xAI has rolled out Grok 4.3, pairing aggressive price cuts with a 1-million-token context window and a new Imagine Agent Mode for long-form creative work. The release tightens the gap with frontier rivals on real-world tasks while undercutting them by roughly an order of magnitude on cost — a deliberate squeeze on OpenAI and Anthropic's enterprise pricing.
Pricing built to undercut frontier rivals
Grok 4.3 ships at $1.25 per million input tokens and $2.50 per million output tokens, a roughly 40% drop on input and 60% drop on output versus Grok 4.20. According to Artificial Analysis, running its full Intelligence Index benchmark suite costs about $395 on Grok 4.3, compared with $3,357 on GPT-5.5 and $4,811 on Claude Opus 4.7. According to Artificial Analysis, the model runs at roughly 195 tokens per second with a December 2025 knowledge cutoff, and reasoning is on by default — billed at the same rate as output tokens.
The model is available through xAI's API, OpenRouter, and inside the Hermes agent. Developers also gain a 1-million-token context window, putting Grok in the same memory tier as Gemini 3.1 and GPT-5.5 for long-document and multi-file workflows.
Benchmarks: ahead of Gemini 3.1, behind GPT-5.5
On the Artificial Analysis Intelligence Index v4.0, Grok 4.3 lands at a composite score of 53, edging Muse Spark and Claude Sonnet 4.6. xAI highlights agentic improvements: τ²-Bench Telecom at 98%, IFBench at 81%, and an 8-point lift on AA-Omniscience accuracy.
The sharpest jump is on GDPval-AA, the benchmark that scores models on real-world knowledge-work tasks. Grok 4.3's Elo climbed 321 points to 1,500, overtaking Google's Gemini 3.1 — though OpenAI's GPT-5.5 still leads by 276 Elo. That places Grok 4.3 closer to the frontier on agentic work than any prior xAI release while keeping costs an order of magnitude below the leaders.
Imagine Agent Mode targets longer creative projects
Alongside the model, xAI launched a beta of Grok Imagine Agent Mode in the web interface. Rather than firing off single prompts, the agent plans, generates, edits, and revises content inside an open workspace. xAI's example use cases include one-minute movies, manga sets, and product stories. The mode is accessible from the input field on the Grok web app and requires a paid account.
Grok 4.3 is positioned as a developer-and-business model, with built-in tools for autonomous web search, X search, Python execution, file search (RAG), and generation of Excel files, PDFs, and PowerPoint decks — moves that bring it closer in capability to OpenAI's Codex and Anthropic's Claude Code workflows.
Implications: a price war, not just a benchmark race
The release reframes the frontier-model conversation around cost-per-intelligence rather than raw leaderboard position. With Grok 4.3 sitting on what Artificial Analysis calls the Pareto frontier — the line where performance and price line up best — xAI is pressuring OpenAI and Anthropic to justify premium pricing on workloads where Grok's 1M context, default reasoning, and tool use are good enough.
For enterprises evaluating multi-model deployments, Grok 4.3 strengthens the case for routing high-volume agentic tasks to cheaper tiers while reserving GPT-5.5 or Claude Opus 4.7 for the hardest reasoning. Combined with the broader May 2026 shift toward governance, compute access, and trust, the launch reinforces that the next phase of AI competition will be won — or lost — on unit economics as much as capability.
