Grok 4.3 Pricing: $1.25 Per Million Tokens Makes xAI's New Model Hard to Ignore

xAI released Grok 4.3 in June 2026 with a clear price signal: $1.25 per million input tokens and $2.50 per million output tokens, each 37.5% and 58.3% cheaper than Grok 4.20 respectively. On paper, Grok 4.3 pricing now undercuts GPT-4.1 by around 37% on input and by a wider margin on output. But the cost story comes paired with a performance claim that is harder to ignore: according to independent testing by Artificial Analysis, Grok 4.3 achieves the highest non-hallucination rate of any frontier model tested — with some caveats worth understanding before you build on it.

What Grok 4.3 API pricing looks like in practice

The headline numbers are $1.25 per million input tokens and $2.50 per million output tokens. Prompt caching — useful for applications that send the same large system prompt or retrieval context on every request — is available at $0.20 per million cached tokens. xAI also offers up to $175 per month in free API credits through its data-sharing program for developers who opt in, which lowers the effective cost further for smaller projects.

At scale, those numbers matter. One million output tokens is roughly 750,000 words of generated text. At $2.50 per million output tokens, a developer generating 100,000 words of output per day would spend about $3.33 daily. The same workload on GPT-4.1 — priced at approximately $8 per million output tokens — runs around $10.67 daily, more than three times higher.

How the benchmarks actually look

Grok 4.3 scored 78% on the Artificial Analysis Omniscience benchmark, which tests factual recall and hallucination resistance across a wide range of questions. That is the highest non-hallucination score of any frontier model tested as of its June 2026 release. On the broader Artificial Analysis Intelligence Index, Grok 4.3 scores 53.2 — outperforming approximately 96% of tracked models.

There is, however, a notable catch: the 78% non-hallucination rate is 8 percentage points lower than Grok 4.20's score on the same test. xAI made deliberate trade-offs to improve agentic performance and reduce inference cost, and some hallucination resistance appears to have been a casualty of that direction. A 78% score is still the current best among deployed models — but it slid down from a higher baseline rather than climbing from a lower one.

Agentic performance and what changed

The area where Grok 4.3 clearly moved forward is real-world agentic task performance. On the GDPval-AA benchmark, Grok 4.3 gained over 300 Elo points versus Grok 4.20 — a jump large enough to represent a qualitative difference, not a marginal one. The model supports configurable reasoning effort (none, low, medium, or high), letting developers trade inference speed against depth of reasoning at the prompt level.

Other capability additions: a 1-million-token context window, native video input alongside text and images, and availability on Amazon Bedrock for enterprise developers who prefer to manage access through their existing AWS accounts rather than a direct xAI billing relationship.

What this means for API developers

Grok 4.3 pricing pushes the competitive floor lower across the frontier tier. When a model scoring in the 96th percentile of the Artificial Analysis Intelligence Index costs $1.25 per million input tokens, it changes the calculation for developers who have been defaulting to GPT-4.1 or Claude Opus for routine tasks.

The practical implication is that mixing models by task type — routing cheaper queries to lower-cost models and reserving expensive models for genuinely complex reasoning — now has a stronger cost argument. That is the same premise behind bring-your-own-key tools like ByteChat, where each provider's API key connects directly with no markup layer, and per-model spend tracking shows you what each model actually costs per conversation.

One caveat worth noting: aggressive launch pricing in AI has a track record of revision. xAI has adjusted Grok API prices before, and Grok 4.3's current rates may not be its steady-state price six months from now.

Frequently asked questions

Is Grok 4.3 cheaper than GPT-4.1?

Yes, by a significant margin as of June 2026. Grok 4.3 costs $1.25 per million input tokens versus approximately $2 per million for GPT-4.1 — about 37.5% cheaper on input. On output, the gap is wider: $2.50 per million for Grok 4.3 versus approximately $8 per million for GPT-4.1. Both figures are subject to change as providers adjust pricing.

What is Grok 4.3's non-hallucination rate?

Grok 4.3 scored 78% on the Artificial Analysis Omniscience benchmark as of its June 2026 release, the highest of any frontier model tested on that benchmark. Note that this is 8 percentage points lower than Grok 4.20's peak score on the same test — the improvement was in agentic task performance and cost, not in factual accuracy.

Does Grok 4.3 support prompt caching?

Yes. Grok 4.3 supports prompt caching through the xAI API at $0.20 per million cached tokens — useful for applications that send large, repeated context (system prompts, retrieved documents) on every request.

Frontier model pricing has moved faster than most enterprise budget cycles anticipated — and Grok 4.3 is another data point in that compression.