Grok 4 Heavy
- Provider
- xAI
- Status
- Superseded
- Context
- 256,000 tok
Grok 4 Heavy is the multi-agent configuration of xAI’s Grok 4, launched alongside it on 9 July 2025. Where standard Grok 4 reasons with a single model, Grok 4 Heavy runs several agents in parallel — a “study group” that compares and cross-checks answers before converging — to push performance on the hardest problems. It was the headline feature of xAI’s premium SuperGrok Heavy tier at $300/month.
The extra compute showed up on the benchmarks: Grok 4 Heavy posted a perfect 100% on AIME 2025, 50.7% on Humanity’s Last Exam (with tools) — over double the best prior tool-free scores at the time — and 61.9% on USAMO 2025 math proofs (DataCamp, xAI). Like standard Grok 4, it has since been superseded by Grok 4.1, 4.2 and 4.3.
Quick specs
| Provider | xAI |
| Released | 9 July 2025 |
| Status | Superseded |
| Architecture | Multi-agent (parallel Grok 4 instances) |
| Access | Consumer only — SuperGrok Heavy ($300/month) |
| Context window | 256,000 tokens |
| AIME 2025 | 100% |
| Humanity’s Last Exam (with tools) | 50.7% |
| USAMO 2025 | 61.9% |
| Best for | The hardest maths and reasoning problems |
| Limitations | $300/month tier; higher latency/cost; superseded |
What Grok 4 Heavy is
Grok 4 Heavy is not a different base model from Grok 4 — it is a way of running it. Multiple Grok 4 agents work the same problem in parallel, share and critique each other’s reasoning, and converge on a best answer, in the spirit of a study group (xAI). That parallelism is what lifts its scores above standard Grok 4, at the cost of more compute, latency and price.
It was positioned as xAI’s top offering for power users and businesses, gated behind the $300/month SuperGrok Heavy subscription rather than sold per-token on the API.
Benchmark performance
xAI-reported launch figures (July 2025); treat vendor numbers as a ceiling.
| Benchmark | Grok 4 Heavy | Standard Grok 4 |
|---|---|---|
| AIME 2025 | 100% | ~91–93% |
| Humanity’s Last Exam (with tools) | 50.7% | 44.4% |
| USAMO 2025 (math proofs) | 61.9% | — |
| ARC-AGI-2 | ~16% | 15.9% |
The multi-agent approach delivered its biggest gains on competition maths and the hardest reasoning sets, where running several agents in parallel pays off (DataCamp). On everyday tasks the difference from standard Grok 4 was smaller — the value was the ceiling, not the average.
Access and pricing
Grok 4 Heavy was consumer-only, exclusive to the SuperGrok Heavy tier at $300/month ($3,000/year) in the Grok app. It was not offered as a standalone per-token API model. xAI carried a multi-agent “Heavy” approach forward into its later premium tiers.
How it compares
- vs standard Grok 4 — same base model, but parallel multi-agent reasoning for higher scores on the hardest problems, at higher latency and cost.
- vs rival “heavy”/Pro modes — comparable in spirit to OpenAI’s Pro (parallel test-time compute) and Google’s Gemini 3 Deep Think; Grok 4 Heavy’s standout was its 100% AIME 2025 and 50.7% HLE at launch.
Known limitations
Expensive and gated. At $300/month and consumer-only, it was out of reach for most users and not an API model. Higher latency. Running multiple agents in parallel is slower and costlier. Superseded. Later Grok models have moved the line on; for current use see Grok 4.3.
FAQ
What is Grok 4 Heavy?
Grok 4 Heavy is xAI’s multi-agent configuration of Grok 4 (July 2025): several agents reason in parallel and cross-check before answering. It scored higher than standard Grok 4 — including 100% on AIME 2025 — and was exclusive to the $300/month SuperGrok Heavy tier.
How is Grok 4 Heavy different from Grok 4?
Same base model, different execution. Grok 4 Heavy runs multiple agents in parallel for better results on the hardest problems, at higher latency and cost; standard Grok 4 uses a single reasoning pass.
How much did Grok 4 Heavy cost?
It was included with the SuperGrok Heavy tier at $300/month ($3,000/year). It was not sold per-token via the API.
Is Grok 4 Heavy still current?
No — it has been superseded by Grok 4.1, 4.2 and 4.3. For current xAI capability see Grok 4.3.
Last verified 18 June 2026. Grok 4 Heavy’s figures are xAI-reported from the July 2025 launch (via xAI and DataCamp). Confirm against xAI’s official pages before relying on specific numbers.