xAI Drops Grok 4.3: The $395 Benchmark Run vs. GPT-5.5's $3,959

Running the full Artificial Analysis Intelligence Index benchmark on Grok 4.3 costs $395. The same benchmark run on GPT-5.5 costs $3,959 — exactly ten times more. The performance gap between them: Grok 4.3 scores 53 on the Intelligence Index; GPT-5.5 scores 60.

Context

xAI’s Grok 4.3 cuts input token costs approximately 40% and output costs approximately 60% relative to Grok 4.20. At $1.25 per million input tokens and $2.50 per million output tokens — with a one-million-token context window and built-in reasoning enabled by default — Grok 4.3 positions itself as the developer-grade agentic model for price-sensitive production workloads, not a frontier model competing at the top of capability benchmarks.

The timing follows a pattern that is reshaping how enterprise teams make model selection decisions. As GPT-5.5, Claude Opus 4.7, and other frontier models consolidate at the high-capability, high-cost tier, mid-tier models with strong price-performance ratios are capturing developer adoption for the majority of production use cases that do not require frontier reasoning — structured data extraction, tool-augmented retrieval, document generation, multi-step workflow execution.

The Detail

Grok 4.3 processes at 100 tokens per second and provides native tool use across web search, X search, Python code execution, and file-based retrieval via RAG. It can generate structured outputs including Excel files, PDFs, and PowerPoint documents, and is available through OpenRouter, the xAI API, and the Hermes agent platform.

On the Artificial Analysis Intelligence Index — a composite benchmark measuring coding, instruction following, reasoning, and knowledge tasks — Grok 4.3 scores 53, ahead of Claude Sonnet 4.6 and behind Claude Opus 4.7 (57) and GPT-5.5 (60). On GDPval-AA, which measures real-world knowledge work performance, Grok 4.3 reaches an Elo score of 1,500 versus GPT-5.5’s 1,776. The Claude Opus 4.7 benchmark run costs $4,811 — making Grok 4.3 approximately twelve times cheaper per equivalent benchmark run.

The cost comparison becomes operationally significant at scale. A production agentic application running 100,000 complex queries per month against GPT-5.5 at roughly $0.40 per query costs approximately $40,000 monthly. The same workload against Grok 4.3 at a tenth of the benchmark run cost — adjusting for the actual token consumption mix — shifts that number substantially. For workloads where the performance gap at the 53–60 Intelligence Index range does not affect outcome quality, the pricing differential is not a trade-off; it is a straightforward engineering decision.

The headline addition to the model release is Grok Imagine Agent Mode (beta), available to paid accounts via the Grok web interface. Imagine is a persistent creative workspace where the model autonomously plans, generates, edits, and revises content across extended multi-step sessions — designed for use cases including short film production, manga series development, and product narrative creation. The session model is what distinguishes Imagine from existing text-to-image or text-to-video tools: the agent maintains project context across an extended workflow rather than operating on a single prompt-response cycle.

This is the same architectural pattern emerging across Runway, Stability AI, and other AI creative tools — but bundled directly into the xAI model API rather than requiring a separate product integration. xAI is positioning Grok 4.3 as both a developer infrastructure component and a creative production tool from the same base model, a product strategy that differs from Anthropic’s explicit separation of Claude’s enterprise and creative use cases.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.