xAI · 2026-03-18
Grok 4
256k context · reasoning · multimodal · tools · closed
Capability
87.3
Coding
81.2
Agents
67.5
Science
90.1
TTFT
410 ms
p50, US-East, streaming
Output speed
138 tok/s
p95 latency
9.10 s
Availability
99.94%
Cost/task
$0.074
Cost/correct task
$0.095
30-day regressions
2
Measured runs
18,420
Last tested 11 min ago
LIVE_MEASUREDEWR 0.925 · human_io 0
TTFT spark · primary endpoint
Category scores
- Math84.2
- Reasoning88.1
- Coding81.2
- Agents67.5
- Multimodal79.4
- Science90.1
- Long context86.0
- Private tasks72.8
Serving stacks
- xAI APILIVE_MEASURED410 ms · 138 tok/s
Reliability on xAI API: invalid JSON 1.8% · broken tools 0.7% · replay 91.2%
| Benchmark | Score | $/task | Median s | Evidence |
|---|---|---|---|---|
| GPQA Diamond public · n=198 | 84.8 | $0.038 | 10.4 | LIVE_MEASURED |
| SWE-bench Verified public · n=500 | 72.4 | $0.048 | 10.0 | LIVE_MEASURED |
| Terminal Bench public · n=80 | 48.8 | $0.048 | 30.0 | LIVE_MEASURED |
| OSWorld public · n=369 | 44.2 | $0.048 | 29.6 | LIVE_MEASURED |
| FrontierMath public · n=300 | 28.4 | $0.038 | 21.1 | LIVE_MEASURED |
| LiveCodeBench public · n=612 | 81.6 | $0.048 | 10.3 | LIVE_MEASURED |
| MMMU public · n=1730 | 76.2 | $0.038 | 10.2 | LIVE_MEASURED |
| LongContext-128k public · n=120 | 86.0 | $0.038 | 10.4 | LIVE_MEASURED |
| GBU Private Rotating — Sep 2026 private · n=240 | 72.8 | $0.038 | 10.1 | LIVE_MEASURED |
| Agent Episodes public · n=96 | 67.5 | $0.048 | 31.7 | LIVE_MEASURED |
| Structured JSON public · n=400 | 91.4 | $0.038 | 10.6 | LIVE_MEASURED |
| CAD Science Lab public · n=64 | 71.2 | $0.038 | 10.0 | LIVE_MEASURED |
Regressions
- NOISEWithin daily variance band. No adapter or tokenizer shift.