Gemini 3

The launch and the narrow documented claim

On 18 November 2025, Google announced Gemini 3, the newest generation of its flagship model family, in a post credited to chief executive Sundar Pichai. The official announcement described two headline artefacts: Gemini 3 Pro, released in preview, and Gemini 3 Deep Think, a higher-compute reasoning mode initially restricted to safety testers and Google AI Ultra subscribers. Demis Hassabis, chief executive of Google DeepMind, and Koray Kavukcuoglu, the company's chief AI architect, were the technical leads named alongside Pichai.

The narrow, documented event is precise and worth stating plainly before any interpretation. On a specific date, a vendor shipped a new model, published a set of self-reported benchmark scores, and simultaneously wired the model into consumer and developer products. That is what can be verified. The broad claim that accompanied it, that Gemini 3 is "the best model in the world" for reasoning, multimodal understanding, and coding, is a marketing judgment about a competitive field that shifts weekly, and it should be read as separable from the launch itself. This report keeps those two levels apart: the dated announcement and its numbers on one side, the historical significance on the other.

The more durable part of the story may not be any single benchmark. It is that Gemini 3 arrived already embedded in Google Search on day one, a distribution move with no real precedent in the company's earlier model releases. That single fact, a frontier model wired into the world's most used piece of software from the moment of release rather than trialed quietly for months first, is the detail most likely to be cited later, long after the specific benchmark percentages have been superseded by whichever model ships next.

Predecessors and the constraint Gemini 3 was built to address

Gemini 3 is the fourth major step in a family that began in December 2023. The original Gemini technical report introduced a natively multimodal architecture trained jointly on text, images, audio, and video rather than stitching separate encoders together after the fact. In 2024 the Gemini 1.5 report pushed a different constraint, long context, demonstrating reasoning across millions of tokens and introducing the mixture-of-experts design that let a large model activate only a fraction of its parameters per token.

The constraint Gemini 3 targets is reasoning depth. That framing did not originate with this release either. In March 2025 Google shipped Gemini 2.5 Pro and described it, in its thinking-model announcement, as "a thinking model, designed to tackle increasingly complex problems," capable of "reasoning through their thoughts before responding." Gemini 2.5 already topped the LMArena leaderboard and shipped with a one-million-token context window. So Gemini 3 inherited native multimodality from 2023, long context from 2024, and explicit test-time reasoning from early 2025. Its distinctive contribution is best understood as pushing each of those inherited capabilities further under a single training run, not inventing any of them. Reading the release this way avoids the common error of treating each new model as a clean break rather than an increment on a documented trajectory.

This layered view also explains why Google's own marketing language for Gemini 3 sounds continuous with what it said about Gemini 2.5 eight months earlier. Both releases describe a model that reasons before answering; the difference between them is depth, reliability, and the range of tasks the reasoning holds up on, not the presence of the behavior itself. A reader who treats every release announcement as though it describes a wholly new capability will overstate how much changed at each step and understate how much of the current model's strength was already present, in weaker form, a year earlier.

How the model is built: sparse experts and native multimodality

Google has not published a full architecture paper for Gemini 3, so specifics come from its model card and developer documentation rather than a peer-reviewed description. The Gemini 3 developer guide and the accompanying model card describe Gemini 3 Pro as a natively multimodal, sparse mixture-of-experts transformer with a knowledge cutoff of January 2025. It accepts text, images, audio, and video as input, up to one million tokens, and produces up to roughly 64,000 output tokens.

The mixture-of-experts idea is older than any Gemini model. It descends from work such as the Switch Transformer, which showed that a network could hold a very large parameter count while routing each token to only a small subset of specialized subnetworks, decoupling total capacity from the compute spent per token. In practice this is why a model can carry a vast reservoir of memorized knowledge yet answer at the latency and cost of something much smaller: only a slice of the network fires for any given input. Applied to Gemini 3, the approach lets Google scale capacity for reasoning and multimodal breadth without a proportional increase in serving cost, which matters because the model was deployed to Search, a surface that answers billions of queries.

Native multimodality is the second structural choice. Because the model was trained from the start across modalities rather than bolting a vision module onto a language model, the same network reasons over a diagram, a video frame, and a paragraph within one representation. That design lineage runs straight back to the first Gemini report and is not new to version three. The practical consequence of this choice shows up less in benchmark tables than in ordinary use: a model built this way can, for example, watch a short video and answer a question about an event thirty seconds in without first converting the clip to a text description and losing detail in the conversion. Bolted-together systems tend to degrade at exactly that kind of cross-modal question, because the visual and textual reasoning were never trained to share a representation in the first place.

The reasoning controls: thinking level, Deep Think, and thought signatures

The most concrete engineering additions in Gemini 3 are the controls that expose test-time reasoning to developers. According to the developer blog, the API introduces a "thinking level" parameter that lets a caller trade latency and cost against reasoning depth, a "media resolution" parameter to control how finely images and video are processed, and stricter validation of "thought signatures," which preserve a model's reasoning state across multi-turn tool-calling conversations.

These are not cosmetic. A thinking-level knob makes the compute a model spends per answer an explicit, billable choice rather than a fixed property of the model, which is a meaningful shift in how reasoning models are operated. Gemini 3 Deep Think sits at the far end of that spectrum: a mode that spends substantially more inference-time computation before responding, released first to safety testers and then to the top consumer tier. The distinction matters for reading the benchmark table below, because several of Google's strongest reported numbers come from Deep Think or from configurations that permit code execution, not from the default Pro settings a typical API call would use. Conflating the two produces an inflated picture of what the base model does out of the box. The honest reading treats "Gemini 3" as a family of operating points, priced and benchmarked differently, rather than a single fixed capability.

The thought-signature mechanism deserves a closer look because it addresses a real, previously unsolved problem in agentic use of reasoning models. When a model reasons across several tool calls in a single task, such as searching, reading a result, and then deciding what to search for next, the intermediate reasoning that connected those steps was previously discarded between calls, forcing the model to re-derive context it had already worked out. Preserving that reasoning state as a signed token that the API validates on the next call lets a multi-step agent stay coherent across a longer chain of actions without repeating work or, worse, contradicting an earlier step. This is an engineering answer to a problem that became visible only once reasoning models were pushed into agentic workflows, and it is one of the few genuinely new mechanisms in the release rather than a scaled-up version of something older.

The reported evidence, and how to read a vendor benchmark table

Google published a dense results table, and the figures are genuinely high, but they are vendor-reported and should be labelled as such. On LMArena, the community-voted comparison arena, Google said Gemini 3 Pro reached 1501 Elo, the first model reported above the 1500 threshold. On GPQA Diamond, a set of graduate-level science questions designed to be resistant to web search, Google reported 91.9 percent for Pro and 93.8 percent for Deep Think, against roughly 65 percent for domain experts with PhDs on the original benchmark.

On Humanity's Last Exam, a 2,500-question benchmark built by over a thousand contributors to sit "at the frontier of human knowledge," Google reported 37.5 percent for Pro without tools and 41.0 percent for Deep Think. Other cited figures included 23.4 percent on MathArena Apex, 45.1 percent for Deep Think on ARC-AGI-2 with code execution, up from 4.9 percent for Gemini 2.5 Pro, 76.2 percent on SWE-bench Verified for coding, and 1487 Elo on WebDev Arena.

Three cautions apply to every row. First, the numbers come from the model's maker, and independent replication lags a launch by weeks. Second, LMArena is decided by human preference votes, so it measures which answers people like, not correctness, and Google itself acknowledged the ranking was preliminary. Third, benchmark scores at this altitude are increasingly vulnerable to contamination, since a January 2025 knowledge cutoff overlaps with the publication of many test sets. High scores are evidence of capability, not proof of it, and the correct posture is provisional acceptance pending outside checks.

There is a further reason to read the ARC-AGI-2 jump with particular care. A rise from 4.9 percent to 45.1 percent across one model generation is an unusually large single-step gain for a benchmark specifically designed to resist the kind of pattern memorization that improves scores without improving underlying reasoning. Part of that jump is attributable to the code-execution configuration rather than to the base model's reasoning alone, since being able to write and run a verification script changes what counts as solving a puzzle. None of this makes the result illegitimate, but it does mean the comparison to the 4.9 percent figure, which was measured under different tool-access conditions, is not a clean apples-to-apples comparison, and coverage that presents it as a ten-times improvement in raw reasoning ability is overstating what changed.

Distribution as the actual event: Search, the app, and developer tools

The feature most likely to matter historically is not a benchmark. It is that Gemini 3 shipped into Google Search's AI Mode on launch day. As CNBC reported, this was the first time Google put a new frontier model into its search engine from day one, rather than testing it in a separate app for months first. Pichai framed the goal as getting "what you need with less prompting."

The same model reached the Gemini app and the full developer stack simultaneously: Google AI Studio, Vertex AI, the Gemini CLI, and third-party coding environments including Cursor, GitHub, JetBrains, and Replit. Pricing, per the developer blog, was set at two dollars per million input tokens and twelve dollars per million output tokens for prompts up to 200,000 tokens during preview.

This breadth is the through-line of the launch. A model that lives only in a chatbot reaches millions of people; a model wired into the default search engine of much of the world reaches billions, most of whom will never know a model version changed. That is a distribution fact, independent of whether the model wins any particular benchmark, and it is the dimension on which Gemini 3 most clearly differs from its own predecessors. VentureBeat's coverage noted that this reach, backed by Google's existing surfaces, was as central to the story as any leaderboard position.

The pricing detail is worth dwelling on because it signals how Google expects the model to be used. Two dollars per million input tokens and twelve dollars per million output tokens places Gemini 3 Pro in a comparable band to the other frontier-tier models on the market at launch, not meaningfully cheaper. That matters because a model this capable, priced at frontier rates, is not obviously the right default for high-volume, low-complexity tasks; the thinking-level parameter exists in part so a developer can dial reasoning down, and therefore cost down, for the large share of requests that do not need the model's full depth. The pricing and the reasoning controls are two parts of the same design decision, not separate announcements.

Generative interfaces and the agentic turn

Two capabilities positioned Gemini 3 as more than a better text engine. The first is generative interfaces. Rather than returning prose, the model can assemble a custom visual layout for a query: an immersive, magazine-style page with images and interactive modules generated on the fly. Google Research described this in its generative UI post as "a rich, custom, visual interactive user experience for any prompt," with two initial experiments, visual layout and dynamic view, offered first to paying subscribers in the United States.

The second is agentic development. Alongside the model, Google launched Antigravity, described as "our agentic development platform, allowing anyone to build in the agent-first era." Antigravity is an IDE plus a command center for running multiple coding agents in parallel across a workspace, using Gemini 3's tool-use and long-horizon planning as the engine. Google reported 54.2 percent on Terminal-Bench 2.0, a test of operating a computer through a terminal, as evidence for this agentic framing.

Both features point the same direction: the model is being sold not as a component you query but as an operator that produces interfaces and executes multi-step work. Whether that framing holds up depends on reliability over long task chains, which benchmarks measure only partially. The capability is real and demonstrated; its dependability at scale is the open question, and it is the right thing to withhold judgment on until independent, adversarial testing accumulates.

Generative interfaces and agentic development also share a less obvious implication: both quietly shift where the interesting engineering problem sits. A model that writes prose has one output to evaluate for correctness. A model that writes an interactive interface, or that decides which of several coding agents should attempt which subtask, produces an output whose correctness cannot be reduced to a single pass or fail check; it succeeds or fails across many small decisions compounding over a session. That is a genuinely harder evaluation problem than a text benchmark, and it is one reason the launch materials lean on demonstrations and case studies for these two features rather than the numeric benchmark rows used for reasoning and coding.

The competitive field: GPT-5.1, Claude, and open-weight pressure

Gemini 3 did not arrive in a vacuum, and its claims are only legible against rivals. At launch it was compared most directly with OpenAI's GPT-5.1 and Anthropic's Claude Opus family. Google's own tables showed Gemini 3 Pro ahead of both on many reasoning and multimodal benchmarks, and independent aggregator Artificial Analysis, cited in contemporaneous coverage, moved Google to the top of its composite index shortly after release. Early reviewers generally described Gemini 3 Pro as marginally ahead of GPT-5.1 and roughly comparable to the strongest Claude models on hard reasoning.

Two competitive facts temper the "clear lead" narrative. First, leads at this frontier are measured in months, not years: each of the three major labs has traded the top position repeatedly, and a launch-day ranking is a snapshot, not a settled order. Second, price and openness matter as much as raw capability. Coverage noted that far cheaper models, including open-weight releases from Chinese labs, were closing the gap on many practical tasks, which constrains how much a frontier lead is worth commercially. Coverage of the launch captured the mood well: state-of-the-art reasoning, yes, but inside an intensely contested market where no lead is durable. Placing Gemini 3 correctly means holding it as one strong entry among several, not as a decisive victory.

The open-weight pressure point deserves more than a passing mention, because it changes what a benchmark lead is worth in practice. When a frontier-priced, closed model leads a cheaper, openly available one by a few points on a reasoning benchmark, many buyers reasonably choose the cheaper option anyway, because the marginal reasoning gain does not justify the marginal cost for their workload. That dynamic does not erase the value of leading a benchmark table, since some tasks genuinely need the extra capability and will pay for it, but it does mean market share and benchmark leadership are no longer the same claim, and a report that treats them as interchangeable will misread how the competitive field actually moves.

Documented limits and open questions

Several honest limits follow directly from what the announcement did and did not say. The grounding record for this event notes that different surfaces, model variants, and access tiers expose different capabilities, and that adoption and benchmark claims remain separable and should be reviewed independently. Both cautions are well founded. The Gemini 3 a free Search user encounters, the Pro model a developer calls through the API, and the Deep Think mode reserved for Ultra subscribers are not the same system, and quoting a Deep Think benchmark as though it describes the default product overstates what most users receive.

The evidence base is also young. At launch, nearly every number originated with Google, and the most impressive results depended on favorable configurations: tool access, code execution, or the higher-compute mode. The reliance on human-preference arenas like LMArena introduces a known confound, since preference is not accuracy. Contamination risk is real given the overlap between the training cutoff and public benchmark releases. None of this implies the model is weak; it implies that confident ranking requires outside verification that takes time to arrive. The disciplined reading is that Gemini 3 is a genuinely strong model whose precise standing was not yet independently settled on the day it shipped.

A further open question concerns safety evaluation under the higher-compute Deep Think configuration. Restricting that mode to safety testers before a wider release is a reasonable precaution, since a model that reasons for longer before answering can, in principle, find its way to outputs that a faster, shallower pass would not have produced, whether useful or harmful. The public record at launch does not include a detailed account of what that safety testing found, only that it occurred, which leaves a real gap between the confidence the launch materials project and what an outside reader can independently confirm about the highest-capability configuration of the model.

Placing Gemini 3 in the history of AI

Gemini 3 is best placed as a consolidation, not a rupture. Its architecture, sparse mixture-of-experts with native multimodality, was established across the earlier Gemini reports. Its explicit reasoning behavior was introduced with Gemini 2.5. Its one-million-token context traces to Gemini 1.5. Version three combined and extended these under one training effort and reported strong results, which is exactly what a mature model line is expected to do. Treating it as an incremental peak on a documented curve is more accurate than treating it as a founding moment.

If Gemini 3 earns a distinct line in an AI timeline, the reason is likely distribution rather than architecture. Putting a frontier reasoning model into the default search path from day one, and pairing it with generative interfaces and an agentic development platform, signals a phase in which model releases are inseparable from the products that carry them to billions of people. The historically useful questions to carry forward are therefore three. Which inherited component did each Gemini version advance? Which capability was genuinely new here versus repackaged from a predecessor? And which of the launch claims survived independent testing over the following months? Answering those keeps the timeline honest, separating the dated, verifiable event, a model shipped on 18 November 2025, from the larger and slower judgment of what it meant.

That separation is the discipline this kind of report has to hold onto as later models arrive and inevitably claim to surpass Gemini 3 on some subset of these same benchmarks. The dated facts, what was announced, what was measured by whom, and what shipped where, do not change with hindsight. The interpretive judgment, how significant the release turns out to have been, is properly revised as independent evidence accumulates, and a reader returning to this record months later should expect the second layer to have moved while the first stays fixed.

References

• Google (2025). A new era of intelligence with Gemini 3.
• Google (2025). Gemini 3 for developers: New reasoning, agentic capabilities.
• Google (2025). Gemini 3 developer guide.
• Google (2025). Google brings Gemini 3 to Search and AI Mode.
• Google (2025). See new Gemini app updates with the Gemini 3 AI model.
• Google Research (2025). Generative UI: A rich, custom, visual interactive user experience for any prompt.
• Google (2025). Google Antigravity.
• Google (2025). Gemini 2.5: Our most intelligent AI model.
• Gemini Team, Google (2024). Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.
• Gemini Team, Google (2023). Gemini: A Family of Highly Capable Multimodal Models.
• Phan et al. (2025). Humanity's Last Exam.
• Rein et al. (2023). GPQA: A Graduate-Level Google-Proof Q&A Benchmark.
• Chiang et al. (2024). Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference.
• Fedus, Zoph and Shazeer (2021). Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity.
• Franzen (2025). Google unveils Gemini 3 claiming the lead in math, science, multimodal and agentic AI benchmarks.
• CNBC (2025). Google announces Gemini 3 as battle with OpenAI intensifies.