Cursor Composer 2 vs Claude Code 2026: Which AI Coding Assistant Reigns Supreme?
📑 Table of Contents
🎯 Quick Verdict
The choice between Cursor Composer 2 vs Claude Code (powered by Opus 4.6) fundamentally reshapes the AI coding landscape in 2026, offering specialized performance at lower costs versus broad reasoning and platform independence.
In the rapidly evolving world of AI-powered software development, the launch of Cursor Composer 2 on March 19, 2026, has ignited a fierce competition against established players like Anthropic’s Claude Code, which leverages the formidable Claude Opus 4.6 model. This head-to-head battle between a code-only, IDE-native model and a generalist reasoning powerhouse fundamentally alters the calculus for developers seeking the best AI assistant for their workflows.
This comparison delves into the critical benchmarks, pricing structures, and unique feature sets of both AI development platforms. Drawing directly from published data, official pricing, and independent developer analyses as of March 2026, we aim to provide a comprehensive, data-driven verdict to help you navigate this crucial decision. Whether you prioritize cost-efficiency, specialized coding prowess, or broad-spectrum AI intelligence, this article will illuminate which tool is best suited for your engineering needs.
⚡ Cursor Composer 2 vs Claude Opus 4.6 — Feature Score Comparison (2026)
Overview
The year 2026 marks a significant inflection point for AI coding assistants, with Cursor Composer 2 stepping onto the scene as a formidable challenger to Anthropic’s flagship Claude Opus 4.6, which powers Claude Code. This section will introduce these two powerful AI development platforms, highlighting their core identities and strategic positioning in the market.
Our comparison is grounded in understanding their fundamental design philosophies. Composer 2 emerges as a highly specialized, cost-optimized solution, whereas Claude Opus 4.6 represents the pinnacle of generalist AI reasoning, extending its capabilities far beyond mere code generation. This dichotomy is central to determining which assistant aligns best with diverse developer needs and organizational objectives.
Cursor Composer 2
Launched on March 19, 2026, Cursor Composer 2 is a proprietary, code-only model developed by Anysphere, the company behind the Cursor IDE. Fine-tuned from the Chinese open-source model Kimi K2.5, Composer 2 is designed for agentic coding within Cursor, excelling at tool use, file edits, and terminal operations. It features a 200,000-token context window with a unique self-summarization mechanism for long-running tasks, aiming to provide superior performance for pure coding at a significantly reduced cost.
Claude Opus 4.6 (for Claude Code)
Claude Opus 4.6 is Anthropic’s top-tier general reasoning model, deployed via the Anthropic API, Claude.ai, and as the core intelligence behind Claude Code, Anthropic’s terminal-based agentic coding CLI. Unlike Composer 2, Opus 4.6 is not confined to coding; it excels in a broad range of intellectual tasks, including complex architectural discussions, technical writing, and analytical reasoning. It boasts an impressive 1 million token context window, making it suitable for tackling the largest and most intricate coding and knowledge-work challenges.
The strategic emergence of Composer 2 as Cursor’s in-house model directly challenges the dependency on third-party AI providers, offering developers a purpose-built coding assistant. This shift forces a re-evaluation of whether a dedicated IDE-native solution or a versatile, platform-agnostic generalist model provides the superior experience for serious software engineering in 2026.
Cursor Composer 2 vs Claude Code: Key Features
The design philosophies of Cursor Composer 2 and Claude Opus 4.6 lead to distinct feature sets, each optimized for different aspects of the developer workflow. Understanding these nuances is crucial for identifying the best fit for your projects.
Self-Summarization & Context Window: Specialized Efficiency vs. Unrestricted Scale
Composer 2 features an innovative self-summarization mechanism that allows it to maintain coherence across extremely long autonomous coding sessions. With a 200,000-token context window, the model periodically compresses its own context to approximately 1,000 tokens, enabling it to track decisions and progress without context degradation over hundreds of actions. This is particularly useful when tackling multi-file refactoring or complex debugging tasks that require deep historical awareness within the Cursor IDE. Conversely, Claude Opus 4.6 offers a massive 1 million token context window at standard pricing, five times that of Composer 2. This immense capacity allows developers to keep entire large codebases, extensive documentation, and detailed conversation histories in full fidelity within context simultaneously, without relying on compression. For tasks involving vast projects or intricate architectural analysis, Opus 4.6’s raw context window size offers unparalleled detail retention.
Multi-Agent Parallel Architecture vs. Platform-Agnostic Deployment
A standout capability of Cursor’s ecosystem is its multi-agent workspace, allowing developers to run up to eight concurrent agents in isolated git worktrees or remote machines, powered by Composer 2. This enables parallel implementation experiments, competitive problem-solving, or splitting large refactoring tasks into simultaneous workstreams. An individual developer can effectively manage multiple parallel feature developments, compressing days of sequential work into a single orchestrated session. This level of concurrent agent interaction is a Cursor-specific advantage. In stark contrast, Claude Opus 4.6’s strength lies in its platform-agnostic deployment. It’s accessible via Anthropic’s API, Claude.ai, Claude Code’s terminal CLI, and even as a selectable model within Cursor itself. This versatility means teams can standardize on Opus 4.6 for IDE assistance, automated CI/CD checks, and API-powered internal tools, eliminating vendor lock-in and allowing a single model investment to cover all development contexts. For more on terminal-based agentic coding, explore our Claude Code vs OpenCode comparison.
Cost-Effectiveness & Business Strategy vs. Broad Capability & Reasoning Depth
Composer 2’s pricing is a significant strategic advantage, coming in at $0.50 per million input tokens, an 86% reduction from its predecessor and one-tenth the price of Opus 4.6. This aggressive pricing allows Cursor to offer generous usage within its subscription tiers, making AI coding assistance more accessible and economically sustainable for large engineering teams. For enterprises with thousands of developers, this cost structure translates into substantial savings and predictable monthly spend. However, this cost efficiency comes with a trade-off: Composer 2 is a code-only model. Claude Opus 4.6, on the other hand, is Anthropic’s flagship general reasoning model. It seamlessly handles complex architectural discussions, technical documentation, code reviews with nuanced explanations, and debugging that requires understanding business logic beyond mere syntax. While its token costs are 10x higher at $5.00/M for input, its ability to tackle the full spectrum of knowledge work surrounding coding provides a genuine practical advantage for technical leads and those whose daily tasks extend beyond pure code generation. For a wider view of AI assistants, check our AI coding assistants guide.
Specialized Code Training vs. Generalist Depth & Trustworthiness
Composer 2 was trained exclusively on code data, a deliberate narrowing of scope that significantly enhances its performance and reduces costs for coding-specific tasks. This specialization is reflected in its superior Terminal-Bench 2.0 and CursorBench scores compared to Opus 4.6. The model is specifically optimized for solving coding problems and interacting with developer environments. For developers who spend the majority of their time writing, refactoring, and debugging code, this focused training leads to highly relevant and accurate outputs. Opus 4.6, by contrast, benefits from its generalist nature as a frontier reasoning model. Its broad training allows it to bring a deeper understanding of problem context, potential implications, and diverse problem-solving approaches that specialized models might miss. While it may not always beat Composer 2 on pure coding benchmarks, its ability to reason through ambiguous requirements, propose high-level architectural solutions, and generate human-like narrative explanations for complex changes makes it a more comprehensive partner for senior developers and architects. Further details on Anthropic’s offerings can be found in our Claude plans comparison guide.
Pricing Comparison
The pricing models for Cursor Composer 2 and Claude Opus 4.6 present one of the most significant differentiators in the AI coding tool market. Composer 2 offers a dramatic cost advantage at the API level, which influences its subscription tiers and enterprise scalability. This financial disparity is a core consideration for individual developers and engineering organizations alike.
At the most granular level, Composer 2’s API input tokens are priced at $0.50 per million, with output tokens at $2.50 per million. This is a substantial 10x cost advantage over Claude Opus 4.6, which charges $5.00 per million for input tokens and $25.00 per million for output tokens. Even Composer 2’s faster variant, priced at $1.50/M input, still undercuts Opus 4.6 by over 3x. This aggressive pricing by Cursor, a direct result of owning and optimizing their model (fine-tuned from Kimi K2.5), allows them immense flexibility in packaging their offerings.
For individual developers, both tools offer an identical entry-level subscription cost of $20/month. Cursor Pro includes generous usage of Composer 2 at this price. Similarly, Claude Pro offers access to Opus 4.6. However, at scale, the differences become stark. Heavy Cursor users might see costs climb to $40-$50/month under a credit-based system, which is still considerably lower than Claude Code’s heavy usage tier. Claude Code users requiring uninterrupted service for intensive tasks often need the Max 20x tier, priced at $200/month, to avoid usage interruptions.
| Plan | Cursor + Composer 2 | Claude Opus 4.6 (Claude Code) |
|---|---|---|
| Free Tier | Free trial usage (within Cursor Pro) | Free trial usage (within Claude Pro) |
| Paid From | $20/month (Cursor Pro) | $20/month (Claude Pro) |
| Best For | Daily IDE coding, cost-sensitive teams | Terminal-first agentic coding, mixed knowledge work |
The enterprise cost picture further amplifies this gap. Cursor, with 50,000 enterprise customers, leverages Composer 2’s low token costs to offer competitive business plans. An engineering director managing a 20-person team might find Cursor’s enterprise pricing for Composer 2 to be significantly more budget-friendly than provisioning Claude Code Premium seats at $150/month per developer. This translates into more included usage per subscription dollar and more predictable monthly expenses for large organizations.
While Opus 4.6’s higher cost reflects its broader general reasoning capabilities, the 10x price differential for raw token usage means that teams performing high-volume, code-centric agentic work will face much higher compute costs with Claude Code. This makes Composer 2 a compelling choice for organizations focused on cost optimization for pure coding assistance, enabling wider AI adoption across entire engineering teams without making the economics unsustainable. For development teams building their full AI stack, exploring AI productivity tools can help optimize non-coding workflows.
Best Use Cases
The distinct strengths of Cursor Composer 2 and Claude Opus 4.6 mean they excel in different scenarios. Matching the right tool to the right problem can significantly enhance developer productivity and project outcomes.
Use Case 1: Daily IDE Coding Work
Problem: A professional developer spends 6-8 hours daily within the Cursor IDE, focused on feature development, bug fixing, and routine code review. They require high-quality agentic assistance at a predictable, sustainable monthly cost to maximize efficiency.
Solution: Use Cursor Composer 2 as the default model within a Cursor Pro subscription ($20/month). Composer 2’s 200K context window, combined with its self-summarization for long sessions, and a Terminal-Bench 2.0 score of 61.7%, makes it the best native model for everyday Cursor workflows. Its 10x lower token cost compared to Opus 4.6 ensures that extensive daily usage remains economically viable within the included subscription usage, reducing the risk of overage charges.
Outcome: Tasks that previously consumed hours can now be completed in minutes, often under 30 seconds. The developer benefits from highly accurate code suggestions, automated refactoring, and seamless terminal interactions, significantly accelerating their development cycle without incurring prohibitive costs. The integration is smooth and optimized for the Cursor environment.
Use Case 2: Terminal-First Agentic Coding
Problem: A developer prefers working primarily in the terminal and needs to delegate complex, multi-hour autonomous coding sessions to an AI agent. They require the agent to handle repository-wide tasks, debug, run tests, and automatically commit changes to Git, allowing the developer to focus on other work.
Solution: Leverage Claude Code with Opus 4.6 through a Claude Pro or Max subscription. Claude Code’s terminal-based agentic loop, powered by Opus 4.6’s expansive 1 million token context window and strong SWE-bench Verified score, is specifically designed for autonomous, repository-level work outside an IDE. This environment allows Opus 4.6 to deeply understand the entire codebase and execute complex, multi-step changes.
Outcome: Developers can initiate comprehensive coding sessions directly from their terminal, confident that Opus 4.6’s broader reasoning capabilities will handle cases where pure coding intelligence needs to be augmented by understanding business logic or architectural implications. This enables the AI to perform significant refactoring or feature implementation autonomously, committing verified results back to the version control system.
Use Case 3: Enterprise Team Cost Optimization
Problem: An engineering director leads a 20-person development team and aims to implement AI coding assistance for all developers. The challenge is to do so without the high per-seat costs of premium alternatives like Claude Code Premium ($150/month per seat) making the annual budget unsustainable.
Solution: Implement a Cursor Business plan with Composer 2 as the default AI model. Cursor already serves 50,000 enterprise customers, including Salesforce, where over 90% of developers use Cursor. Composer 2’s 10x lower token cost compared to Opus 4.6 directly translates into more included usage per seat at a significantly lower overall per-developer cost. The 86% cost reduction from Composer 1.5 allows Cursor to offer more value within its subscription tiers.
Outcome: The entire engineering team can adopt AI coding assistance within a manageable budget. This widespread adoption leads to double-digit improvements in cycle time, PR velocity, and code quality, enhancing overall team productivity and business sustainability by reducing overage charges and ensuring predictable monthly spend at enterprise scale.
Use Case 4: Mixed Coding and Knowledge Work
Problem: A technical lead requires AI assistance that extends beyond basic code generation. Their daily tasks involve complex architectural discussions, drafting technical documentation, writing client-facing technical explanations, and performing analytical reasoning, alongside coding. They need a single AI that can handle this diverse workload.
Solution: Utilize Claude Opus 4.6 via Claude Pro or Max, accessing it through Claude Code for coding sessions and Claude.ai for all non-coding knowledge work. Opus 4.6’s broad general reasoning capability covers the full spectrum of a technical lead’s daily output. Its ability to understand complex narratives, synthesize information, and articulate nuanced explanations makes it invaluable for tasks where Composer 2’s code-exclusive scope would fall short.
Outcome: A single, powerful AI model streamlines all work contexts, eliminating the need to switch between different tools or models based on task type. Opus 4.6 provides coherent, high-quality assistance for both code-centric and conceptual tasks, proving essential for professionals whose AI needs extend significantly beyond pure software engineering.
Use Case 5: Parallel Multi-Agent Feature Development
Problem: A developer needs to implement a large, complex feature that can logically be broken down into several parallel workstreams, such as API layer development, frontend component creation, test suite generation, and technical documentation. They want to execute these workstreams simultaneously rather than sequentially to accelerate delivery.
Solution: Employ Cursor’s multi-agent workspace with Composer 2, running up to 8 concurrent agents in isolated git worktrees. Each agent can be assigned a specific workstream (e.g., one for API, one for frontend, one for tests). The developer then monitors the parallel execution, reviews the individual outputs, and merges the best results from each concurrent run.
Outcome: One developer can effectively manage five to eight parallel implementation experiments simultaneously. This drastically compresses the development timeline, transforming what would typically be sequential days or weeks of work into a single, orchestrated session. This parallel agent architecture is a unique, Cursor-native capability, providing a compelling advantage for complex feature work that allows for clear decomposition.
Pros and Cons
✅ Pros
- Cursor Composer 2: Scores 61.7% on Terminal-Bench 2.0, outperforming Opus 4.6 at 58.0%. This statistically meaningful lead on a benchmark directly measuring real-world agentic coding capability confirms its high performance for CLI-based development tasks, especially for developers deeply integrated into the Cursor ecosystem.
- Claude Opus 4.6: Features a massive 1 million token context window, five times larger than Composer 2’s. This allows it to hold entire large codebases and extensive conversation histories in full fidelity, eliminating the detail loss inherent in compression-based approaches and making it ideal for the most architecturally complex projects.
❌ Cons
- Cursor Composer 2: It is exclusively available within the Cursor IDE and cannot be accessed via an external API or used in other development environments. This “Cursor-only” limitation means teams using AI across diverse surfaces (CI/CD pipelines, custom internal tools) require separate model investments for non-IDE contexts, limiting its versatility.
- Claude Opus 4.6: At $5.00 per million input tokens, it is 10x more expensive than Composer 2 ($0.50/M) for input tokens. For teams performing high-volume agentic coding, this significant cost differential compounds quickly, potentially making it economically unfeasible for widespread adoption purely for standard coding tasks where Composer 2 performs comparably or better.
Final Verdict
The choice between Cursor Composer 2 vs Claude Code (powered by Opus 4.6) in 2026 isn’t a simple matter of one being unequivocally “better.” Instead, it represents a fundamental decision between two distinct AI development philosophies: highly specialized, cost-effective coding assistance versus broad, generalist AI reasoning with extensive platform flexibility. Both models excel in their chosen domains, making the “best” choice highly dependent on a developer’s specific workflow, team needs, and budget constraints.
For developers who are fully immersed in the Cursor IDE and whose primary need is pure coding execution, Cursor Composer 2 is the unequivocal choice. Its superior performance on coding-specific benchmarks like Terminal-Bench 2.0 (61.7% vs Opus 4.6’s 58.0%) at one-tenth the token cost makes it an economic and performance leader for code-centric tasks. The innovative self-summarization mechanism and unmatched parallel multi-agent architecture further solidify its position as a powerhouse for complex, long-running coding projects. While concerns about self-reported benchmarks and early reliability issues warrant cautious optimism, its trajectory of improvement and cost reduction is undeniable. For developers optimizing for pure coding output and cost within a single IDE, Composer 2 offers unparalleled value.
Conversely, for developers who operate across diverse platforms, require AI assistance beyond just coding, or have established workflows around Claude Code’s terminal-native agentic architecture, Claude Opus 4.6 remains the superior system-level choice. Its enormous 1 million token context window ensures full fidelity for the largest codebases and most extensive conversations. Its platform independence—available via API, Claude.ai, Claude Code CLI, and even within Cursor—provides unmatched versatility across the entire development lifecycle, from IDE to CI/CD pipelines. While its 10x higher token cost is a significant hurdle for raw coding volume, Opus 4.6’s generalist capability for architectural reasoning, technical documentation, and complex analytical tasks makes it indispensable for technical leads and those whose work spans the full spectrum of knowledge work surrounding software development. Its depth of reasoning for the hardest, most ambiguous engineering challenges is currently unmatched by any code-only model.
The pragmatic recommendation for most serious developers in March 2026, as echoed by NivaaLabs analysis, is to run both. A Cursor Pro subscription at $20/month grants access to Composer 2 as the default model, with Opus 4.6 also available within the same IDE. This hybrid approach allows developers to leverage Composer 2 for the 80% of daily work that is pure coding execution and switch seamlessly to Opus 4.6 for the 20% requiring broader reasoning, deep architectural insights, or maximal context window. This strategy combines the cost-efficiency and specialized performance of Composer 2 with the generalist intelligence and versatility of Opus 4.6, delivering the best of both worlds at a combined cost often lower than relying solely on Claude Code Max. For developers building out their comprehensive toolkit, reviewing other AI coding assistants can further refine their stack.
❓ Frequently Asked Questions
Does Cursor Composer 2 actually beat Claude Opus 4.6?
Yes, on Terminal-Bench 2.0 (61.7% vs 58.0%) and CursorBench (61.3% vs 58.2%), Composer 2 beats Opus 4.6 according to Cursor’s official data. However, these are largely self-reported scores, and independent third-party verification is pending. GPT-5.4 currently leads both on Terminal-Bench 2.0 at 75.1%.
What is Cursor Composer 2 based on?
Composer 2 is a fine-tuned variant of the Chinese open-source model Kimi K2.5. Cursor continued pre-training the base model using exclusively code data and optimized it with reinforcement learning for long-horizon agentic coding tasks specifically within the Cursor environment.
How much cheaper is Composer 2 than Claude Opus 4.6?
Composer 2 costs $0.50 per million input tokens and $2.50 per million output tokens. This is a 10x cost advantage compared to Opus 4.6, which is priced at $5.00/$25.00 per million tokens. Even Composer 2 Fast is over 3x cheaper.
Can I use Composer 2 outside of Cursor?
No, Composer 2 is exclusively available within the Cursor IDE. It is specifically tuned for Cursor’s agent workflow and tightly integrated with its tool stack. You must be an active Cursor subscriber running the Cursor IDE to use Composer 2.
Should I switch from Claude Code to Cursor because of Composer 2?
Not necessarily. If you use a terminal-first workflow with Claude Code’s CLI, Composer 2 is unavailable. However, if you’re evaluating new environments, a Cursor Pro subscription ($20/month) allows you to use both Composer 2 and Opus 4.6 within the same IDE, letting you leverage each where it excels.
Latest Articles
Browse our comprehensive AI tool reviews and productivity guides
Cursor Composer 2 vs Claude Code 2026: Which Wins?
Cursor Composer 2 vs Claude Code 2026: Which Wins?
Compare Cursor Composer 2 vs Claude Code (Opus 4.6) in 2026. Discover benchmarks, pricing, key features, and which AI coding tool wins for developers.
Cursor Composer 2 Review 2026: Is It the Best AI Coding Model?
Cursor Composer 2 Review 2026: Is It the Best AI Coding Model?
Explore Cursor Composer 2 in 2026: a deep dive into its unique features, pricing, and performance to assess if it's the leading AI coding model for developers.
Best AppSumo AI Deals in 2026: Lifetime Tools Worth Every Penny (Updated...
Best AppSumo AI Deals in 2026: Lifetime Tools Worth Every Penny (Updated Monthly)
AppSumo's AI category is exploding in 2026 — but not every lifetime deal is worth grabbing. We've curated the best AI tools available right now with real prices, honest verdicts, and exactly who each deal is for.
8 Real Ways Companies Are Using Synthesia in 2026 That Will Genuinely...
8 Real Ways Companies Are Using Synthesia in 2026 That Will Genuinely Surprise You
From a Cannes Lion-winning Messi campaign to banks running AI-powered angry customer simulations — here are 8 real Synthesia projects that go way beyond "training videos.
Grammarly Review 2026: Who Actually Needs It (And Who Doesn’t)
Grammarly Review 2026: Who Actually Needs It (And Who Doesn’t)
Grammarly is one of the most widely installed writing tools on the internet — but is it actually worth paying for in 2026? We break down exactly who gets real value from Grammarly Pro, who's fine staying free, and who should skip it entirely.
Canva Pro vs Free 2026: Is the Upgrade Actually Worth $15/Month?
Canva Pro vs Free 2026: Is the Upgrade Actually Worth $15/Month?
Canva's free plan is genuinely generous — but Pro unlocks tools that change what's possible. We break down every key difference by user type so you know exactly whether $15/month is worth it for you.
Writesonic vs Jasper 2026: Stop Guessing — This Comparison Ends the Debate
Writesonic vs Jasper 2026: Stop Guessing — This Comparison Ends the Debate
Writesonic and Jasper are both solid AI writing tools — but they're built for completely different workflows. Features, pricing, SEO, and GEO compared so you know exactly which one belongs in your stack.
Semrush Review 2026: Is It Worth It? Features, Pricing, Pros & Cons...
Semrush Review 2026: Is It Worth It? Features, Pricing, Pros & Cons Honestly Assessed
Semrush is still the most comprehensive SEO platform available in 2026 — but at $139.95/month for the entry plan, it is not right for everyone. This honest review covers every pricing plan, the new Semrush One AI Visibility feature that tracks your brand in ChatGPT and Perplexity, how it compares to Ahrefs and SE Ranking, and exactly who should and should not pay for it. Plus: a full breakdown of the $200 affiliate commission and 120-day cookie that makes it one of the best programs in digital marketing.
7 Best ChatGPT Alternatives in 2026: We Tested Every One (Free +...
7 Best ChatGPT Alternatives in 2026: We Tested Every One (Free + Paid Options)
ChatGPT is no longer the best AI at any single thing — and in March 2026, the alternatives have genuinely caught up. Whether you need better coding (Claude), smarter research with citations (Perplexity), a completely free option with no daily limits (DeepSeek), or a 2 million token context window at no cost (Google Gemini), there is a more specialised tool for every job. We tested all 7 — here is exactly which one is right for your workflow.
Cursor Composer 2 vs Claude Opus 4.6 vs Sonnet 4.6 in 2026:...
Cursor Composer 2 vs Claude Opus 4.6 vs Sonnet 4.6 in 2026: Which Model Should You Use for Agentic Coding?
Cursor Composer 2 vs Claude Opus 4.6 vs Sonnet 4.6 — The smartest dev tools just leveled up. See which AI model actually codes, plans, and ships like a teammate—not just a chatbot.
Cursor vs Windsurf vs Claude Code in 2026: Which AI Coding Tool...
Cursor vs Windsurf vs Claude Code in 2026: Which AI Coding Tool Should You Use?
Cursor vs Windsurf vs Claude Code is the defining AI coding tool comparison of 2026 — three tools built on fundamentally different philosophies, targeting overlapping developer audiences at nearly identical price points, but delivering very different day-to-day experiences
Claude Dispatch Review 2026: Anthropic’s Remote AI Agent — Setup, Use Cases,...
Claude Dispatch Review 2026: Anthropic’s Remote AI Agent — Setup, Use Cases, Limits & Is It Worth It?
Claude Dispatch launched March 17, 2026 — send tasks from your phone, your desktop executes them locally, you come back to finished work. Setup takes 2 minutes. Current reliability is ~50% on complex tasks. Here is everything you need to know before relying on it.