or misreference. No action taken.
• US Congressional probe: no July update. The April probe into Anysphere's use of Chinese AI models had no public July resolution. Impact on model availability remains undisclosed.

Plans

Pricing notes:

• On-demand usage is billed in arrears at the same API rates once included usage is consumed. Requests are never downgraded in quality or speed. Source: Cursor – Pricing
• Teams plans add the $0.25/MTok Cursor Token Rate for third-party model requests (direct selection or Auto Balance/Intelligence routing). Auto Cost and first-party Cursor models (Grok 4.5, Composer 2.5) are exempt. Source: Cursor – Pricing
• Opting in to regional data residency incurs a 10% uplift on model pricing for eligible models. Source: Cursor – Pricing
• Cursor's own guidance: daily Tab users and limited Agent users "often stay within $20", daily Agent users typically cost $60 to $100/mo, and power users "often $200+/mo". Source: Cursor – Pricing
• Regional data residency is now documented as available with a 10% pricing uplift. Source: Cursor – Pricing

Source: Cursor – Pricing , Cursor – Pricing

API Pricing

Cursor meters subscription usage at each model's per-million-token rate. Grok 4.5 and Composer 2.5 draw from the first-party Cursor Models pool. All third-party models draw from the Other Models pool. All figures are per 1M tokens.

First-party Cursor Models (Cursor Models pool)

Source: Cursor – Grok 4 5 , Cursor – Pricing

Other Models (third-party)

Source: Cursor – Pricing , Cursor – Grok 4 5

Model Performance / Benchmarks

CursorBench 3.2 was published on July 8 with full numeric results (score, cost/task, tokens/task, steps/task) for 53 model configurations. This is the first time Cursor has published extractable numeric benchmark data for its proprietary models. Source: Cursor – Cursorbench

Grok 4.5 training data contamination. Cursor disclosed that "an earlier snapshot of the Cursor codebase was unintentionally included in training" for Grok 4.5. "The exact score impact is unclear." CursorBench tasks are derived from real Cursor sessions, meaning the model may have seen solutions or codebase patterns during training. All Grok 4.5 scores carry an asterisk on the CursorBench page. Cursor says the data has been removed for future models and they are working on a larger CursorBench update. Source: Cursor – Grok 4 5 , Cursor – Cursorbench

Router performance claims (from the blog, not independently verifiable):

• Auto Intelligence lands near Fable 5 on user satisfaction at ~60% lower cost, and lifts satisfaction ~15% over Opus 4.8 at nearly the same cost.
• Auto Balance lands above Opus 4.8 on satisfaction at ~36% lower cost.
• Cost per commit: Auto Intelligence $6.76, Auto Balance $4.63, GPT-5.6 Sol (matched cost, lower satisfaction), Opus 4.8 $7.34, Fable 5 $12.69.
• Early access: 3 enterprise accounts saved 30-50% vs Opus 4.8, with no quality decrease.

Source: Cursor – Router

The Grok 4.5 blog post also references SWE-Bench Pro and Terminal-Bench scores as chart images (not extractable numbers). The footnote states these are self-reported scores for third-party models, and SWE-Bench multilingual GPT-5.5 score is from Cursor's internal run. Source: Cursor – Grok 4 5

Latest News

Grok 4.5 Launch (July 8, 2026)

Grok 4.5 launched as Cursor's most intelligent model and the first jointly trained by Cursor and SpaceXAI. It is a mixture-of-experts model trained on trillions of tokens of Cursor data plus broader STEM content, positioned for long-running tasks across software engineering, data science, finance, and legal work. Base pricing: $2/$6 per MTok; fast: $4/$18. Plans include significant usage with double usage for the first week. New cybersecurity safeguards added. Source: Cursor – Grok 4 5

Grok 4.5 Model Card Published (July 14, 2026)

Cursor published the official model card documenting capability benchmarks and safety evaluations, available as a PDF. Source: Cursor – Grok 4 5 Model Card

Cursor Router Launch (July 22, 2026)

Auto mode was replaced by Cursor Router with three optimization modes (Intelligence, Balance, Cost). The router was trained on 600k+ live requests and evaluated via online A/B tests across millions of requests. It is on by default for Teams plans. Admins can restrict modes, set defaults, and allow/block underlying models. Available across desktop, web, iOS, CLI, and SDK. Source: Cursor – Router

Cursor Start for India (July 28, 2026)

A new plan for developers in India at INR 649/mo (tax inclusive), payable with UPI. Cursor reports 3M+ developers in India, its third largest market, with more agent requests per developer than anywhere else. Includes Grok 4.5 (fixed medium effort, non-fast), Composer 2.5 (non-fast), cloud agents, iOS app, and plugins. Excludes Other Models pool, on-demand usage, Bugbot, Auto, Automations, and SDK. Source: Cursor – Cursor Start India

Cursor for iPad (July 29, 2026)

iPad app launched on all paid plans with split-screen sidebar chats, full PR review surface (comments, checks, approvals), Apple Pencil markup, and an Inbox. Bitbucket and Azure DevOps SCM support added to both iPhone and iPad. Multi-PR sessions and team switching added. Source: Cursor – Ipad

Side Chats and Conversation Search (v3.11, July 10, 2026)

Side chats via /side or /btw enable research tangents without interrupting the main agent. Conversation search across agent transcripts via Cmd+K with a local search index. Redesigned project and repo pickers with combined Remote Machines menu. New cloud agent hooks for conversation-level events. Source: Cursor – Side Chat

Cursor in Slack Improvements (July 17, 2026)

Slack integration now shares a plan before starting, supports multi-repo environments, and can read from and post to other channels and threads mid-task. Source: Cursor – Slack Improvements

SpaceX acquisition: no July close

The definitive $60B all-stock acquisition agreement signed June 16 has not closed as of July 31. The deal remains expected to close in Q3 2026. Source: TechCrunch – Spacex To Acquire Cursor For 60B In Stock Days After Blockbuster Ipo

SpaceXAI training partnership confirmed via Grok 4.5

The April SpaceXAI partnership blog post (leverage Colossus infrastructure to scale up model training) produced its first public output with Grok 4.5. The model confirms that Cursor telemetry data flows into jointly trained models. Source: Cursor – Spacex Model Training , Cursor – Grok 4 5

US Congressional probe: no July update

The April US House probe into Anysphere's use of Chinese AI models had no public July resolution. Impact on Cursor's ability to use models like DeepSeek remains undisclosed.

Community Signals

Grok 4.5: Massive HN Engagement (776 points, 1,502 comments)

The HackerNews thread for the Grok 4.5 launch reached 776 points and 1,502 comments, making it the highest-engagement Cursor model launch in the project's tracking window. The discussion was dominated by the SpaceXAI training connection, the CursorBench data contamination disclosure, and comparative testing against Claude models.

• mholt (Caddy creator): "Of the 3 models I tried, Grok did the best at making an iOS app I wanted for personal use (a bike computer with specific qualities). (Claude just gave up and did an HTML/CSS implementation but I insisted on native SwiftUI+Metal.) Grok definitely fumbles sometimes, but I have been surprised what it CAN intuit versus me having to micromanage it." News – Item
• Schiendelman: "I agree. There's no chance Grok is better than Claude Code for this. And Claude is never so badly misaligned that it gives up and switches stacks." News – Item
• enraged_camel: "Same. A few months ago I pointed Opus 4.6 at a mid-size Vue app and told it to create the iOS equivalent using SwiftUI, and it nailed it. I broke the process down to phases and reviewed each phase, but within about ten days I had a functioning iOS app that had full feature parity." News – Item

Thread: News – Item (776 points, 1,502 comments). Source: Cursor – Grok 4 5

Cursor Router: Auto Mode Billing Surprise Warning

A "Tell HN" post on July 23 warned that Cursor's Auto mode change creates a billing risk for users who do not actively switch to Auto Cost.

• subhobroto: "Cursor's 'Auto', to me, used to mean 'cheap flat rate, don't worry which model ran'. That assumption is now a financial mishap waiting to happen as of today: 2026-07-22... Balance and Intelligence now bills at whatever model the router actually picks and draws from your API usage tier. So if you're still on 'Auto' and see 'GPT-5.6 Sol (Auto Balanced)' billed to your account under On-Demand, that's now expected." News – Item

Thread: News – Item (5 points). Source: Cursor – Router

Competing Routers Emerge: Tokenless (YC S26) Launches

The model router space is heating up. Tokenless (YC S26) launched on July 29 with 70 points and 60 comments, claiming to match Claude Fable 5 at half the cost. This follows the Weave Router (216 points in June). Both are external competitors to Cursor Router.

• rohaga: "We've been able to develop a version of the router that matches the performance of Claude Fable 5 at half the cost... Our approach queries multiple models at once and uses their progress to make decisions (this technique is novel AFAIK)... Switching models doesn't destroy the cache if the routing algorithm is aware of when the cache is hot/cold." News – Item

Thread: News – Item (70 points, 60 comments). Source: Usetokenless

CursorBench Contamination: Community Notes

The CursorBench footnote about Grok 4.5's training data contamination was noted in the HN thread.

• j-bu (separate HN post): "Cursorbench: Grok 4.5 better than GPT-5.5, at ~half the cost" News – Item

This post linked directly to cursor.com/cursorbench, but did not prominently surface the contamination caveat, which was only visible in the blog footnote and the asterisk on the leaderboard.

Failed sources

Reddit JSON endpoints (old.reddit.com/r/cursor) returned HTTP 403 for all attempts, consistent with the block seen in June. July Reddit sentiment could not be fetched programmatically. HN coverage above was used instead. The SpaceX acquisition close news search on HN returned zero results for July 2026, confirming no close announcement has been made.

Enterprise Readiness

Transparency Gaps