With tools variant. No-tools score is 56.8%. 11.6 pts ahead of Opus 4.6 with tools (53.1%).

Safeguard-affected (~Mythos 5). GPT-5.5 scores 41.4% on standard config.

With tools. Still HLE-with-tools SOTA, above GPT-5.5 Pro (57.2).

Highest self-reported HLE score. With tools. Leads GPT-5.2-Thinking by 28%. First Chinese model to top HLE.

Contemplating mode (multi-agent parallel reasoning). First model from Meta Superintelligence Labs. Accepts voice, text, image inputs. Strong multimodal + health, weaker on agentic tasks vs Mythos/GPT-5.4.

With tools, full set.

With tools. Second only to GPT-5.4 Pro (58.7).

With tools, full set. Up from GLM-5.1 (52.3).

HLE score with tool use (Deep Research Max). Uses extended search/reasoning time.

New open-weight SOTA on HLE with tools. Exceeds GPT-5.4 (52.1%) and Opus 4.6 (53.0%).

With tools, full set.

Score as cited in Kimi K3 blog (source: Anthropic official). Fable 5 hit fallbacks on 35% of tasks.

HLE score. Strong reasoning for a 21B-active MoE.

Hardest reasoning benchmark. With tool use. Corrected score per Feb 23 update.

With tools. Near GPT-5.4 Pro (58.7) and Opus 4.6 (54.7).

With tools, from Anthropic system card.

With tools from Anthropic system card.

Strong HLE score, competitive with frontier models on this extreme reasoning test.

Tool-augmented evaluation variant. Base (no tools) score is 30.5%. Higher than Claude Opus 4.5 tools score.

Leads non-Mythos models on HLE. Close to Fable 5 safeguarded score.

With tool access. Nearly 2x the CoT-only score. Impressive for a 27B model — beats several frontier models.

Without tools. Google specialized reasoning mode. Also scored 84.6% on ARC-AGI-2.

With tools, Max reasoning. Below leaders but competitive.

HLE with tools. Flash-tier model beating Step 3.5 (35.7) and DeepSeek V4 Flash (45.1).

Full set text+MM. Highest among non-Mythos models.

HLE with tools. Inkling-Small preview (276B MoE, 12B active) matches its larger sibling on reasoning.

Open-weights 975B MoE (41B active). First model from Thinking Machines Lab (Mira Murati). Multimodal from scratch.

Score as cited in Kimi K3 blog (source: OpenAI official).

Without tools. Google DeepMind official benchmarks page. Leads frontier models on HLE standard eval.

Top of Scale AI HLE leaderboard. Standard evaluation, no tools.

HLE-Full (no tools).

No tools. Still below Opus 4.6 (46.9) and Gemini 3.1 Pro (44.4).

Open-weight model with tools, competitive with GPT-5.1 High (42.7%).

No tools. Below GPT-5.5 Pro (43.1).

No tools. Below Opus 4.6 (46.9) and Gemini 3.1 Pro (44.4).

Ties GPT-5.5 (41.4). Text-only subset.

Text-only subset. Up from GLM-5.1 (31). 10-point jump.

Full set text+MM. Below Opus 4.7 (46.9). Up from Gemini 3 Flash (33.7).

No tools. Below GPT-5.4 (39.8%) and Gemini 3.1 Pro (44.4%).

Second-highest on HLE behind GPT-5.4 Pro (44.3%). Codex-native agent combining frontier coding with general reasoning.

No tools, from Anthropic system card. Between Claude Opus (40.0%) and Gemini (44.4%).

No tools, Max reasoning. Below GPT-5.5 (41.4) and Gemini 3.1 Pro (44.4).

Second on Scale AI HLE leaderboard. Standard evaluation, no tools.

Text-only. Below GLM-5.2 (40.5) and Qwen3.7-Max (41.4).

xhigh thinking config on Scale AI HLE leaderboard.

No tools, Max reasoning, 13B activated. Near Kimi K2.6 (34.7).

HLE without tools. Improved from K2.5 (30.1%) but below GPT-5.4 (39.8%) and Opus 4.6 (40.0%).

Frontier intelligence at Flash speed. Outperforms Gemini 2.5 Pro on HLE without tools.

Full set text+MM. Lowest among compared frontier models.

Standard evaluation on Scale AI HLE leaderboard.

No-tools. Slight improvement over GLM-5 (30.5%). 754B open-weight MoE.

Text-only subset. With tools: 50.4% (comparable to Kimi K2.5 51.8%).

HLE text-only. Inkling 975B MoE (41B active), Thinking Machines Lab first open-weights release.

HLE text-only variant.

Standard evaluation on Scale AI HLE leaderboard.

With search tools enabled. Without search: 19.5%.

Chinese frontier model. Strong HLE showing for open-weight model.

Chain-of-thought only. With tools jumps to 48.5%. Shows HLE requires tool use for small models.

Frontier reasoning benchmark. Third-party verified via Artificial Analysis.

Solid HLE for a small dense model.

HLE no tools. Low but non-zero for Flash-tier.

Open MoE with competitive HLE score. Beats Qwen 3.5 35B-A3B (22.4) is slightly lower.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

Without tools. Dense 31B model. Strong HLE for an open-weight model.

HLE without tools. Behind frontier models (Opus 4.6: 30.7, GPT-5.2: 31.4).

HLE no-tool score; lower than Qwen3.5-35B (22.4), showing frontier reasoning still difficult for 3B active params.

With search tools enabled. Without search: 8.7%.

Without tools. 30B-A3B MoE, 3.6B active params.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

HLE no tools. Low but non-zero for 12B.