Cursor's Composer 2.5 Cuts Bugbot Review Times to 90 Seconds
Cursor has updated its Bugbot code review agent with the proprietary Composer 2.5 model, increasing speed by 3x and lowering average execution costs by 22%.
Cursor released a major architecture update to its Bugbot review agent that cuts average processing times from five minutes to 90 seconds. The performance gains rely on Cursor’s proprietary Composer 2.5 model, which replaces third-party frontier models to handle high-concurrency tasks directly. For developers using automated code review, this shifts the feedback loop from an asynchronous wait to a near-real-time step before opening a pull request.
Performance and Cost Metrics
The transition to Composer 2.5 changes the operational metrics across speed, accuracy, and price. Internal benchmarks for the June 2026 release show measurable improvements in baseline capabilities.
| Metric | Previous Baseline | June 2026 Update |
|---|---|---|
| Average Latency | ~5 minutes | ~90 seconds (3x faster) |
| Bugs Detected / Review | 0.56 | 0.62 (+10%) |
| Pricing Model | $40/month flat fee | $1.00 - $1.50 per PR |
The shift to usage-based billing lowers the average cost per run by 22%. The internal model stack allows Bugbot to process full repository context without hitting the rate limits or latency ceilings associated with external API calls to Claude or GPT.
Pre-Push and Differential Reviews
You can now run Bugbot locally before pushing commits. Using the /review command triggers a local analysis within the editor. The direct commands /review-bugbot and /review-security allow developers to target specific vulnerability classes without waiting for CI/CD pipelines to execute.
When you push the code and open a pull request on GitHub or GitLab, Bugbot performs differential review syncing. If the remote diff matches the local diff already processed, the agent skips redundant cloud execution. It leaves a comment referencing the local run rather than charging for a duplicate pass. Incremental feedback configurations also force the agent to review only newly added code in long-running pull requests, preventing repeated comments on ignored suggestions.
Enterprise Compliance and Custom Stores
To support organizational security requirements, the agent now respects model block lists. You can specify which AI models are permitted to process certain code segments, isolating sensitive proprietary logic from generalized cloud models.
Concurrent with the agent updates, Anysphere released updates for the TypeScript and Python Cursor SDK. Enterprises can configure custom tools and persistence stores, allowing teams to route review metadata directly into internal observability platforms. These updates are live via Cursor 3.7+ and the web interface at cursor.com/agents, with command-line support scheduled for a future release.
If your team utilizes multiple AI coding assistants, configure Bugbot’s incremental feedback rules immediately. Relying on local /review commands before pushing commits will maximize the new 90-second latency window and prevent unnecessary usage charges on duplicate diffs.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Scale Claude Code Across Enterprise Monorepos
Learn how to deploy Claude Code in multi-million line monorepos using hierarchical context, language server protocol integration, and on-demand skills.
Cursor Composer 2.5 Hits 79.8% on SWE-bench Multilingual
Cursor released Composer 2.5, an agentic coding model utilizing targeted reinforcement learning to match Claude Opus 4.7 performance on sustained tasks.
Claude Code Skills Cut Internal Token Usage by 47% at Anthropic
Anthropic’s internal deployment of structured skill directories in Claude Code reduces token overhead while formalizing hundreds of automated workflows.
Cursor Adds Multi-Repo Support to Cloud Agent Environments
Cursor's updated Cloud Agent Development Environments introduce multi-repo capabilities, layer caching, and scoped egress for autonomous coding tasks.
Cursor's Autoinstall Bootstraps RL Training Environments
Learn how Cursor uses previous model generations to automate reinforcement learning environment setups, mock dependencies, and verify target commands.