First independent results land for Inkling and Grok 4.5
LiveBench profiles Thinking Machines' debut model mid-pack at its highest effort setting, and Grok 4.5 enters DeepSWE effectively even with Claude Sonnet 5.
Event date Published
Inkling's first LiveBench profile: mid-pack overall, math-leaning
Inkling, Thinking Machines' open-weights reasoning model released July 15, posted its first LiveBench results, evaluated at its highest reasoning-effort configuration. It debuts twenty-first of 32 models overall at 71.7 on a board led by GPT-5.6 Sol at 82.4. Its highest score is Mathematics at 88.4, which still ranks eighteenth of 32 there; its best category rank is eleventh, on Instruction Following. Lower-effort configurations could score differently.
- LiveBench overall
- 71.7 #21 of 32 Board leader GPT-5.6 Sol at 82.4.
- LiveBench Mathematics
- 88.4 #18 of 32 Inkling's highest score, not a board standout.
- LiveBench Instruction Following
- 70.1 #11 of 32 Its best category rank.
- LiveBench Reasoning
- 78.4 #24 of 32 Logic and inference problems requiring multi-step deduction.
- LiveBench Coding
- 71.0 #25 of 32 Code generation and completion tasks with verifiable solutions.
- LiveBench Agentic Coding
- 47.7 #17 of 32 Coding tasks run through an agentic harness.
- LiveBench Data Analysis
- 72.8 #18 of 32 Table-join and format-conversion tasks, answered without writing code.
- LiveBench Language
- 73.5 #26 of 32 English reading comprehension of recent text.
Grok 4.5 enters DeepSWE even with Claude Sonnet 5
xAI's Grok 4.5 received its first DeepSWE result: 53.76% of tasks resolved, ninth of the 16 models on the board and effectively even with Claude Sonnet 5 at 53.85% — a gap under 0.1 points. The board leader, GPT-5.6 Sol, resolved 72.7% of tasks. This is one agent-run coding benchmark of 16 models, not a broad coding-quality ranking.
- Grok 4.5
- 53.76% #9 of 16 Effectively even with Claude Sonnet 5.
- Claude Sonnet 5
- 53.85% #8 of 16 The gap to Grok 4.5 is under 0.1 points.
- GPT-5.6 Sol
- 72.7% #1 of 16 Board leader.