Enterprise & RAG Leaderboard
The axis academic benchmarks miss: how dependable a model is in retrieval-augmented and agentic deployments. Four axes, all oriented higher = better — hallucination is shown inverted as Grounding so the table never reads backwards. They come from different evaluations and are not a combined score: read each column on its own. Only curated flagships are scored; — marks a genuine gap, not a zero.
| # | Model | Grounding ↑▼100 − hallucination | Tool-Use ↑↕τ-bench | τ³-Banking ↑↕policy KB + tool calls | Knowledge ↑↕AA-Omniscience, −100…100 |
|---|---|---|---|---|---|
| 1 | DeepSeek-V3.2CN | 93.7 | 71.3 | — | — |
| 2 | GPT-6 SolUS | 93.5 | — | — | 27.1 |
| 3 | GPT-5.4 ThinkingUS | 93.0 | — | — | — |
| 4 | GPT-5.4 ProUS | 91.7 | 80.1 | — | — |
| 5 | DeepSeek-V4-ProCN | 91.4 | — | — | — |
| 6 | GPT-6 AstraUS | 91.3 | — | 41.4 | 43.4 |
| 7 | GPT-5.5US | 90.7 | — | — | — |
| 8 | GLM-5CN | 89.9 | 82.1 | — | — |
| 9 | Gemini 3.1 ProUS | 89.6 | 76.5 | — | — |
| 10 | Qwen3.5CN | 89.3 | 77.5 | — | — |
| 11 | Claude Opus 4.5US | 89.1 | — | — | — |
| 12 | Claude Opus 4.7US | 88.0 | — | — | — |
| 13 | ChatGPT 5.1US | 87.9 | 74.2 | — | — |
| 14 | Claude Opus 4.6US | 87.8 | 84.8 | — | — |
| 15 | GPT-5.6 SolUS | 87.6 | — | 44.3 | — |
| 16 | Gemini 3US | 86.4 | 75.3 | — | — |
| 17 | Kimi K2.5CN | 85.8 | 74.2 | — | — |
| 18 | Mistral 3 FamilyEU | 85.5 | 70.2 | — | — |
| 19 | Grok 4.1US | 80.8 | 74.6 | — | — |
| 20 | GPT-5.3 CodexUS | — | 77.8 | — | — |
| 21 | DeepSeek-V4-Pro-0813CN | — | — | 39.6 | — |
| 22 | Qwen3.8-MaxCN | — | — | 51.3 | — |
| 23 | Qwen3.8-2.4T-A95BCN | — | — | 49.1 | — |
| 24 | Grok 4.6US | — | — | 50.7 | — |
| 25 | Qwen3.6 PlusCN | — | 76.8 | — | — |
| 26 | Claude Fable 5US | — | 89.2 | 38.1 | 43.3 |
| 27 | GLM-5.3CN | — | — | 50.3 | 14.3 |
| 28 | Kimi K3CN | — | — | 46.0 | 19.7 |
| 29 | Claude Opus 5US | — | — | 42.1 | — |
| 30 | Claude Fable 5.1US | — | — | 47.2 | 43.5 |
| 31 | Muse Spark 1.3US | — | — | 47.0 | 25.0 |
| 32 | Grok 4.7US | — | — | — | 32.0 |
| 33 | Claude Opus 5.5US | — | — | — | 46.4 |
33 flagships scored · coverage: Grounding 19/33 · Tool-Use 14/33 · τ³-Banking 12/33 · Knowledge 9/33. Click a column to sort.
Inverted Vectara HHEM hallucination rate (100 − rate). HHEM scores faithfulness when summarizing a given document — a proxy for RAG grounding, not full open-domain factuality.
τ-bench multi-turn task completion with policy adherence (retail/airline). Frozen: the authority stopped updating the task set. Kept because it is the only tool-use figure the 2025 and early-2026 models carry.
τ-bench’s successor family, banking domain, run by Artificial Analysis. Find the right policy in ~700 linked documents, then execute the tool calls it prescribes. The only axis here that scores 2026-H2 models — and far harder than the τ-bench column, so never read the two side by side.
AA-Omniscience Index, 6,000 factual questions across 42 topics. A wrong answer costs a point, saying “I don’t know” costs nothing, so the scale runs from −100 to 100 and 0 means as many right as wrong. It measures what the model knows without a document and whether it admits gaps — unlike Grounding, which checks fidelity to a text it was given.
- Function-Calling · Berkeley BFCL v4 (overall accuracy) — column not shown. Column is not rendered: ACTIVE_COLUMNS drops any metric with this status. BFCL has scored almost none of the frontier since 2025, and standalone call accuracy overlaps with the Tool-Use axis. Kept collectable so a revival needs no backfill — but it is not advertised anywhere the reader can see, and the page/SEO copy must not promise it.
- Tool-Use · τ-bench (tau1 retail+airline aggregate) — the authority has moved on to τ³-Banking, carried by the `tau3_banking` column. Versions score different task sets, so their numbers are not interchangeable: adopting the successor means a new column here, never an overwrite of these values.
Enterprise reliability is measured by a handful of small teams, and their leaderboards stop and restart. A blank recent row here means nobody independent has scored the model — which is itself worth knowing before shipping it into a RAG pipeline.
The axes come from different evaluations with different methodologies and snapshot dates — read them as directional signals per axis, not a single combined ranking. Hover any score for its source.
Origin badges (US · CN · EU · CA) make the geopolitics explicit: scored here are 21 US, 11 China, 1 EU, 0 Canada. The frontier is a US–China race at the top, while the enterprise/RAG axis is exactly where the EU/Canada players (Mistral, Cohere) compete — when the canonical boards list them.