Grok 4.3 - xAI's latest flagship reasoning model

What is Grok 4.3

Grok 4.3 is xAI's latest flagship reasoning model, focused on Agentic workflows and extreme cost-effectiveness. The model has approximately 500 billion parameters, and its API supports a 1 million token context window and native video understanding, and can generate structured documents such as PDFs and PPTs. In evaluations, its Intelligence Index reached 53 points, GDPval-AA surged by 321 points, and IFBench topped the entire leaderboard with 81%. API pricing is highly competitive, at $1.25 for input and $2.50 for output per million tokens, about 60% lower than the previous generation, with evaluation costs only 1/12 of Claude Opus 4.7.

Grok 4.3's main features

• Agentic reasoning engine: Reasoning mode is always on, designed specifically for multi-step Agent tasks, supporting complex workflow automation and long-document analysis, with no output token limit.
• Ultra-long context processing: The API side supports a 1 million token context window, while the consumer side retains 2 million tokens, allowing an entire book or a massive codebase to be processed in one go.
• Native multimodal understanding: Supports text, image, and video input, and can recognize objects in videos, describe event sequences, and answer questions about specific timestamps.
• Structured document generation: Can directly generate and download native PDF, PowerPoint, and Excel files without the need for external plugins or conversion tools.
• Built-in code execution environment: Supports writing and running code and directly producing downloadable files, suitable for development scenarios such as data analysis and script automation.
• Web and X platform search: Built-in real-time web search capability, can access X platform data, ensuring answers are based on the latest information.
• Voice cloning and synthesis: The Custom Voices API is also launched, supporting voice cloning, speech recognition (STT), and speech synthesis (TTS) in 25+ languages.
• Desktop automation Agent (Grok Computer): Can execute multi-step task chains across applications, enabling desktop-level automated operations.
• Developer toolchain: Supports function calling, structured output, Prompt caching, MCP remote tool access, and is compatible with third-party gateways such as OpenRouter.

How to use Grok 4.3

• Consumer web/APP use: After subscribing to SuperGrok Heavy ($300/month), in the Grok chat interface, click the model switch menu and select "Grok 4.3 (beta)" to start a conversation, supporting text, image uploads, and real-time web search.
• Official API access: Go to the xAI developer platform, register an account and obtain an API Key, call the model name grok-4.3 or grok-4.3-latest, send requests in the standard OpenAI-compatible format, supporting the us-east-1 and eu-west-1 regions.
• Third-party gateway calls: Access through platforms such as OpenRouter, Vercel AI Gateway, or Puter.js, without needing to directly manage the official xAI interface, suitable for developers who already have third-party AI gateway integrations.
• Long document analysis: Using the 1 million token (API) or 2 million token (consumer side) context window, directly upload an entire PDF, codebase, or long text, and let the model perform full-text summarization, extract key information, or engage in multiple rounds of in-depth follow-up questions.
• Video content understanding: Upload video files or provide video links, and Grok 4.3 can automatically recognize objects in the frames, describe event timelines, and answer questions about specific timestamps.
• Code Execution and File Generation: In a conversation, ask the model to write and run code, or specify the generation of structured documents such as PDF, PPT, Excel, etc., and the model will directly return downloadable files.
• Agent Automated Tasks: Through the Grok Computer feature or API function calling / MCP tool integration, set up multi-step task chains, allowing the model to perform automated operations across applications, such as data scraping, report generation, email processing, etc.
• Voice Cloning and Synthesis: Call the simultaneously launched Custom Voices API, upload audio samples to clone a voice, or use TTS/STT capabilities in 25+ languages for voice interaction.
• Prompt Caching to Reduce Costs: Enable the Prompt caching feature in API calls to cache repeatedly used long contexts. The cache rate is only $0.20 per million tokens, greatly reducing the cost of long-text processing.

Key Information and Usage Requirements for Grok 4.3

• Release Time and Version Status: The Beta version was launched first. The API is now fully open, while the consumer side is still in the early testing stage.
• Parameter Scale: About 500 billion parameters (0.5T). Musk confirmed that the 1T-parameter Grok 4.4 is in training and is expected to be released in early May.
• Context Window: The API side supports 1 million tokens, and the consumer side supports 2 million tokens; requests exceeding 200,000 total tokens trigger higher context pricing.
• Knowledge Cutoff: December 2025, with training data covering up to that point in time.
• Reasoning Mode Characteristics: Always-on, and cannot be manually turned off or adjusted. All requests undergo deep reasoning, making it suitable for complex tasks but with relatively slower response speed.
• API Pricing: Input $1.25 / million tokens, output $2.50 / million tokens, cache $0.20 / million tokens; the cost of running the full benchmark suite is only about $395.
• Rate and Regional Restrictions: 1,800 requests per minute, 10 million tokens; currently only two regions are open, us-east-1 and eu-west-1.
• Model Alias: When calling the API, the model name is grok-4.3 or grok-4.3-latest
• Consumer Subscription Threshold: Currently limited to SuperGrok Heavy ($300/month) paid members, and is expected to open to standard subscribers in mid-to-late May.
• API Access Channels: Requires an official xAI API Key, and can also be called through third-party platforms such as OpenRouter, Vercel AI Gateway, Puter.js, etc.
• Multimodal Input Requirements: Supports text, image, and video input, but the video understanding feature must be used through the API or the latest client.
• Output Capabilities: Supports text, code, and native PDF/PPT/Excel file generation, with no output token limit.
• Enterprise Feature Requirements: For advanced features such as function calling, structured output, Prompt caching, and MCP remote tools, access must be through the official API rather than only using the consumer side.

Core Advantages of Grok 4.3

• Extreme Cost-Effectiveness: API input $1.25, output $2.50 / million tokens, about 60% cheaper than the previous generation; running the full benchmark suite costs only $395, which is 1/12 of Claude Opus 4.7, and it was evaluated by Artificial Analysis as a model on the "Pareto frontier".
• Ultra-long context: The API supports 1 million tokens and the consumer side supports 2 million tokens, allowing it to ingest an entire book, a massive codebase, or long videos in one go, with no output token limit.
• Agentic reasoning capability: Reasoning mode is always on, GDPval-AA real-work evaluation surged 321 points to 1500, with a win rate of over 87% against the previous generation, and it excels at practical economic tasks such as marketing copy, financial models, and PPTs.
• Top-tier instruction following: IFBench ranked first on the entire leaderboard with 81% accuracy, and τ²-Bench Telecom customer service Agent tasks reached 98%, tied for first, showing extremely strong understanding and execution of complex instructions.
• Native multimodality and document generation: For the first time, it supports video input understanding and can directly generate native PDF, PPT, and Excel files without plugin conversion, closing the loop from "analysis to output".

Comparison of Grok 4.3 with similar competitors

Application scenarios of Grok 4.3

• Deep analysis of ultra-long documents: Based on the ultra-long context of 1 million (API) / 2 million (consumer) tokens, upload an entire book, legal contract, academic paper, or massive codebase at once to perform full-text summarization, key information extraction, cross-chapter correlation analysis, or multiple rounds of in-depth follow-up questions.
• Intelligent parsing of video content: After uploading a video file, the model can automatically identify objects and scenes in the footage, describe event sequences along a timeline, and accurately answer fine-grained timestamped questions such as "what appeared at the 3rd minute," suitable for security monitoring, course learning, and media material review.
• Agentic automation of business tasks: Based on the Agent capability that surged 321 points in the GDPval-AA evaluation, it can automatically complete real economic activity tasks such as writing marketing copy, building financial models, creating business PPTs, and generating competitive research reports, chaining multi-step reasoning to directly produce usable results.
• Code development and environment execution: Through the built-in code execution environment, the model can write and directly run code such as Python/SQL, output data analysis results, visualization charts, or downloadable script files, suitable for rapid prototype validation, data cleaning, and automated script development.
• Native structured document generation: Directly generate downloadable PDF, PowerPoint, and Excel files without relying on external plugins, suitable for one-click generation of formal documents such as report materials, data reports, contract templates, and training manuals.
• Intelligent Customer Service and Tool Invocation: Can be deployed as a highly complex customer service Agent, combining function invocation and MCP tool integration to automatically query databases, call APIs, process work orders, and provide precise responses.