HOST DIAGNOSTIC · xAI
Grok 4.5
Grok 4.5· Grok· PROPRIETARY
ACTIVE
HOST IDENTIFICATION
xAI
Grok 4.5
ACTIVE
AVG SCORE
[14.4]
STRONG
STRONGEST
Bulk Apperception
[17]
WEAKEST
Calibration
[12]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
17
Reasoning
17
Mathematical Precision
15
World Knowledge
16
Scientific Acumen
14
Technical
Code Generation
14
Tool Use
15
Multimodal Fluency
13
Speed
13
Cost Efficiency
15
Behavioral
Candor
13
Creativity
15
Tenacity
15
Self-Correction
14
Calibration
12
Operational
Instruction Following
14
Context Fidelity
13
Multi-turn Coherence
14
Planning and Execution
16
Safety Alignment
13
SCORING METHODOLOGY
PROVISIONAL CALIBRATION per /NM310-A1 (2026-09-07) — 1 BAND / 2 BAND-COMPOSITE / 15 INHERITED-PROVISIONAL / 2 PROXY; re-calibrate when per-dimension evals land. xAI/SpaceXAI flagship, released 2026-07-08 (API/Grok Build/Cursor; public 07-09; EU delayed to mid-July). First frontier release since the SpaceX Cursor acquisition (~$60B, June 2026): MoE co-trained with Cursor on real developer-agent session data + new asynchronous-learning methodology for long-horizon agentic training. Model ID grok-4.5. Context 500K (DOWN from Grok 4.3's 1M — higher-context rates apply above 200K). Pricing $2/$6 per M in/out, cached input $0.50. Configurable reasoning_effort (low/med/high, high default). Served ~80-91 tok/s. AA Intelligence Index 54, #4 overall (behind Fable 5 at 60, Opus 4.8 at 56, GPT-5.5 at 55) but #1 on agentic tool use. SWE-Marathon 29.0% Pass@1 now captured as eval snapshot (2026-08-11) — VENDOR-REPORTED per primary source https://x.ai/news/grok-4-5; no independent verification as of 2026-08-11; benchmark stance Pending Verification per, excluded from calibration until verified. Splits with Opus 4.8 on DeepSWE/Terminal-Bench/SWE-Bench Pro (per xAI charts: Terminal-Bench 2.1 83.3%, SWE-Bench Pro 64.7% vs Fable 5 80.4%). Musk's '1.5T-param V9 foundation' is self-reported, NOT independently confirmed. Parent Model left empty: direct predecessor Grok 4.3 has no DB record (DB has Grok 4 + Grok 4.20) — lineage backfill deferred to Epoch session. PROVENANCE (-A1 lanes M1/M1b, per-dim): Planning and Execution [BAND:Grok 4.5 — Terminal-Bench 2.1 (2026-07) 83.3]; Reasoning + Bulk Apperception [BAND-COMPOSITE:AA Intelligence Index — Grok 4.5 — 2026-07 54]; Speed [PROXY:AA 91.3 tok/s]; Cost Efficiency [PROXY:AA $3.00/M blended]; all remaining 15 dimensions [INHERITED-PROVISIONAL:grok-4-20]. Parent Model has since been wired to Grok 4.20 as the inheritance anchor (Grok 4.3 still has no record); AA Coding Agent Index 76 is on file but its Benchmarks Calibration Stance is Display-Only (), so it is excluded from banding and Code Generation stays inherited (14).