Intel / article

Xai Launches Grok 4 3 With Improved Agentic Performance And Lower Pricing

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Apr 30, 2026

Context

xAI's Grok 4.3 scores 53 on the Artificial Analysis Intelligence Index, 4 points above the latest version of Grok 4.20 and just ahead of Muse Spark and Claude Sonnet 4.6 (artificialanalysis.ai). It is also cheaper to evaluate, though not by as much as its price cut suggests. Input cost 37.5% less and output tokens 58.3% less than Grok 4.20 0309 v2, but Grok 4.3 writes about 44% more output tokens to complete the index. The net result is $395 to run the full suite, about 20% less than its predecessor, which Artificial Analysis says puts it on the Pareto frontier for intelligence versus cost.

The largest single improvement is on GDPval-AA, Artificial Analysis's test of real-world agentic tasks. Grok 4.3 reaches an Elo of 1500, up 321 points from 1179, passing Gemini 3.1 Pro Preview, Muse Spark, GPT-5.4 mini and Kimi K2.5. Elo ratings are relative, and a gap converts into an expected head-to-head win rate: Grok 4.3 still sits 276 points behind GPT-5.5, which Artificial Analysis translates into a win rate of about 17% against it. Grok 4.3 also gains 5 points on τ²-Bench Telecom, a customer support test, to reach 98%, and holds the 81% IFBench instruction-following score of Grok 4.20.

Not every measure moved the same way. On AA-Omniscience, Grok 4.3's accuracy rose 8 points, but its non- rate fell by 8 points, so Grok 4.20 0309 v2 still leads that measure. The newer model answers more questions correctly while doing worse than its predecessor at avoiding hallucinated answers.

Key takeaways · AI-distilled

Grok 4.3 costs $395 to run the Artificial Analysis Intelligence Index, about 20% less than Grok 4.20 0309 v2, even though it uses about 44% more output tokens; token prices 37.5% lower for input and 58.3% lower for output drive the saving.
On GDPval-AA, Grok 4.3 reaches an Elo of 1500 but still trails GPT-5.5 (xhigh) by 276 points, an expected win rate of about 17% under the standard Elo formula.
Grok 4.3 gains 5 points on τ²-Bench Telecom to 98% and holds 81% on IFBench, pointing to strength in instruction following and customer support agent tasks.
Its AA-Omniscience accuracy rose 8 points while its non-hallucination rate fell 8 points, so the earlier Grok 4.20 0309 v2 still leads Artificial Analysis's non-hallucination measure.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.

Read the source artificialanalysis.ai