xAI drops Grok 4.3 with steep price cuts and an Imagine agent mode for creative projects
Key Points
- Elon Musk's xAI has released Grok 4.3, a developer-focused AI model capable of autonomously performing tasks like web searches, code execution, and document creation.
- While Grok 4.3 outperforms its predecessor in real-world knowledge work benchmarks, it still falls short of the leading models from OpenAI and Anthropic.
- xAI also launched a beta version of "Grok Imagine Agent Mode," aimed at enabling cohesive creative production workflows.
xAI released Grok 4.3, a cheaper and more capable model aimed at practical tasks. The release also includes a new agent-based image generation tool.
xAI developer Eric Jiang says Grok 4.3 is built for developers and businesses, with a focus on speed, low cost, and tool calls. The model can handle web search, X search, Python code execution, and file search (RAG) on its own, and it can generate Excel files, PDFs, and PowerPoint decks.
Grok 4.3 runs at 100 tokens per second and has a one-million-token context window. Pricing is $1.25 per million input tokens and $2.50 per million output tokens, with a knowledge cutoff of December 2025. Reasoning is now built in by default: Grok 4.3 "thinks" before answering every request, and reasoning tokens are billed at the same rate as regular output tokens. The model is available through OpenRouter, the xAI API, and the Hermes agent.
Much lower prices, mid-tier benchmark scores
According to independent benchmarking service Artificial Analysis, Grok 4.3 scored 53 on the Intelligence Index. That's slightly above Muse Spark and Claude Sonnet 4.6, four points above the previous Grok 4.20, and well behind the flagship models from OpenAI and Anthropic.
Input costs drop about 40 percent and output costs about 60 percent compared to Grok 4.20. That puts Grok 4.3 on what Artificial Analysis calls the Pareto frontier, where performance and cost line up best. A full benchmark run costs $395, compared to $3,959 for GPT-5.5 and $4,811 for Claude Opus 4.7.
On GDPval-AA, a benchmark that tries to measure AI performance on real-world knowledge work tasks, Grok 4.3's Elo score jumps 321 points to 1,500. That puts it ahead of Google's Gemini 3.1, though OpenAI's GPT-5.5 still leads by 276 Elo points.
Other benchmarks are less kind. Andon Labs, which has AI models run a snack vending machine, reports setbacks on autonomous agent tasks, noting that the model sometimes sits idle instead of taking action. "It seems to have narcolepsy problems, preferring to sleep for multiple days in a row over taking actions," Andon Labs writes. Val's AI ranks Grok 4.3 first on CaseLaw and first on CorpFin but found it struggles on general coding benchmarks and harder math problems, where it lands in 13th place.
New Agent Mode for Grok Imagine
xAI also released Agent Mode for Grok Imagine, now in beta on the Grok web interface. Rather than running off single prompts, the mode handles longer creative projects. An AI agent plans, generates, edits, and revises content in an open workspace, which xAI says works for things like a one-minute movie, a manga set, or product stories.
Agent Mode is accessible through the Grok web interface and can be turned on in the input field at the bottom left. A paid account is required.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now