Grok 4.3 cuts input costs 37.5%

Grok pricing is one of those details that sounds boring until you remember how AI products actually get built. Every prompt costs money. Every retry costs money. Every agent loop that searches, plans, calls tools, and comes back with an answer costs a lot more. So when xAI pushed Grok 4.3 into its developer stack on April 30, the interesting part was not just “new model.” It was “new default model with meaningfully cheaper tokens.” ### What actually changed? xAI’s docs now steer developers toward Grok 4.3 as the general-purpose model to use, calling it the company’s most intelligent and fastest model. At the same time, independent benchmarking from Artificial Analysis says Grok 4.3 arrived with input token prices about 37.5% lower and output token prices about 58.3% lower than Grok 4.20. ### Why do token prices matter so much? Because token pricing is the meter running underneath the whole app. Input tokens cover the user prompt and conversation history. Output tokens cover the model’s answer. xAI also bills reasoning tokens and tool use separately in some workflows, which means a model that is cheaper at the token layer can materially change the economics of high-volume products. ### Is this just a price cut? Not really — the more important claim is price-per-intelligence. Artificial Analysis puts Grok 4.3 at 53 on its Intelligence Index, 4 points above the latest Grok 4.20 variant it compared against. The same analysis says Grok 4.3 costs about $395 to run its benchmark suite, roughly 20% less than Grok 4.20 0309 v2, even though Grok 4.3 used about 44% more output tokens in that evaluation. That means xAI did not just slash price tags. It improved capability while lowering the cost to get that capability. ### Where did the biggest performance jump show up? In agent-style tasks — the stuff companies increasingly care about. Artificial Analysis says the biggest jump was on GDPval-AA, where Grok 4.3 reached an Elo of 1500, up 321 points from 1179 for Grok 4.20 0309 v2. It also says Grok 4.3 hit 98% on τ²-Bench Telecom and held 81% on IFBench. Basically, the gains look strongest where the model has to follow instructions, work through multi-step tasks, and behave like a useful software worker instead of a chatbot with vibes. ### So is Grok 4.3 now the leader? Not across everything. Artificial Analysis says Grok 4.3 sits on the intelligence-versus-cost Pareto frontier, which is the nerdy way of saying it is hard to find another model that is both smarter and cheaper at the same time. But it also says Grok 4.3 still trails GPT-5.5 (xhigh) on GDPval-AA by 276 Elo points, with an expected win rate of only about 17% in that matchup. So this is not “xAI won.” It is “xAI got much more competitive on the cost curve.” ### Is there any catch? Yes — better scores did not come cleanly everywhere. Artificial Analysis says Grok 4.3 gained 8 points on its Omniscience Accuracy measure, but lost 8 points on the paired non-hallucination measure versus Grok 4.20. So the tradeoff may be familiar: more reach, more assertiveness, but not a simple across-the-board reduction in mistakes. ### Why does this matter for builders? Because cheaper strong models change behavior. Teams stop over-optimizing every prompt. They test more variants. They let agents take an extra step or two. They can afford more context, more retries, and broader rollout. When the default model gets cheaper without obviously stepping down in quality, the product roadmap loosens up. ### Bottom line? Grok 4.3 looks less like a flashy headline release and more like an economic one. xAI seems to be trying to make Grok the model developers reach for by default — not just because it is strong, but because the bill hurts less when you use it at scale.