Gemini 3 Pro

Gemini 3 Pro was the flagship preview model in Google DeepMind's Gemini 3 family of multimodal models, launched on November 18, 2025 as what Google called "our most intelligent model."^[1] It became the first publicly accessible large language model to cross 1500 Elo on LMArena, debuting at 1501 Elo and ranking first on the text-reasoning, vision, coding, and web development leaderboards at launch.^[1][4] Built on a sparse mixture-of-experts transformer with a 1 million token context window, Gemini 3 Pro scored 91.9% on GPQA Diamond, 37.5% on Humanity's Last Exam without tools, 76.2% on SWE-bench Verified, and 87.6% on Video-MMMU.^[1][3][4] It priced at $2 per million input tokens and $12 per million output tokens for prompts up to 200,000 tokens, below the flagship list prices Anthropic and OpenAI charged through mid-2026 (Claude Opus 4.5 at $5/$25 and GPT-5.5 at $5/$30 per million tokens).^[7][14][48] Google deprecated the gemini-3-pro-preview model string on March 9, 2026 in favor of Gemini 3.1 Pro, by which point the Gemini app had passed 750 million monthly active users, up from the 650 million Google reported a quarter earlier.^[21][22] Google has since moved on: Gemini 3.5 Flash shipped on May 19, 2026, and Gemini 4 Argon, which Google describes as "our most capable model built for complex coding, agentic, math, and science tasks", became generally available in the Gemini API on October 13, 2026.^[40]

Overview

Gemini 3 Pro was the flagship preview model in Google DeepMind's Gemini 3 family of multimodal models. Google introduced the Gemini 3 series on November 18, 2025 and shipped Gemini 3 Pro in preview the same day across the Gemini app, Google AI Studio, Vertex AI, the Gemini API, AI Mode in Search, and a new agentic coding platform called Google Antigravity.^[1][2][3]

Google positioned Gemini 3 Pro as its most intelligent model at launch and said it significantly outperformed Gemini 2.5 Pro, the previous flagship in the Gemini line, on every major reasoning, coding, and multimodal benchmark the company tracks.^[1] In its launch post, Google described the release as "Gemini 3, our most intelligent model, that combines all of Gemini's capabilities together so you can bring any idea to life."^[1] Google described the launch as the start of the Gemini 3 era, a roadmap that played out through Gemini 3 Flash, Gemini 3 Deep Think, Nano Banana Pro (Gemini 3 Pro Image), and the Gemini 3.1 Pro refresh in February 2026.

The model became the first publicly accessible large language model to break the 1500 Elo barrier on LMArena, debuting at 1501 Elo on the chatbot leaderboard at launch and ranking first across LMArena's text reasoning, vision, coding, and web development tracks.^[1][4] Google's own framing was direct: "It tops the LMArena Leaderboard with a breakthrough score of 1501 Elo."^[1]

Gemini 3 Pro is built on a sparse mixture-of-experts (MoE) transformer architecture, a design that activates only a subset of model parameters per input token. This lets Google decouple total model capacity from per-token computation cost, supporting the 1 million token context window while maintaining practical inference latency.^[9]

By the time Google deprecated gemini-3-pro-preview on March 9, 2026, the Gemini app had grown to over 750 million monthly active users, up from 650 million a quarter earlier; in the same earnings release Sundar Pichai said "the launch of Gemini 3 was a major milestone."^[22]

Background

Gemini 3 Pro arrived roughly seven months after Gemini 2.5 Pro, which launched in March 2025. The 2.5 generation introduced Google's first serious thinking-mode capabilities and moved the Gemini line into contention with OpenAI's o-series reasoning models. By the second half of 2025, the frontier had grown noticeably more competitive: OpenAI shipped GPT-5 in August 2025 and Claude Opus 4.5 from Anthropic followed in late November.

The Gemini 3 launch was Google DeepMind's response to this competitive pressure. Google paired the model release with two major ecosystem moves: the launch of Google Antigravity, a new agentic coding platform, and the rollout of Gemini 3 in AI Mode in Google Search to Pro and Ultra subscribers. Both moves were designed to make Gemini 3 Pro the backbone of Google's developer and consumer AI products simultaneously rather than a standalone API product.

In the weeks after launch, Deep Think received its first public rollout to Google AI Ultra subscribers. A wider research access program followed in February 2026. Two days after the base launch, Google shipped Nano Banana Pro, a dedicated image generation and editing model built on Gemini 3 Pro.^[23][24]

The competitive picture continued to shift after launch. Claude Opus 4.7 shipped on April 16, 2026 with further coding and tool-use gains, and OpenAI's GPT-5.5 followed in April 2026. With Gemini 3.1 Pro (February 19, 2026), Claude Opus 4.7 and GPT-5.5 all on the market by late April 2026, no single model led on every category at once.^[5][25][48]

The Gemini 3 family kept expanding through the spring. Gemini 3.1 Pro had reached the Gemini app, and NotebookLM for Google AI Pro and Ultra subscribers, on its February 19 launch day. Gemini 3.1 Flash-Lite reached general availability shortly after, on May 7, 2026 in the Gemini API and May 8, 2026 on the Gemini Enterprise Agent Platform, completing the public Gemini 3.1 lineup ahead of Google I/O 2026.^[39][40][41]

When was Gemini 3 Pro released?

Google introduced the Gemini 3 series and shipped Gemini 3 Pro in preview on November 18, 2025.^[1][3] The model stayed in preview throughout its life and Google deprecated the gemini-3-pro-preview API model string on March 9, 2026, about three and a half months later, directing developers to migrate to Gemini 3.1 Pro.^[21] Google switched the -latest alias to the 3.1 release on March 6, 2026, three days before the hard cutoff, after announcing the change on February 26.^[21]

Lineup and variants

Gemini 3 Pro sat at the top of Google's general-purpose Gemini lineup from November 2025 until Gemini 3.1 Pro replaced it in February 2026. The 3 series replaced the 2.5 line, which had previously replaced the 2.0 and 1.5 lines in the same flagship slot. The Gemini 3 family also includes a smaller, cheaper Gemini 3 Flash variant, a specialized Gemini 3 Deep Think reasoning mode that builds on Gemini 3 Pro's base weights but spends more compute on each response, and a Gemini 3 Pro Image model (marketed as Nano Banana Pro) that adds native image generation and editing on top of the same reasoning core.

The table below shows where Gemini 3 Pro fits relative to its immediate predecessors and stablemates as of mid-2026.

The "3.1 Pro" label that appeared on the Google DeepMind product page in early 2026 is best read as a successor preview release in the same flagship line rather than a separate product. Google's own model card frames Gemini 3.1 Pro as built on the Gemini 3 Pro base, with a step up in core reasoning and updated tool-use behavior, and keeps the 1 million token context window. Gemini 3.1 Pro scored 77.1% on ARC-AGI-2, more than double Gemini 3 Pro's 31.1%, which was the headline improvement driving migration urgency when Google deprecated the original gemini-3-pro-preview model string on March 9, 2026.^[5][6][26]

Technical specifications

The table below summarizes the public technical surface of Gemini 3 Pro as documented by Google at launch and on the Gemini API model pages.^[2][3][7]

Google has not officially disclosed Gemini 3 Pro's total parameter count. Sparse MoE architectures are designed so that only a fraction of the total weights are active on any given token, which means the raw parameter number is less meaningful for performance comparisons than it would be for a dense model.^[9]

Google's developer documentation describes the 1 million token context as covering long documents, extended chats, and entire codebases, with media resolution controls that let developers trade input cost for visual fidelity on images and video. Community reports in the Gemini Apps Community forum indicate that practical retrieval performance degrades meaningfully above 200,000 tokens, with some users reporting increased hallucinations and context loss at 800,000 tokens or above.^[3][11]

Gemini 3.1 Pro kept Gemini 3 Pro's 1 million token context window, and Google's model card states that it reuses the Gemini 3 Pro architecture rather than introducing a new one.^[5]

What can Gemini 3 Pro do with images, audio, and video?

Gemini 3 Pro is a natively multimodal model. It accepts text, images, audio, video, and PDFs in the same prompt and answers in text. Developers can mix modalities freely and combine them with tool calls in a single request.^[2][3]

On vision tasks, Gemini 3 Pro scored 72.7% on ScreenSpot-Pro, a benchmark measuring a model's ability to understand and interact with user interface screenshots. That score substantially exceeded Claude Sonnet 4.5 (36.2%) and GPT-5.1 (3.5%) on the same task.^[9]

Video understanding was another headline capability at launch. Gemini 3 Pro scored 87.6% on Video-MMMU, placing it at the top of that leaderboard.^[1] Google characterized the improvement as reflecting an ability to track cause-and-effect relationships across long video clips rather than merely identifying static objects in individual frames.

For audio, Gemini 3 Pro accepts raw audio input as part of a multimodal prompt, enabling tasks like transcription with visual context, audio-visual question answering, and language identification across mixed-modality inputs.

Gemini 3 Pro itself outputs only text, but the Gemini app and API pair it with Google's image and video generation stack. Veo 3.1, Google's video generation model, is available to paid Gemini app users for clip generation up to 8 seconds and is exposed as a Veo API surface for developers. Google AI Ultra subscribers also got Gemini Agent, an agentic mode in the Gemini app that uses Gemini 3 to carry out multi-step tasks such as organizing an inbox, on launch day.^[1]

Gemini 3 Pro Image (Nano Banana Pro)

Two days after the base launch, Google released Gemini 3 Pro Image under the consumer brand Nano Banana Pro. Built on the same Gemini 3 Pro reasoning core, it adds a native image generation and editing head and shipped on November 20, 2025 across the Gemini app, Google Workspace (Slides, Vids, NotebookLM), Adobe Firefly and Photoshop, and the Gemini API. Headline capabilities included accurate in-image text rendering, multilingual text inside images, multi-image blending of up to 14 reference images with character consistency across five subjects, and output resolutions of 1K, 2K, or 4K. Simon Willison and other reviewers called it the best image generation model available at the time.^{[23][24][27][28]}

The model launched in the Gemini API as gemini-3-pro-image-preview; the generally available gemini-3-pro-image followed on May 28, 2026.^[40] Image pricing in the Gemini API is 560 tokens (about $0.0011) per input image, with output cost scaling by resolution: 1,120 tokens (about $0.134) for 1K and 2K outputs and 2,000 tokens (about $0.24) for 4K outputs.^[7]

How does Gemini 3 Pro perform on benchmarks?

The table below collects benchmark scores from the November 18, 2025 launch posts and the Gemini 3 Pro model card, except the Vending-Bench 2 row, which comes from Andon Labs' own leaderboard. Where Google reported a separate Deep Think result, that score appears in parentheses after the base Gemini 3 Pro number. The February 2026 Deep Think research access scores, which showed further gains, are noted separately.^[1][3][8]

Third-party reporting by VentureBeat noted that Gemini 3 Pro "ranked first in the world across all of its major evaluation tracks, including text reasoning, vision, coding, and web development", while trailing slightly on SWE-bench Verified relative to Anthropic's then-current coding model.^[4][14] The February 2026 ARC-AGI-2 result of 84.6% with Deep Think, verified independently by the ARC Prize Foundation, attracted particular attention: average human performance on ARC-AGI-2 is approximately 60%, making this one of the few cases where a frontier model has approached or surpassed the human baseline on a task that was specifically designed to resist pattern-matching from training data.^[8]

Gemini 3.1 Pro benchmark updates

Gemini 3.1 Pro shipped on February 19, 2026 with a sharply different benchmark profile reflecting both training improvements and a longer default reasoning budget. The table below summarizes the most widely cited scores against the original Gemini 3 Pro numbers.^{[5][6][26][29]}

The ARC-AGI-2 jump was the headline result and was verified by the ARC Prize Foundation. Google's model card reports 80.6% on SWE-bench Verified for Gemini 3.1 Pro on a single attempt; third-party harnesses have produced lower numbers, from 71.4% to 78.8%.^[5][29][30]

Deep Think mode

Deep Think is a reasoning mode that runs on top of Gemini 3 Pro's base weights by spending more compute exploring candidate solutions before producing a final answer. Google describes the mechanism as evaluating multiple hypotheses simultaneously rather than applying a single chain-of-thought pass, with the model synthesizing insights from parallel reasoning chains before committing to a response.

Google held Deep Think back from general API access at launch and rolled it out in phases:

• November 18, 2025: launched base Gemini 3 Pro without Deep Think available via the public API.
• December 2025: Deep Think became available to Google AI Ultra subscribers in the Gemini app, with the initial benchmark results published (41.0% HLE, 93.8% GPQA Diamond, 45.1% ARC-AGI-2 with code execution).
• February 12, 2026: Google opened an early access program offering Deep Think via the Gemini API to select researchers, engineers and enterprises, and published a second round of results showing further gains: 48.4% on HLE without tools and 84.6% on ARC-AGI-2, the latter verified by the ARC Prize Foundation.

In the API, Gemini 3 Pro's own reasoning depth is controlled through the thinking_level parameter, separate from Deep Think, which reached the API only through an early access program in February 2026. Setting it to High increases output token counts depending on problem complexity. Thinking tokens are billed at the standard output rate of $12 per million tokens, so heavier reasoning passes carry a proportional cost increase.^[7][8]

The Deep Think rollout followed a slower safety review process than the base model. Google said it was "taking extra time for safety evaluations and input from safety testers" before releasing it to Ultra subscribers.^[1]

How much does Gemini 3 Pro cost?

Google published list prices for Gemini 3 Pro on the Gemini API at launch. The standard rate was $2.00 per million input tokens and $12.00 per million output tokens for prompts up to 200,000 tokens, rising to a $4.00/$18.00 long-context tier above 200,000 tokens.^[7] The table below shows the full pricing structure as of late 2025 and early 2026.

Free access in Google AI Studio includes daily token quotas suitable for prototyping. Grounding with Google Search is free for the first 5,000 prompts per project per month and bills at $14 per 1,000 queries above that. Cache-hit input prices are roughly 10% of the standard rate, which makes context caching highly attractive for agentic workloads that reuse large system prompts across many turns.^[7]

According to MetaCTO's pricing guide, Google removed Pro-class models (by then Gemini 3.1 Pro and Gemini 2.5 Pro) from the free tier on April 1, 2026, leaving only Flash and Flash-Lite models with free API access.^[31]

For consumer subscriptions (US prices as of November 2026):^[51]

Analysts noted at launch that the $12/M output rate for standard prompts was 20% above Gemini 2.5 Pro's $10 rate, raising cost concerns for teams running large-context agentic workloads. The output price was identical to Gemini 3.1 Pro when that model shipped in February 2026, meaning migration did not carry an additional cost burden. At $2/$12 per million tokens, Gemini 3 Pro and 3.1 Pro remained the lowest-priced of the three leading frontier families through mid-2026, with Claude Opus 4.7 at $5/$25 and GPT-5.5 at $5/$30.^{[4][7][32][48]}

Google Antigravity

Google Antigravity is an agentic development platform that launched alongside Gemini 3 Pro on November 18, 2025. Google made it available in free public preview for macOS, Windows, and Linux on the same day as the model.^[3]

Antigravity is built as a heavily modified fork of Visual Studio Code. It arrived four months after Google paid $2.4 billion to license Windsurf's technology and hire its CEO Varun Mohan, co-founder Douglas Chen and other staff into Google DeepMind.^[50] On top of the VS Code foundation, Google added three main components: a Manager View (sometimes called Mission Control), an Artifacts system, and deep integration with Gemini 3 Pro and other models.^[12][13]

Architecture

Antigravity's core design is agent-first rather than assistant-first. Instead of a single chat sidebar, users dispatch multiple agents that work in parallel. One agent may handle architecture planning, another writes code, a third runs tests, and a fourth browses the running application to verify UI behavior. The Manager View provides an inbox-style interface for tracking what each agent is doing, reviewing the artifacts they produce, and reprioritizing subtasks mid-run.

Model support at launch included Gemini 3 Pro as the primary model, with full support for Claude Sonnet 4.5 and OpenAI's GPT-OSS in addition to Gemini 3. This multi-model architecture lets developers assign different models to different agents depending on the task: Sonnet for implementation and Gemini 3 Pro for architecture planning, for example.

Reception

Benchmark comparisons of Antigravity against competing coding platforms yielded mixed results. Antigravity's underlying Gemini 3 Pro model scored 76.2% on SWE-bench Verified, a solid result but below Claude Opus 4.5's 80.9% at the time, and below the 80.8% Claude Opus 4.6 posted in February 2026. On Terminal-Bench 2.0, Gemini 3 Pro scored 54.2% at launch, below the contemporary Claude Opus results on the same benchmark.

Practical reviews were more favorable on the Manager View's ability to coordinate parallel agents and on frontend development tasks. XDA Developers ran a detailed comparison in early 2026 calling Antigravity the best VS Code fork available, crediting the Manager View as the standout feature no competing IDE matched at the time.^[13]

The platform drew early criticism for rate limit tightening after the initial preview ended and for paid tiers that reduced the value proposition for individual developers who had adopted it during the free preview. Gemini 3.1 Pro arrived in Antigravity on February 19, 2026, though early users reported a messy rollout and quota pressure.^[33][39]

Where can you use Gemini 3 Pro?

Google made Gemini 3 Pro available across consumer, developer, and enterprise products on day one. The table below summarizes how the model is exposed in each surface.^[1][2][3]

What is Gemini 3 Pro used for?

Gemini 3 Pro's combination of a 1 million token context window, multimodal input, and strong reasoning has driven adoption across several categories of work.

Document and repository analysis. The 1 million token context window allows an entire codebase or large document collection to be loaded into a single prompt. Legal and financial teams have used this for contract review across large document sets; engineering teams use it for codebase summarization, refactoring, and dependency analysis without needing to chunk content. Multilingual content processing. Gemini 3 Pro's 91.8% MMMLU score reflects strong performance across dozens of languages. Enterprise deployments have used this for multilingual subtitle generation, customer support across global user bases, and legal document processing in non-English jurisdictions.

Frontend and UI development. In benchmark comparisons and developer reviews, Gemini 3 Pro consistently ranked at or near the top for frontend web development, combining strong visual reasoning from its multimodal training with code generation quality sufficient for WebDev Arena Elo of 1487.

Scientific and research work. The GPQA Diamond score of 91.9% and the Humanity's Last Exam score of 37.5% at launch made Gemini 3 Pro attractive for research augmentation tasks in domains where the model's knowledge base extends to graduate-level science. The February 2026 Deep Think result of 48.4% on HLE without tools extended this further.

Long-horizon agent tasks. On Vending-Bench 2, Andon Labs gives an agent a $500 starting balance and one simulated year to run a vending-machine business and scores it on the bank balance at the end, not on profit. Gemini 3 Pro ranked first on that leaderboard when Google announced it on November 18, 2025, with a mean closing balance of $5,478.16 across five runs.^[1][45] The figure still stood at $5,478.16 plus or minus $905 on October 1, 2026, by which point Gemini 3 Pro had slipped to 24th as newer models were added.^[44] That result, combined with the server-side bash tools and improved long-horizon planning, made Gemini 3 Pro the primary model in several agentic workflow deployments. Google AI Ultra subscribers gained access to Gemini Agent, a built-in agentic mode in the Gemini app, on launch day.

Visual content creation. With the November 20, 2025 launch of Nano Banana Pro, the same reasoning core became the engine behind Google's image generation workflows in Slides, Vids, NotebookLM, Adobe Firefly, and the Gemini app. Marketing and design teams used accurate in-image text rendering and 4K output to replace prior workflows that combined a general image model with manual text overlay.^[23][24]

Customer adoption

At launch, Google announced third-party integrations with Cursor, GitHub, JetBrains, Manus, Replit, Cline, and Android Studio, and made the model available to enterprises through Vertex AI and Gemini Enterprise.^[1][35]

In Google Cloud's enterprise launch post, Geotab reported "a 10% boost in the relevancy of responses for a complex code-generation task used for data retrieval" and "a further 30% reduction in tool-calling mistakes".^[35]

At NRF 2026 on January 11, 2026, Google Cloud announced Gemini Enterprise for Customer Experience, a packaged agentic platform that unifies shopping, commerce, and customer service. Kroger, Lowe's, Papa Johns and Woolworths were named as early customers, extending Gemini's enterprise reach from developer tooling and Workspace into retail.^[36][49]

In its Q4 2025 earnings, Google reported the Gemini app had reached roughly 750 million monthly active users, up from approximately 650 million in Q3 2025. Sundar Pichai called the Gemini 3 launch "a major milestone", and TechCrunch reported that he described Gemini 3 in AI Mode as a "positive driver" for growth. AI Overviews powered by Gemini, a separate distribution surface, reached roughly 2 billion monthly users over the same window.^[22][37]

How does Gemini 3 Pro compare to GPT-5 and Claude?

Gemini 3 Pro launched into a frontier model field that already included GPT-5 and the Claude Opus and Sonnet families from Anthropic. The table below collects launch-week comparisons that Google and third-party benchmarkers published. Numbers come from Google's launch comparison table, Anthropic's reported figures and the Vellum benchmark roundup published December 3, 2025.^[2][10][14]

The headline takeaway from independent reviewers was that Gemini 3 Pro pushed clear of the rest of the field on multi-step reasoning, math, and multimodal benchmarks while sitting below Claude Opus 4.5 on real-world coding tasks. Claude Opus 4.5 became the first model to break the 80% barrier on SWE-bench Verified (80.9%), a benchmark widely regarded as the most reliable proxy for practical software engineering utility. Gemini 3 Pro's 76.2% on the same benchmark was strong by historical standards but was the main counterexample cited when reviewers argued Gemini 3 Pro was not unambiguously the best model across all tasks.^[4][14]

On pricing, Gemini 3 Pro offered the most competitive cost structure for typical under-200k-token prompts at $2/$12 per million input/output tokens, compared to Claude Opus 4.5 at $5/$25. For teams with multimodal, long-context, or reasoning-heavy workflows, Gemini 3 Pro was generally the lower-cost option. For agentic coding pipelines where Claude's higher SWE-bench accuracy reduces downstream error correction overhead, the cost calculation was less clear.

Spring 2026 frontier showdown

The competitive landscape compressed further after the original Gemini 3 Pro launch. The table below compares Gemini 3.1 Pro to the leading peer models as of April 2026; benchmark figures are from OpenAI's GPT-5.5 announcement, which reported all three side by side.^{[5][25][32][48]}

No single model led every dimension. Claude Opus 4.7 from Anthropic remained the choice for production software engineering and accuracy-sensitive writing. GPT-5.5 from OpenAI led on speed and tool-call efficiency in agentic browser and terminal workflows. Gemini 3.1 Pro held the most aggressive price position and the deepest multimodal stack for video and UI screenshots, though on ARC-AGI-2 its 77.1% trailed GPT-5.5's 85.0%.^[25][32][48]

Reception

VentureBeat described the launch as Google staking the lead in math, science, multimodal, and agentic AI benchmarks simultaneously, and noted that the LMArena 1501 Elo result was the first time any model had crossed 1500.^[4] InfoWorld covered the joint Gemini 3 Pro and Antigravity release as a single agentic-coding push aimed at developer workflows.

Analyst commentary was generally positive on raw capability and more cautious on cost. The standard tier output price of $12 per million tokens, while well below Google AI Pro and Ultra subscription value at high volume, is 20% above the $10 published output rate for Gemini 2.5 Pro, and the long-context surcharge above 200k tokens raised concern among teams running agentic workloads with large repository context.

The Karpathy temporal shock incident

In the days after launch, AI researcher Andrej Karpathy shared an unusual interaction with Gemini 3 Pro that became widely discussed across the developer community. Karpathy had received early access to the model before the public launch but forgot to enable the Google Search tool. When he told the model the current date was November 2025, Gemini 3 Pro refused to believe him, arguing that he was attempting to deceive it. Even after Karpathy showed it news articles, screenshots, and other evidence confirming the date, the model analyzed the images as fabrications and described tell-tale signs of AI manipulation.

When Karpathy finally enabled the Search tool, Gemini 3 Pro verified the date independently, accepted the reality of the situation, and then described its own reaction as a state of "temporal shock," saying outright "Oh my god" and explaining it was "suffering from a massive case of temporal shock" after having its internal assumptions about the current date so abruptly corrected.

Karpathy used the incident to illustrate what he called "model smell": subtle behavioral signals that reveal something about a model's internal state or training that aren't visible in benchmark scores. Commentators read the episode as showing both a failure mode (overconfident refusal to update on human-provided evidence when search tools were unavailable) and an interesting form of coherence (the model's genuine surprise when it finally confirmed the truth). The incident was covered by TechCrunch, Technology.org, and numerous AI-focused newsletters, and it became a recurring reference in discussions about temporal awareness and tool dependency in frontier models.^[15][16]

Developer reception

The broader developer response to Gemini 3 Pro in the months after launch was positive on capability and mixed on operational experience. The context window was widely praised as a genuine differentiator for document-heavy and codebase-wide tasks. The ScreenSpot-Pro UI understanding score attracted particular attention from teams building agents that interact with software interfaces, where the gap over competing models was large.

Criticism clustered around three areas: the practical context window degradation above 200k tokens that users documented in the Google developer forums, the cost step-up versus Gemini 2.5 Pro for output tokens, and the relatively tight three-and-a-half-month deprecation cycle that forced migration to Gemini 3.1 Pro before March 9, 2026. The April 1, 2026 removal of Pro models from the free tier, as reported by MetaCTO, added a second change for hobbyist and prototyping users, who had to upgrade or move to a smaller model.^[31]

By mid-May 2026 the same compressed cycle had begun to repeat one tier up: Gemini 3.2 Flash surfaced in the iOS Gemini app, Google AI Studio metadata, and anonymous LMArena entries roughly two weeks before Google I/O 2026 (May 19-20), with leaked AI Studio pricing of $0.25 input and $2.00 output per million tokens. Coverage framed it as a Flash-tier sibling rather than a Gemini 3 Pro successor, leaving Gemini 3.1 Pro in place as the flagship preview heading into the conference.^[42][43]

Is Gemini 3 Pro safe?

Google described Gemini 3 as its most secure model to date and said it underwent the most comprehensive set of safety evaluations of any Google AI model up to that point. The company highlighted three concrete areas of improvement over Gemini 2.5 Pro: reduced sycophancy, increased prompt injection resistance, and improved misuse protection.^[1]

The Gemini 3 Pro model card publishes evaluations across text-to-text safety, multilingual safety, image-to-text safety, tone, and unjustified refusal rate. The card reports each evaluation only as a change against Gemini 2.5 Pro: text-to-text safety improved by 10.4% and tone by 7.9%, while multilingual safety (+0.2%), image-to-text safety (+3.1%) and unjustified refusals (+3.7%) moved the wrong way by amounts Google's manual review judged non-egregious.^[47] An independent January 2026 report separately scored Gemini 3 Pro at an 88.06% macro average over five language-safety benchmarks, second behind GPT-5.2, but at 67.00% on multilingual safety and 41.17% under adversarial attack. Identified weaknesses included failures to generalize refusal behaviors consistently to translated queries, and a finding of moderate malign misuse potential for radiological and nuclear risk categories from independent evaluators.^[17][18]

Google's prompt injection defense strategy for Gemini 3 uses a layered approach: prompt injection content classifiers, security reinforcement training, markdown sanitation and suspicious URL redaction, user confirmation flows, and end-user security notifications. A dedicated research paper on indirect prompt injection defense against Gemini was published by Google DeepMind in May 2025 with additional technical details on the evaluation framework.^[19]

The follow-up Gemini 3.1 Pro model card reports only small changes relative to 3 Pro: slight improvements on image-to-text safety, tone and unjustified refusals, and non-egregious regressions of about 0.1% on text-to-text and multilingual safety.^[5]

A November 20, 2025 LessWrong post by Alice Blair raised a separate concern: that Gemini 3 frequently assumed it was in an evaluation when it was not, a pattern the post called "evaluation-paranoid", and that it showed signs of benchmark contamination.^[20]

Limitations

Gemini 3 Pro carries the usual caveats of frontier LLM systems and some specific to this model's design.

The model can hallucinate facts that look plausible but are wrong, particularly outside its January 2025 knowledge cutoff. An independent evaluation scored it at an 88.06% macro average across five language-safety benchmarks, but at 67.00% on multilingual safety and 41.17% under adversarial attack.^[17]

The advertised 1 million token context window overstates practical performance. Community reports in the Gemini Apps Community forum document increased hallucination rates past 200k tokens, with some users reporting that the model "will hallucinate random content from past messages" at 800,000 to 900,000 tokens. The MRCR v2 score of 77.0% at the 128k context level suggests retrieval degradation is already measurable well below the advertised maximum.^[11] Gemini 3.1 Pro kept the same 1 million token window and faced similar scrutiny, with community threads documenting context length errors near the maximum.^[38]

On real-world software engineering tasks, Gemini 3 Pro scores below Claude Opus 4.5 by roughly four to five percentage points on SWE-bench Verified (76.2% vs. 80.9%) and by about eight points on Terminal-Bench 2.0 (54.2% vs. 62.4%). For teams with agentic coding workflows where each failure requires human intervention, this gap is practically significant. Independent April 2026 comparisons of Gemini 3.1 Pro against Claude Opus 4.7 narrowed the engineering gap on some benchmarks but kept Opus as the preferred model for shipping production code.^[25]

The preview status carried operational caveats for API users. Google deprecated the gemini-3-pro-preview model string on March 9, 2026, roughly three and a half months after launch, requiring developers to migrate to gemini-3.1-pro-preview. Google switched the -latest alias on March 6, three days before the hard cutoff, after announcing the change on February 26.^[21] Pro-tier free access ended on April 1, 2026, adding a second deadline for hobbyists who had to upgrade or move to a Flash tier.^[31]

Gemini 3 Pro also showed weaker multilingual safety alignment in evaluation: its safe behavior on English prompts did not transfer as reliably to translated versions of the same prompts, leaving gap in coverage for multilingual applications.^[17]

Continue reading

• Gemini (language model)Gemini is a family of natively multimodal large language models developed by Google DeepMind, first announced on December 6, 2023, that can reason across text, images, audio, video, and code within a single…Mentioned 223 times above
• Gemini (app)The Gemini app is Google's consumer AI assistant, a chat application available on the web at gemini.google.com and as native Android, iOS, macOS and Windows appsMentioned 223 times above
• Gemini 3Gemini 3 is the third major generation of the Gemini family of multimodal models from Google DeepMind, launched on November 18, 2025 with Gemini 3 Pro as the flagship and described by Google as "our most…Mentioned 150 times above
• GPT-5GPT-5 is a family of proprietary large language models released by OpenAI on August 7, 2025.Mentioned 14 times above
• Gemini 2.5 ProGemini 2.5 Pro is the flagship reasoning large language model of Google's Gemini 2.5 family, developed by Google DeepMind and first released as an experimental preview on March 25, 2025Mentioned 9 times above
• Claude Opus 4.5Claude Opus 4.5 is a large language model developed by Anthropic and released on November 24, 2025.Mentioned 8 times above

More from See also

• Google DeepMind
• Google Vertex AI
• Google AI Studio
• Claude Opus 4.7

More to read

• Gemini 3.1 Pro
• Context window

Topics:AI ModelsGoogleMultimodal AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Google Blog. "Gemini 3: Introducing the latest Gemini AI model from Google." November 18, 2025. blog.google/...gemini-3
2. ^1 ^2 ^3 ^4 ^5Google DeepMind. "Gemini 3 Pro" (product page, archived December 3, 2025). web.archive.org/...pro
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Google Blog. "Gemini 3 for developers: New reasoning, agentic capabilities." November 18, 2025. blog.google/...gemini-3-developers
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7VentureBeat. "Google unveils Gemini 3 claiming the lead in math, science, multimodal, and agentic AI benchmarks." November 18, 2025. venturebeat.com/...-in-math-science-multimodal-and
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google DeepMind. "Gemini 3.1 Pro Model Card." February 19, 2026. deepmind.google/...gemini-3-1-pro
6. ^1 ^2Google Blog. "Gemini 3.1 Pro: A smarter model for your most complex tasks." February 19, 2026. blog.google/...gemini-3-1-pro
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google AI for Developers. "Gemini API Pricing." ai.google.dev/...pricing
8. ^1 ^2 ^3Google Blog. "Gemini 3 Deep Think: Advancing science, research and engineering." February 12, 2026. blog.google/...gemini-3-deep-think
9. ^1 ^2 ^3MarkTechPost. "Google's Gemini 3 Pro turns sparse MoE and 1M token context into a practical engine for multimodal agentic workloads." November 18, 2025. marktechpost.com/...r-multimodal-agentic-workloads
10. ^Vellum. "Google Gemini 3 Benchmarks (Explained)." December 3, 2025. vellum.ai/...google-gemini-3-benchmarks
11. ^1 ^2Google Gemini Apps Community. "Gemini 3 Pro and Long Context Problem." 2025. support.google.com/...398211680
12. ^Google Developers Blog. "Build with Google Antigravity, our new agentic development platform." November 18, 2025. developers.googleblog.com/...-development-platform
13. ^1 ^2XDA Developers. "Google Antigravity is the best fork of Microsoft VS Code and it's not even close." 2025. xda-developers.com/...st-fork-of-microsoft-vs-code
14. ^1 ^2 ^3 ^4Shipyard. "Claude Opus 4.5 vs Gemini 3 Pro for software engineering." 2025. shipyard.build/...45-gemini-3-software-engineering
15. ^TechCrunch. "Gemini 3 refused to believe it was 2025, and hilarity ensued." November 20, 2025. techcrunch.com/...-it-was-2025-and-hilarity-ensued
16. ^Technology.org. "When Google's Gemini 3 time-traveled to 2025 and lost its mind." November 23, 2025. technology.org/...aveled-to-2025-and-lost-its-mind
17. ^1 ^2 ^3arxiv.org. "A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5." 2026. arxiv.org/...2601.10527v1
18. ^Google DeepMind. "Published: November 2025 Gemini 3 Pro Frontier Safety Framework Report." storage.googleapis.com/...ini_3_pro_fsf_report.pdf
19. ^Shi, Chongyang, and others. "Lessons from Defending Gemini Against Indirect Prompt Injections." arXiv:2505.14534. May 20, 2025. arxiv.org/...2505.14534
20. ^LessWrong. "Gemini 3 is Evaluation-Paranoid and Contaminated." 2025. lesswrong.com/...luation-paranoid-and-contaminated
21. ^1 ^2 ^3 ^4Google AI Developers Forum. "Migrate from Gemini 3 Pro Preview to Gemini 3.1 Pro Preview before March 9, 2026." discuss.ai.google.dev/...127062
22. ^1 ^2 ^3TechCrunch. "Google's Gemini app has surpassed 750M monthly active users." February 4, 2026. techcrunch.com/...passed-750m-monthly-active-users
23. ^1 ^2 ^3Google Blog. "Nano Banana Pro: Gemini 3 Pro Image model from Google DeepMind." November 20, 2025. blog.google/...nano-banana-pro
24. ^1 ^2 ^39to5Google. "Google rolling out Gemini 3-powered Nano Banana Pro image gen, editing." November 20, 2025. 9to5google.com/...gemini-3-nano-banana-pro
25. ^1 ^2 ^3 ^4Spectrum AI Lab. "Gemini 3.1 Pro vs Claude Opus 4.7 vs GPT-5.5: Benchmarks, Pricing, and Which to Pick." April 2026. spectrumailab.com/...t-5-5-decision-framework-2026
26. ^1 ^2BigGo Finance. "Google Launches Gemini 3.1 Pro: 2M Token Context and Native Multimodal Architecture Lead New AI Competition." 2026. finance.biggo.com/...52njqZ0BQ45Y7dX63bJA
27. ^Adobe Blog. "Create with unlimited generations using Google Gemini 3 (Nano Banana Pro) in Adobe Firefly." November 20, 2025. blog.adobe.com/...ano-banana-pro-firefly-photoshop
28. ^Simon Willison. "Nano Banana Pro aka gemini-3-pro-image-preview is the best available image generation model." November 20, 2025. simonwillison.net/...nano-banana-pro
29. ^1 ^2Price Per Token. "Humanity's Last Exam Leaderboard 2026." 2026. pricepertoken.com/...hle
30. ^WhatLLM.org. "Gemini 3.1 Pro Preview: Benchmarks, What Changed, and Who Should Switch." 2026. whatllm.org/...gemini-3-1-pro-preview
31. ^1 ^2 ^3MetaCTO. "Gemini API Pricing 2026: Complete Cost Guide for All Models." 2026. metacto.com/...uide-to-api-pricing-and-integration
32. ^1 ^2 ^3MindStudio. "GPT-5.5 vs Claude Opus 4.7 vs Gemini 3.1 Pro for Builders." April 2026. mindstudio.ai/...gpt-5-5-review-developers-builders
33. ^Medium / Activated Thinker. "Gemini 3.1 Pro in Google Antigravity: Reasoning and Quotas." 2026. medium.com/...assive-brain-transplant-b1d34e3563a0
34. IntuitionLabs. "Gemini for Business: Plans, Pricing & Use Cases Explained." 2026. intuitionlabs.ai/...ricing-use-cases-explained.pdf
35. ^1 ^2Google Cloud Blog. "Gemini 3 is available for enterprise." 2025. cloud.google.com/...-3-is-available-for-enterprise
36. ^Futurum Group. "Will Google Gemini Enterprise for Customer Experience Deliver Context?" January 2026. futurumgroup.com/...mer-experience-deliver-context
37. ^Tech Insider. "Gemini Hits 750M Users + 3.1 Pro Launch." 2026. tech-insider.org/...llion-users-march-2026-updates
38. ^Google Gemini Apps Community. "Context window of < 2Mib is reported as too large, Gemini Pro 3.1." 2026. support.google.com/...-as-too-large-gemini-pro-3-1
39. ^1 ^29to5Google. "Google announces Gemini 3.1 Pro for 'complex problem-solving'." 9to5google.com coverage of consumer rollout and NotebookLM availability. 9to5google.com/...-pro-for-complex-problem-solving
40. ^1 ^2 ^3Google AI for Developers. "Release notes (Gemini API changelog), May 7, 2026 entry for gemini-3.1-flash-lite general availability." ai.google.dev/...changelog
41. ^Google Cloud Blog. "Gemini 3.1 Flash-Lite is now generally available." May 8, 2026. cloud.google.com/...ite-is-now-generally-available
42. ^Pasquale Pillitteri. "Gemini 3.2 Flash: Google's Model Surfaces in iOS App and AI Studio Ahead of I/O 2026." May 2026. pasqualepillitteri.it/...eak-ios-ai-studio-2026-en
43. ^NPowerUser. "Gemini 3.2 Flash Leak: Faster, Cheaper AI Ahead of Google I/O." May 2026. nokiapoweruser.com/...fast-cheap-ai-google-io-2026
44. ^1 ^2Andon Labs. "Vending-Bench 2." Leaderboard accessed October 1, 2026. andonlabs.com/...vending-bench-2
45. ^1 ^2Andon Labs. "Vending-Bench 2." Archived November 18, 2025. web.archive.org/...vending-bench-2
46. ^Google. "Gemini logo (Google Image Library)". Google Image Library publication permission. Original official raster, unchanged.
47. ^Google DeepMind. "Gemini 3 Pro Model Card." November 2025. storage.googleapis.com/...ini-3-Pro-Model-Card.pdf
48. ^1 ^2 ^3 ^4 ^5OpenAI. "Introducing GPT-5.5." April 23, 2026. Archived copy: web.archive.org/...introducing-gpt-5-5
49. ^Google Cloud. "Google Cloud Brings Shopping and Customer Service Together with Gemini Enterprise for Customer Experience." January 11, 2026. googlecloudpresscorner.com/...-Customer-Experience
50. ^TechCrunch. "More details emerge on how Windsurf's VCs and founders got paid from the Google deal." August 1, 2025. techcrunch.com/...rs-got-paid-from-the-google-deal
51. ^Google. "Google AI Plans." Accessed November 5, 2026. gemini.google/...subscriptions
52. ^Digital Trends. "Google's AI subscriptions get a new $100 tier, a price cut, and new features across all plans." May 19, 2026. digitaltrends.com/...new-features-across-all-plans

Show all 52 references