Model Details
Grok 4 (March 2026)
Quick answer: Grok 4 (March 2026) is xAI's first-quarter 2026 frontier model, scoring 89.5% on ARC-AGI, 20.1% on Agent's Last Exam, and 8.0% on LHTB. Priced competitively at $1.25/$2.50 per 1M tokens with multimodal capability, it is xAI's most capable model tracked on Benchgen.
At a Glance
Where Grok 4 (20) leads
- 89.5% ARC-AGI — near-frontier abstract reasoning performance
- 20.1% Agent's Last Exam — competitive with leading 2025-era models
- $1.25/$2.50 per 1M tokens — aggressive pricing vs comparable frontier models
- Multimodal vision support
- December 2025 knowledge cutoff — highly current
Where it lags
- 8.0% LHTB — lower long-horizon task benchmark performance
- Agent's Last Exam 20.1% — below GPT-5.6 Sol (52.7%) and GPT-5.6 Terra (50.4%)
- Closed weights
Best for: Cost-efficient frontier reasoning with ARC-AGI-level capability; multimodal applications requiring a current knowledge cutoff at $1.25/$2.50 pricing.
What Grok 4 (March 2026) Is
Grok 4 is xAI's March 2026 flagship model — the fourth generation of the Grok series and xAI's most capable model to date. Its release coincides with the competitive frontier at the beginning of 2026, positioning it against GPT-5, Gemini 2.5 Pro, and Claude 4.
The model's ARC-AGI score of 89.5% places it in the upper tier of abstract visual reasoning capability — near GPT-5.6 Sol's 97.5% and above most models at this price point ($1.25/$2.50 per 1M). Its 20.1% Agent's Last Exam score indicates strong performance on complex multi-step agent tasks.
Grok 4's pricing is notably aggressive: $1.25/$2.50 per 1M tokens is lower than Gemini 2.5 Pro ($1.25/$10), o4-mini ($1.10/$4.40), and comparable to the most competitive frontier models. xAI appears to be pricing Grok 4 to gain market share against established providers.
Specifications
| Field | Value |
|---|---|
| Organization | xAI |
| Parameters | Undisclosed |
| License | Proprietary (API only) |
| Release date | March 9, 2026 |
| Knowledge cutoff | December 2025 |
| Modality | Text + Vision (multimodal) |
Pricing
| Input (per 1M tokens) | Output (per 1M tokens) | |
|---|---|---|
| xAI API | $1.25 | $2.50 |
Pricing per xAI pricing page. Output pricing ($2.50) is notably lower than comparable frontier models.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| ARC-AGI | 89.5% | Benchgen evaluation | 2026-07 |
| Agent's Last Exam | 20.1% | Benchgen evaluation | 2026-07 |
| LHTB | 8.0% | Benchgen evaluation | 2026-07 |
Grok 4 vs Alternatives
| Model | ARC-AGI | Agent's Last Exam | Price (in/out per 1M) |
|---|---|---|---|
| Grok 4 | 89.5% | 20.1% | $1.25 / $2.50 |
| GPT-5.6 Sol | 97.5% | 52.7% | $5 / $30 |
| GPT-5.6 Terra | 96.5% | 50.4% | $2.50 / $15 |
| o3 | 88.0% | — | $10 / $40 |
| Gemini 2.5 Pro | — | — | $1.25 / $10 |
Grok 4 vs o3: near-identical ARC-AGI (89.5% vs 88.0%) at $1.25 vs $10 input — a 8× cost advantage. GPT-5.6 Sol dominates on both benchmarks at $5/$30. Grok 4 is the strongest option in the $1-2 input tier on ARC-AGI.
Use Grok 4 via API
from openai import OpenAI
client = OpenAI(
api_key="YOUR_XAI_API_KEY",
base_url="https://api.x.ai/v1",
)
response = client.chat.completions.create(
model="grok-4",
messages=[{"role": "user", "content": "Solve this visual reasoning problem..."}],
)
print(response.choices[0].message.content)Related Models & Benchmarks
- Compare: GPT-5.6 Sol · o3 · Gemini 2.5 Pro
- Benchmarks: ARC-AGI · Agent's Last Exam · LHTB
Frequently Asked Questions
What is Grok 4?
Grok 4 is xAI's March 2026 frontier model, scoring 89.5% ARC-AGI and 20.1% Agent's Last Exam. Priced at $1.25/$2.50 per 1M tokens with multimodal support.How much does Grok 4 cost?
$1.25 per 1M input tokens and $2.50 per 1M output tokens — among the most competitive pricing at the frontier tier.How does Grok 4 compare to o3?
Grok 4 scores 89.5% vs o3's 88.0% on ARC-AGI, at approximately 8× lower input cost ($1.25 vs $10 per 1M). o3 has stronger coverage on other benchmarks (HLE: 20.32%) where Grok 4 has not been evaluated.What is Grok 4's knowledge cutoff?
Grok 4's training knowledge cutoff is December 2025 — among the most recent of any model tracked on Benchgen.Specs and scores from xAI's official Grok 4 announcement (March 9, 2026) and Benchgen evaluations. Pricing cited to the xAI API pricing page. Last updated 2026-07-24.