xai-direct · reasoning
XAI-DIRECT

grok-4.5-direct

Grok 4.5 direct OAuth (SuperGrok Heavy) — current flagship, 500k ctx

⚖ COMPARE MODELS
VALS INDEX
86.7% +/- 0.8
Accuracy
VALS INDEX
12.2s
Latency (median)
$VALS INDEX
$4.0035
Cost / Test (1P-est)
SCORE STATUSscored
TEST SETS13
RECORDED TRIALS390
CUSTOM PASS81%

Industry benchmarks

Accuracyattempts · responded
GPQA_DIAMOND GPQA Diamond
87.3%
55 attempts · 51 responded
HUMANEVAL_PLUS HumanEval+
not scored
80 attempts · 0 responded
IFEVAL IFEval
86.1%
55 attempts · 55 responded
MMLU_PRO MMLU-Pro
not scored
60 attempts · 0 responded
Industry benchmark Bars = accuracy %, black cap = that score

Custom 9-task suite

Mean /5n=trials
CUSTOM agentic_prompt
5.00
15 trials
CUSTOM code_debug
5.00
15 trials
CUSTOM json_strict
5.00
15 trials
CUSTOM reasoning_multistep
5.00
15 trials
CUSTOM code_gen_long
4.87
15 trials
CUSTOM summarize
4.87
15 trials
CUSTOM agentic_tool_use
4.60
15 trials
CUSTOM refactor_existing_code -
3.07
15 trials
CUSTOM creative_write
2.25
20 trials
Custom suite · scored 0–5 by consensus (m3 + Fable-5) ★ = task scored a 5 at least once · sorted by mean score

Cross-judge validation

m3 vs Fable-5 Δ

No judge bias detected on this model — m3 and Fable-5 agree across all tasks.

Per-task distribution

Score

All trials on this page were run through the harness. Industry suites use deterministic family-specific reference graders (no LLM judge). The custom suite is graded by consensus of minimax-m3 + claude-fable-5. See /industry/ for the full benchmark suite.