Top AI Companies By Intelligence

Anthropic keeps a lean lineup — 6 models in this dataset, less than half of OpenAI's 13 — and Claude Opus 4.5 is the one that carries the company to #2, scoring 49 in FactsFigs' identical-prompt testing, just two points off the top. Anthropic's models have built their reputation on careful, well-reasoned answers and strong coding performance rather than the widest possible feature set, and Opus 4.5's score reflects that focus: consistent output quality across long, multi-step prompts, the kind of task where staying coherent over many turns matters more than raw speed.

Google tracks more models than any other creator in this dataset — 14 — yet its best score, 48 from Gemini 3 Pro Preview (high), still lands third, one point behind Anthropic and three behind OpenAI. That gap between portfolio size and peak score fits Google's strategy: Gemini ships across Search, Workspace, and Android as much as it competes as a standalone flagship, so the lineup spans from lightweight Flash variants built for speed up to the high-effort Pro configuration scored here. In FactsFigs' testing, Gemini's strongest showings came on multimodal prompts — reading images, tables, and mixed context — more than on pure text reasoning.

Z AI's GLM-4.7 is the highest-scoring model from a five-model lineup, landing at 42 — six points behind Google and fifteen off the top score. GLM models have built a name as capable open-weight options, particularly for bilingual Chinese-English tasks and lower-cost deployment compared with closed frontier labs, and GLM-4.7's score in FactsFigs' testing held up on general reasoning prompts even without the multi-tier model ladder that OpenAI, Anthropic, and Google maintain.

DeepSeek V3.2 scored 41 in FactsFigs' testing, tying xAI's Grok 4 for fifth place, out of an 8-model lineup — the third-widest in this ranking behind Google and OpenAI. DeepSeek built its reputation on open-weight models that approach frontier-lab reasoning at a fraction of the training and inference cost, and V3.2's score reflects that: it held its own on math and step-by-step reasoning prompts against labs running far larger infrastructure budgets.

Grok 4 ties DeepSeek V3.2 exactly at 41, drawn from a 7-model lineup. xAI's models lean on real-time access to X (formerly Twitter) data and a more conversational, less-filtered response style rather than chasing the widest benchmark suite, and that shows up in FactsFigs' testing as a relative strength on current-events and reasoning-under-ambiguity prompts — the kind of question where up-to-the-minute context matters more than a structured coding task.

Moonshot AI's Kimi fields the leanest lineup of the top seven — just 3 models — with Kimi K2 Thinking topping the group at 40. Kimi's models are built around agentic tool use: chaining searches, running code, and planning multi-step tasks inside a single response rather than answering in one shot, and that specialization is exactly where K2 Thinking held its ground against flagship models from labs with far larger model portfolios.

MiniMax fields just two models in this dataset, and MiniMax-M2.1 is the stronger of the pair, scoring 39 — tied with Xiaomi for the eighth and ninth spots. MiniMax has positioned its models as efficient agentic workers built for fast tool-calling and structured task execution rather than the deepest possible reasoning chains, trading some peak score for lower latency and cost per task.

Xiaomi's MiMo-V2-Flash ties MiniMax-M2.1 at 39, also from a two-model lineup. True to the "Flash" naming, Xiaomi's models are tuned for speed and on-device efficiency rather than maximum reasoning depth — a fit for Xiaomi's hardware business, where models need to run well on phones and lightweight edge deployments rather than top a leaderboard.

KwaiKAT closes the top ten with a single model in this dataset — KAT-Coder-Pro V1 — scoring 36. The name gives away the focus: it's built specifically for coding tasks rather than general-purpose use, and that narrow, single-model approach explains both its lower peak score against nine generalist rivals and its sharper edge on code completion and debugging prompts, where broad conversational ability matters less than getting the syntax right.

FactsFigs' intelligence score is not MMLU, not LMSYS Arena, and not any other public leaderboard — it's a private ranking built from identical-prompt testing across ten AI creators' flagship models, scored in-house. OpenAI's GPT-5.2 (xhigh) topped that testing at 51, with Anthropic's Claude Opus 4.5 (49) and Google's Gemini 3 Pro Preview (48) close enough behind that the order could shift with the next model release from either rival.

Model count tells a separate story from peak score. Google tracks the most models in this dataset (14) and OpenAI the second-most (13), yet neither leads on raw intelligence — KwaiKAT closes the ranking with a single model and still lands a respectable 36. A large model lineup signals product breadth, not necessarily a sharper flagship, and reading the two numbers together is closer to the real picture than either alone.