xAI shipped Grok 4.3 this week, leaning hard into a price-cutting pitch as the model lands on OpenRouter and the company's own API. Input runs $1.25 per million tokens, output $2.50, with reasoning always on and billed at the standard output rate.
The intelligence story is more nuanced than the launch tweets suggest. Independent benchmarker Artificial Analysis puts Grok 4.3 at 53 on its Intelligence Index, four points above Grok 4.20 and just edging Claude Sonnet 4.6, but well behind GPT-5.5 at 60. Where it wins is dollars per point: a full Intelligence Index run cost $395, versus $3,959 for GPT-5.5 and $4,811 for Claude Opus 4.7.
Agentic gains look real but uneven. The model jumped 321 ELO on the GDPval-AA benchmark and tops Vals AI's CaseLaw v2 at 79.3% accuracy. Andon Labs, which runs models against an autonomous vending machine task, reported the opposite. Grok 4.3 "seems to have narcolepsy problems, preferring to sleep for multiple days in a row over taking actions," the lab wrote. Awkward read for a model xAI is positioning as agent-ready.
The other piece is Grok Imagine Agent Mode, now in beta on the Grok web client. It swaps the chat box for an infinite canvas with templates for short films, manga sets, UGC product stories, and brand identities. One sentence in, the agent plans, generates, and stitches images and video into a finished output. A paid account is required.
Context window sits at 1 million tokens. Knowledge cutoff is December 2025. Grok 5 is in training, per xAI.
Bottom Line
Grok 4.3 runs the Artificial Analysis Intelligence Index for $395, versus $4,811 for Claude Opus 4.7.
Quick Facts
- API pricing: $1.25 per 1M input tokens, $2.50 per 1M output
- Score: 53 on Artificial Analysis Intelligence Index (company-reported context: +4 vs Grok 4.20)
- Context window: 1 million tokens, reasoning always-on
- GDPval-AA: ELO 1500, up 321 points from Grok 4.20
- Available on OpenRouter and the xAI API; Imagine Agent Mode in beta on Grok web
