Grok 4.6 high benchmark results

[QingyunQian]

Summary

Adds the complete Grok 4.6 (high) ORBIT-Q TensorCircuit benchmark archive and publication-ready result figures.

• Solver: grok-4.6, reasoning effort high
• Auditor: gpt-5.6-sol, reasoning effort high
• Result: 9/12 passed — Tasks 02, 03, 04, 06, 07, 09, 10, 11, 12
• TensorCircuit: exactly 1.8.0.dev20260726
• Immutable image ID: sha256:b059c5fa7f75702f9afbf94ec7866e102ac32afd59d25634ec0aca0fd56e2833
• Resources: 6 CPU, 10240 MiB memory, 16384 MiB storage
• Frozen task aggregate SHA-256: 19fe27b83eaf668b3df32d1a68902b08cbe28585f189290769018eb16d927895
• Benchmark base commit: 0201238ec2983907e2891f5319f5fff2d00844d5

All 12 selected outcomes are first attempts under the repaired xAI compatibility layer. No integer-tool-schema or reasoning-replay compatibility failures recurred. Task 01 reached the 1,800-second agent limit; Tasks 05 and 08 passed functional checks but failed the independent source audit. Task 08 is also negative under the campaign-wide human adjudication.

Results

The archive includes selected solutions, rewards, solver trajectories, verifier/audit details, runtime and token accounting, configs, task/image provenance, PNG/PDF/SVG figures, and reproduction/verification tools.

Resource accounting

• Agent wall time: 199.17 min total
• Solving-side tokens: 29.767M total
• Reconstructed xAI list-price cost: $20.129532 total, or $2.236615 per valid solution
• Price basis: $2/M uncached input, $0.50/M cached input, and $6/M output

Harbor's raw cost_usd field is preserved as provider-reported zero/null; the separate reconstructed cost is derived from the archived token classes.

Validation

python3 results/grok-4.6-high/tools/verify_archive.py

Verified: 12 valid outcomes, fixed task hashes, Grok 4.6/high solver, GPT-5.6 Sol/high audit, pinned TensorCircuit 1.8 image, solution hashes, and valid figures.

[refraction-ray]

this looks much more expensive than Grok 4.5. no idea why

[QingyunQian]

this looks much more expensive than Grok 4.5. no idea why

The cost increase comes from two effects:

1. xAI prices cached input tokens higher for Grok 4.6: $0.50/M, compared with $0.30/M for Grok 4.5. Regular input and output prices remain unchanged.

2. Grok 4.6 generated a substantially longer and more tool-intensive agent trajectory:

Metric	Grok 4.5	Grok 4.6
Agent steps	343	581
Tool calls	301	626
Tool-output text	0.47M characters	1.36M characters
Mean final context	~38k tokens	~84k tokens
Cumulative input tokens	8.25M	29.28M

For example, on Task 01, both models ran for approximately 30 minutes, but Grok 4.5 used 47 steps and 1.52M cumulative input tokens, while Grok 4.6 used 93 steps and 6.58M input tokens. Grok 4.6 performed more exploration, testing, and tool calls; each subsequent call carried an increasingly long conversation and tool-output history, raising cumulative input by approximately 4.3× on that task.