by Google · Gemini 3 family · best for frontier reasoning + long-context on Google Cloud
Gemini 3.1 Pro is Google DeepMind's flagship reasoning model, launched 2026-02-19 in preview to validate the release before general availability. As of 2026-05-28 it remains the headline model in the Gemini app (Google AI Pro and Ultra) and the top reasoning option on the Gemini API and Vertex AI, even though its API/Vertex surface is still governed by Pre-GA Offerings Terms. It posts the highest public GPQA Diamond score of any proprietary model (94.3%, no tools), pairs that with a real 1M-token context (2M on Vertex for enterprise), and grounds answers in live Google Search. For a buyer: if you want frontier reasoning plus the deepest enterprise-cloud and live-data integration, this is Google's answer — accept that the API is technically still pre-GA.
| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 44.4% | deepmind.google 2026-02-19T00:00:00.000Z |
| MMLU | 92.6% | deepmind.google 2026-02-19T00:00:00.000Z |
| MMMU | 80.5% | deepmind.google 2026-02-19T00:00:00.000Z |
| TAU-bench | 90.8% | deepmind.google 2026-02-19T00:00:00.000Z |
| LMArena Elo | 1501 | facebook.com 2026 |
| GPQA Diamond | 94.3% | deepmind.google 2026-02-19T00:00:00.000Z |
| LiveCodeBench | 2887% | deepmind.google 2026-02-19T00:00:00.000Z |
| Terminal-Bench | 68.5% | deepmind.google 2026-02-19T00:00:00.000Z |
| MRCR Long Context | 84.9% | deepmind.google 2026-02-19T00:00:00.000Z |
| SWE-bench Verified | 80.6% | deepmind.google 2026-02-19T00:00:00.000Z |
| Artificial Analysis Index | 57 | artificialanalysis.ai 2026-05-28T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The strongest case yet to standardize on Google's AI stack — if we can live with a flagship that's still technically pre-GA.”
Gemini 3.1 Pro gives a CTO frontier reasoning plus the deepest enterprise integration on the market: Vertex VPC-SC, CMEK, residency controls, audit logging, and Workspace/BigQuery grounding no rival matches. The 1M-2M context removes RAG complexity for many workloads. Two strategic risks: it remains under Pre-GA terms (support and stability caveats), and adopting Vertex deepens Google Cloud lock-in. At $2/$12 with cache discounts and the 2M context, long-context TCO beats Claude Opus and GPT-5 Pro. Roadmap confidence is high given 3.5 Pro is already incoming.
“Google's wedge is live data plus long context — 3.1 Pro turns the Search and Workspace moat into a model-level advantage.”
In market terms, 3.1 Pro competes at the very top on reasoning (GPQA 94.3%) while owning a differentiation axis rivals can't copy: native Google Search grounding and Workspace/BigQuery integration. That positions it as the default for any org already inside Google's ecosystem and for use cases where freshness and citation matter. Its competitive moat is distribution (the Gemini app, Workspace, Android, Cloud) more than raw benchmark leads, several of which Opus 4.6 and GPT-5.3-Codex contest. Market timing is strong, but 3.5 Pro looming may stall procurement decisions.
“Cleanest frontier pricing in the category — until a prompt crosses 200K and the per-token rate quietly doubles.”
$2/$12 standard is competitive with GPT-5 and ~40% under Claude Opus 4.7 for long-context work. Explicit caching cuts input reads to $0.20, and batch halves rates. The catch is the two-tier model: above 200K tokens, input/output jump to $4/$18, and cached storage runs $4.50/1M tokens/hour — easy to under-model in RAG and streaming pipelines. Thinking tokens bill as output and can balloon spend on Deep Think. Vertex billing folds into existing Google Cloud invoices, simplifying procurement. Predictable once the tier boundary and thinking-token behavior are understood.
“Function calling and structured output just work, and 1M context lets me skip half my RAG plumbing.”
For a builder, the Gemini API and Vertex surface is the smoothest Google has shipped — clean SDKs (Python, TS, Go, Java, Dart), reliable function calling, response-schema structured output, built-in code execution, and Search grounding. The 1M-2M context collapses many agent loops into a single call. Genkit and Google ADK ease orchestration. Friction points: preview-tier RPD caps until spend gates clear, and high TTFT makes tight interactive loops painful. AI Studio's prompt history helps debug tool calls. Migration across Gemini tiers is a model-name swap.
“Brilliant on hard problems and live questions; just don't expect it to answer fast in Deep Think.”
In the Gemini app on AI Pro/Ultra, 3.1 Pro is genuinely excellent for research, planning, and analysis, and Search grounding makes it strong on current events where most rivals guess. Conversation quality is high and multimodal understanding (PDFs, images, video) is a daily advantage. Downsides users feel: Deep Think latency (10-30s), stricter refusals on edgy prompts than Claude or ChatGPT, and a tone that reads slightly clinical. The 2026 UX overhaul (native macOS app, cleaner mobile) fixed many prior complaints; Trustpilot remains mixed, mostly about caps and policy.
“A 'flagship' still under Pre-GA terms, with a 200K price cliff and TTFT north of 30 seconds — read the asterisks.”
The GPQA 94.3% headline is real, but several touted leads are contested: Opus 4.6 edges SWE-bench, GPT-5.3-Codex leads specialized coding, and 3.5 Flash beats 3.1 Pro on agentic benchmarks Google itself publishes. "Released" overstates it — this is preview validating before GA, with limited support. Long-context marketing glosses over MRCR recall decay near 1M and the doubled >200K pricing. AA hallucination data on the family is non-trivial, so Search grounding is doing real work to cover knowledge gaps. None of this makes it bad — it makes the "best at everything" framing marketing, not fact.
The full research notes behind this review — verified against primary sources.
Gemini is a sparse mixture-of-experts family, but Google does not publish parameter counts, expert counts, layer counts, or attention details for 3.1 Pro — all are honestly null. What is disclosed and verifiable: a native 1M-token context window (2M on Vertex), 65,536 max output tokens, a January 2025 knowledge cutoff, and native multimodal input across text, image, audio, and video in a single context. The model uses a thinking/Deep Think mechanism whose reasoning depth scales with difficulty and can be budgeted. Long-context retrieval is strong through the documented MRCR v2 128K average of 84.9%, though recall degrades as prompts approach the 1M ceiling.
Gemini 3.1 Pro is Google's strongest model on reasoning (cap_reasoning 9.5) and scientific knowledge — GPQA Diamond 94.3% is the highest proprietary score on record, and HLE 44.4% leads most of the field. Coding is now frontier-class (cap_coding 9.0): SWE-bench Verified 80.6% and LiveCodeBench Pro 2887 Elo. Math is excellent (cap_math 9.0) via Deep Think. Agentic/tool use (cap_agentic 8.5) is strong on tau2-bench (Retail 90.8%, Telecom 99.3%) and BrowseComp (85.9%) but trails Gemini 3.5 Flash on Terminal-Bench and MCP Atlas. Long context (cap_long_context 9.5) is a defining strength — true 1M/2M with usable retrieval. Vision and document/OCR are class-leading (cap_vision 9.5, cap_document_ocr 9.0) on MMMU-Pro 80.5% and mixed text-and-figure PDFs. Multilingual is broad (cap_multilingual 9.0, MMMLU 92.6%). The standout differentiator is real-time data (cap_realtime_data 9.5): Google Search grounding is the deepest live-retrieval integration of any frontier model. Creative writing (8.0) is good but slightly more clinical than Claude; instruction following (8.5) and function calling (9.0) are reliable; safety calibration (8.5) is strong but stricter on some prompts than OpenAI/Anthropic.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| GPQA Diamond | 94.3% | More than 2x ARC-AGI-2 reasoning vs Gemini 3 Pro | #1 proprietary, no tools | card |
| Humanity's Last Exam | 44.4% (51.4% w/ search+code) | New high for Google | Within ~6 pts of Opus 4.7 | card |
| ARC-AGI-2 | 77.1% | More than double Gemini 3 Pro | Class-leading (verified) | card |
| SWE-bench Verified | 80.6% | Up from ~63.8% (2.5 Pro) | Roughly tied Opus 4.6, behind GPT-5.3-Codex | card |
| Terminal-Bench 2.0 | 68.5% | New | Behind Gemini 3.5 Flash (76.2% on 2.1) | card |
| LiveCodeBench Pro | 2887 Elo | New high for Gemini | Top of public board | card |
| MMMU-Pro | 80.5% | +pts vs 2.5 Pro | Class-leading multimodal | card |
| MMMLU (multilingual) | 92.6% | Strong | Frontier | card |
| tau2-bench (Telecom / Retail) | 99.3% / 90.8% | New | Class-leading tool use | card |
| BrowseComp | 85.9% | New | Class-leading agentic search | card |
| MRCR v2 (128K) | 84.9% | Strong | Long-context leader | card |
| Artificial Analysis Index | 57 | New high for Gemini | Top tier | AA |
(MMLU-Pro, AIME 2025, MATH-500, HumanEval, IFEval, SimpleQA are not on Google's official card and are left null rather than sourced from inconsistent third parties.)
At ~142.7 output tokens/sec (Artificial Analysis median) and a high ~30.8s time-to-first-token, Gemini 3.1 Pro feels like what it is: a deliberate reasoning model. TTFT is dominated by the thinking phase, especially in Deep Think mode, where end-to-end latency can run 10-30s. For interactive chat this is noticeably slower than the Flash tier; for batch reasoning, long-document analysis, and agent loops the latency is well amortized. Streaming smooths perceived latency once the first token lands. Teams needing snappy UX should route easy turns to Flash and reserve 3.1 Pro for hard ones.
| Surface | Cost | Notes |
|---|---|---|
| API input (<=200K) | $2.00 / 1M tok | Standard tier |
| API output (<=200K) | $12.00 / 1M tok | Thinking tokens billed as output |
| API input (>200K) | $4.00 / 1M tok | Long-context tier |
| API output (>200K) | $18.00 / 1M tok | Long-context tier |
| Cached input | $0.20 (<=200K) / $0.40 (>200K) | + $4.50 / 1M tok / hour storage |
| Batch (in/out, <=200K) | $1.00 / $6.00 | ~50% off; async |
| Search grounding | 5,000 prompts/mo free (shared across Gemini 3), then $14 / 1,000 queries | Live Google Search |
| Direct UI | $19.99/mo (AI Pro); $100 & $200/mo (Ultra) | Formerly Gemini Advanced |
| Free tier | None | Pricing page: "Not available" for 3.1 Pro |
| Rate limits | Tiered; preview Tier 1 capped | Raises after cumulative spend gates |
Proprietary, closed-weights. Available via the Gemini API (Google AI Studio), Vertex AI, and Gemini Enterprise, plus Gemini CLI, Google Antigravity, and Android Studio. Vertex AI is the enterprise path: VPC Service Controls, CMEK, audit logging, regional pinning (US, EU, Asia), and data-residency commitments, plus the 2M-token context mode. No open weights, no self-hosting. OpenRouter resells API access. The trade-off is Google Cloud lock-in: IAM, Vertex tooling, and the two-tier pricing model are sticky once adopted, but the Workspace and BigQuery grounding is a moat no other provider matches.
Governed by Google's Frontier Safety Framework with configurable safety filters and adjustable thresholds on Vertex. Paid API and Vertex inputs are not used to train models; the AI Studio free tier may be. Opt-out is available. Compliance coverage is broad: SOC 2, HIPAA, GDPR, ISO 27001, FedRAMP, CCPA. Content moderation is built in. Refusal behavior is stricter than OpenAI and Anthropic on some safety-adjacent categories, which enterprises often see as a feature and creative users as friction.
SDKs in Python, TypeScript, Go, Java, and Dart; integrations with LangChain, LlamaIndex, Vercel AI SDK, Genkit, and Google ADK. It powers the Gemini app, Google AI Pro/Ultra, NotebookLM, Gemini CLI, Google Antigravity, Android Studio, and Workspace surfaces (Docs, Sheets, Gmail). Distribution through Android, Chrome, Search, and Workspace makes Gemini's effective reach dominant regardless of standalone API share.
Not formally. It launched 2026-02-19 in preview to validate before GA, and as of 2026-05-28 the API/Vertex surface is still under Pre-GA Offerings Terms. It is, however, the production-default model in the consumer Gemini app.
Prompts over 200K tokens bill at $4.00 input / $18.00 output per 1M (vs $2/$12 under 200K). Cached reads rise to $0.40 and storage is $4.50/1M tokens/hour.
The extended 2M context is rolling out on Vertex AI for enterprise; the standard Gemini API exposes 1M.
No for paid API and Vertex inputs. The free AI Studio tier may use inputs to improve products; opt-out is available.
3.5 Flash beats 3.1 Pro on Terminal-Bench, MCP Atlas, and CharXiv at lower cost and 1.5x speed. Use Pro when pure reasoning, HLE-class problems, or long-context recall dominate.
Google Search grounding gives 5,000 free prompts/month (shared across Gemini 3), then $14 per 1,000 queries — the deepest real-time integration of any frontier model.
No. Gemini is closed-weights, API/Vertex only.
Stronger creative tone and edges SWE-bench/expert tasks; weaker live-data access and long-context cost. Opus leads LMArena (~1504); 3.1 Pro Preview sits just behind.
Comparable frontier reasoning and ranks above 3.5 Flash on the AA Index; weaker native video ingestion and no Workspace/Search-grounding moat. The two trade frontier leads benchmark by benchmark.
Google's own sibling beats 3.1 Pro on agentic/coding tasks (Terminal-Bench 76.2%, MCP Atlas 83.6%) at lower price and higher speed, but loses on pure reasoning, HLE, and long-context recall. The honest internal split: Flash for agents, Pro for reasoning.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-05-27