Model components / Fine-tuned / chat models

Gemini Pro

Google DeepMind's flagship Gemini reasoning tier, natively multimodal with a 1M+ token context window and a thinking mode. The higher-capability counterpart to Gemini Flash. Tracked as a tier rather than a point release, with the measured release named in the score note. Generations to date are Gemini 2.5 Pro (2025), Gemini 3 Pro (Nov 2025) and Gemini 3.1 Pro (2026, in preview at the time of scoring). Closed weights, available through the Gemini API, AI Studio, Vertex AI and the Gemini apps.

Tier product. Scored against Gemini 3 Pro, the generally available Pro release. Gemini 3.1 Pro Preview leads LiveCodeBench Pro and posts 80.6% on SWE-bench Verified, but preview checkpoints are excluded from the map as separate products because they are not stable; it is noted here so the tier's coding standing is not lost. Supersedes the separate gemini-2-5-pro, gemini-3-pro and gemini-3-1-pro-preview entries. The capability axis cites only the DeepMind model index, which carries no benchmark figures, so the numbers in its note have no source behind them and it is flagged. Verified 2026-08-13 via the DeepMind model index.

Openness

1 high confidence

1.0

weights: closed
data: closed
code: closed
license: Proprietary(API-only)

Closed by construction and stable across generations. Google's own model index files Gemini under API and app access while placing Gemma separately under "Open models", so the line is drawn by the vendor rather than inferred. Measured against Gemini 3 Pro, the GA Pro release at the time of scoring.

https://deepmind.google/models/gemini/ recorded 2026-08-13
Google's model index lists Gemini under API and app access and files Gemma separately under a heading of "Open models", with no weight download for Gemini

Adoption

5 medium confidence

5.0

Distributed through the Gemini API, AI Studio, Vertex AI and the Gemini apps, whose combined reach is far above the >10M band floor. Attributed to the surfaces the Pro tier powers rather than a standalone per-model count, so directional.

https://deepmind.google/models/gemini/ recorded 2026-08-13
Pro tier offered across the Gemini app, Google AI Studio, the Gemini API and Vertex AI; no per-model usage or user figure published

Capability

5 medium confidence

5.0

Frontier tier. DeepMind's own Gemini Pro page carries a benchmark table in body text: Gemini 3.1 Pro (Thinking High) reads GPQA Diamond 94.3%, Humanity's Last Exam 44.4% with no tools, ARC-AGI-2 77.1%, Terminal-Bench 2.0 68.5%, LiveCodeBench Pro 2887 Elo and SWE-bench Verified 80.6% at a single attempt, against Sonnet 4.6, Opus 4.6 and GPT-5.2 in the same table. It tops that comparison set on Humanity's Last Exam, GPQA Diamond, ARC-AGI-2, Terminal-Bench and LiveCodeBench Pro, which is what the 5 rests on. Two limits on the reading: the 80.6% on SWE-bench Verified is second to Opus 4.6's 80.8% rather than leading, and no LMArena placement is claimed, because arena.ai's text board carries no Gemini Pro entry in its top ten. Confidence stays medium.

https://deepmind.google/models/gemini/pro/ recorded 2026-08-14
DeepMind's Gemini Pro page, Performance table in body text. Gemini 3.1 Pro (Thinking High) - Humanity's Last Exam 44.4% no tools, GPQA Diamond 94.3% no tools, ARC-AGI-2 77.1%, SWE-Bench Verified 80.6% single attempt, LiveCodeBench Pro 2887 Elo, Terminal-Bench 2.0 68.5%, MMMLU 92.6%. Comparison columns Sonnet 4.6, Opus 4.6, GPT-5.2 - Opus 4.6 leads SWE-Bench Verified at 80.8%.
https://deepmind.google/models/gemini/ recorded 2026-08-13
Pro presented as the flagship reasoning tier of the Gemini family; a product listing only, carrying no benchmark figures.

Verified 2026-08-13