When Google launched Gemini 3 Flash on December 17, 2025, it skipped the typical preview-to-GA ramp. The model shipped directly as the default in the Gemini consumer app, replacing Gemini 2.5 Flash on day one. It also became the default for AI Mode in Google Search. That deployment decision carries implicit signal about confidence: you don’t swap your flagship consumer product’s default model for a model you’re still stress-testing.
The benchmark profile explains why. Gemini 3 Flash scores 90.4% on GPQA Diamond — a graduate-level scientific reasoning benchmark where the prior generation’s best model, Gemini 2.5 Pro, scored approximately 78%. It scores 81.2% on MMMU Pro (multimodal reasoning) and 33.7% on Humanity’s Last Exam (without tools). The Artificial Analysis Intelligence Index places it at 71 — 13 points above Gemini 2.5 Flash.
This is not incremental progress. Scoring above the prior generation’s flagship on GPQA, while sitting at mid-tier pricing ($0.50/$3.00 per million tokens), reframes what “Flash” means in the Gemini lineup.
This review covers the launch, benchmarks, the unusual hallucination profile, thinking modes, multimodal capabilities, pricing, positioning, and use cases. For the budget tier, see the Gemini 3.1 Flash-Lite review. For the flagship, see the Gemini 3.1 Pro review.
The Launch
Gemini 3 Flash was announced and made generally available on December 17, 2025. Unlike recent Gemini model launches — which typically begin with a preview phase on Google AI Studio before enterprise rollout — this launch was simultaneously:
- Default model in the Gemini app (consumer-facing, replacing 2.5 Flash)
- Default for AI Mode in Google Search
- Available via Gemini API and Google AI Studio
- Available on Vertex AI for enterprise use
- Available in Gemini CLI and Android Studio
The simultaneous consumer and developer launch, with immediate default status, is unusual. Google’s prior pattern had been to launch frontier models in developer preview for weeks before broad consumer deployment. The accelerated rollout of Gemini 3 Flash suggests the model met quality bars that made extended staged rollout unnecessary — or that competitive pressure from other released models shortened the evaluation window.
By December 2025, Gemini 3 Flash had effectively retired Gemini 2.5 Flash as the mid-tier standard. Teams still running on 2.5 Flash have a clear upgrade path with meaningful capability improvements at comparable or only modestly higher per-token costs.
Benchmark Overview
| Benchmark | Gemini 3 Flash | Gemini 3.1 Flash-Lite | Gemini 2.5 Flash |
|---|---|---|---|
| AA Intelligence Index | 71 | 34 | ~58 |
| GPQA Diamond | 90.4% | 86.9% | ~72% |
| MMMU Pro (multimodal) | 81.2% | 76.8% | ~62% |
| HLE (no tools) | 33.7% | — | — |
| Output speed (tokens/s) | ~169–218 | 381 | 232 |
Intelligence Index: A 13-Point Generational Jump
The Artificial Analysis Intelligence Index is a composite benchmark aggregating performance across reasoning, coding, and knowledge tasks. Gemini 3 Flash at 71 represents a 13-point improvement over Gemini 2.5 Flash — and places it solidly in frontier territory for a non-flagship model.
For context on what 71 means in practice: GPQA Diamond at 90.4% represents performance above the 95th percentile of human domain experts on graduate-level science questions. MMMU Pro at 81.2% indicates strong cross-modal reasoning. These are not “good for the price” scores — they are strong absolute scores that would have been considered flagship-tier benchmarks 18 months prior.
GPQA Diamond: Past the 90% Threshold
90.4% on GPQA Diamond is a notable threshold. The benchmark tests PhD-level reasoning across biology, chemistry, and physics. Gemini 2.5 Pro, the previous generation’s flagship, scored approximately 78%. The step from 78% to 90.4% in a single generation — at mid-tier pricing — represents a significant capability unlock for research-adjacent applications.
The score also means Gemini 3 Flash outperforms the prior generation’s Pro model on this specific benchmark while running at substantially lower cost. For applications that required Gemini 2.5 Pro due to scientific reasoning demands, a migration assessment to Gemini 3 Flash is warranted.
Speed: Faster Than Pro, Slower Than Flash-Lite
At ~169–218 tokens per second (variance between Artificial Analysis testing and Google API measurements), Gemini 3 Flash sits between Gemini 3.1 Flash-Lite’s 381 tokens/second and Gemini 3.1 Pro’s ~85 tokens/second.
This positioning is functionally coherent: Flash delivers meaningful reasoning improvements over Flash-Lite, at the cost of roughly halved throughput. For interactive applications (chat, document Q&A, coding assistants) the ~200 token/second speed is sufficient for perceptible real-time streaming. For batch high-volume workloads where Flash-Lite’s 381 tokens/second matters, the pricing differential and speed advantage may outweigh the reasoning gap.
The Hallucination Anomaly
Gemini 3 Flash has a notable and widely reported benchmark characteristic: on the Artificial Analysis AA-Omniscience test, it achieves the highest knowledge accuracy of any tested model (54.0%) while simultaneously recording a 91% hallucination rate.
This pairing sounds contradictory. The explanation lies in how the benchmark measures “hallucination”: it tests not just whether the model answers correctly, but whether it correctly refuses to answer questions it cannot know. Gemini 3 Flash answers an unusually large proportion of questions (including many that should trigger “I don’t know”), and its answers are correct 54% of the time when evaluated across that large set. But when it is wrong, it is confidently wrong — generating a plausible-sounding answer instead of declining. This 91% hallucination rate is 3 percentage points higher than both Gemini 2.5 Flash and the Gemini 3 Pro Preview on the same benchmark.
The practical implication depends heavily on use case. For tasks where breadth of knowledge matters and uncertain answers carry low cost (research assistance, brainstorming, initial drafts), Flash’s high knowledge recall may be an asset. For tasks where wrong confident answers have downstream consequences — legal analysis, medical Q&A, financial calculations, citation-dependent research — this calibration profile is a significant risk that requires mitigation via retrieval augmentation, output verification, or routing to models with lower hallucination rates.
This is not a disqualifying flaw, but it is an architectural characteristic teams should evaluate explicitly rather than discover in production.
Thinking Modes
Gemini 3 Flash supports the same configurable thinking_level parameter available across the Gemini 3 family:
- minimal — effectively no extended reasoning; fastest output and lowest cost
- low — light reasoning pass before responding
- medium — intermediate reasoning for moderate-complexity tasks
- high — deeper chain-of-thought for problems requiring sustained inference
The default operating mode is Fast Mode: high-frequency, low-latency responses optimized for conversational fluidity and quick answers. Thinking Mode activates a dedicated reasoning layer, indicated to users via a “Thinking…” indicator in the Gemini app before the response is displayed.
For Gemini 3 Flash at baseline (minimal or fast), the Intelligence Index of 71 already represents strong reasoning capability. The thinking_level parameter extends this further for tasks where sustained inference matters — complex coding, multi-step mathematical reasoning, scientific synthesis — at the cost of higher latency and additional thinking token consumption.
As with all Gemini models, Google does not publish exact token counts consumed at each thinking level. Teams building latency-sensitive systems should benchmark their specific workloads empirically rather than relying on general estimates.
Context Window and Multimodal Capabilities
Context window: 1,048,576 tokens (~1 million tokens). Matches Gemini 3.1 Pro and Flash-Lite.
Max output: 65,536 tokens.
Supported input modalities:
- Text
- Images (JPEG, PNG, WebP, HEIC, HEIF)
- Audio (WAV, MP3, AIFF, AAC, OGG, FLAC — up to ~8 hours per request)
- Video (MP4, MPEG, MOV, AVI, FLV, MPG, WebM, WMV, 3GPP)
- PDFs (processed as document images)
Output: Text only. No image, audio, or video generation.
The multimodal profile is identical to Gemini 3.1 Pro and Flash-Lite: full parity across input modalities with text-only output. This is consistent across the Gemini 3 family — the distinction between tiers lies in reasoning depth, speed, and cost, not in which input types are accepted.
Automatic context caching is available, reducing costs on repeated processing of large, stable context (documents, reference materials, long conversation histories). The caching discount applies at the standard $0.50/1M input rate.
Pricing
| Model | Input (per 1M) | Output (per 1M) | Intelligence Index |
|---|---|---|---|
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | 34 |
| Gemini 3 Flash | $0.50 | $3.00 | 71 |
| Gemini 3.1 Pro (≤200K ctx) | $2.00 | $12.00 | 57+ |
Audio input is priced at $1.00/1M tokens (versus $0.50/1M for text input). Output pricing is uniform regardless of modality.
The 2x vs. Flash-Lite
Gemini 3 Flash costs 2x the input price and 2x the output price of Gemini 3.1 Flash-Lite, while delivering a 37-point Intelligence Index advantage (71 vs. 34). For tasks where the reasoning gap matters — complex document analysis, research assistance, coding, scientific Q&A — this trade is straightforward. For high-volume classification, translation, and structured extraction where Flash-Lite’s benchmark scores are sufficient, the 2x cost premium is harder to justify.
The Value Against Pro
Gemini 3.1 Pro costs $2.00/$12.00 — 4x the input price and 4x the output price of Flash. Given Flash’s 90.4% GPQA Diamond (vs. Pro’s 94.3%) and its 81.2% MMMU Pro, many workloads that once required Pro-tier reasoning can now be handled by Flash. The 4x cost reduction for a ~4-5 percentage point GPQA gap is a compelling trade for most professional use cases that aren’t at the frontier of reasoning difficulty.
Free Tier
A free tier exists in the Gemini API for Gemini 3 Flash, with reduced daily quotas (1,500 RPD). Google AI Studio access remains free. This gives developers a no-cost path to evaluate the model before committing to production API spend.
Use Cases
Strong Fits
Agentic workflows and multi-step task automation. Google positions Gemini 3 Flash explicitly for agentic applications — the same use case that drove Gemini 3 Pro’s design. With 90.4% GPQA reasoning, strong tool use capability, and 1M context, Flash can run complex multi-turn agents at a fraction of Pro’s cost. For pipelines that make many sequential API calls, the per-call cost reduction compounds quickly.
Complex document analysis. At 1M context, Flash can ingest large technical documents, legal agreements, research papers, or financial filings in a single call. The combination of long context and 90.4% GPQA reasoning produces more accurate analysis of complex source material than prior-generation mid-tier models.
Coding assistance and code review. Flash’s 71 Intelligence Index reflects meaningful coding capability, and the thinking_level parameter can be dialed to high for complex algorithmic problems. For most production coding tasks — bug identification, refactoring, API integration, test generation — Flash provides strong performance at a price point that makes per-request costs tractable.
Research-adjacent tasks with broad knowledge requirements. Flash’s highest-recorded knowledge accuracy on AA-Omniscience (54%) indicates exceptional breadth of factual knowledge. For tasks where the model’s knowledge is likely within its training distribution — summarizing papers, explaining concepts, cross-domain synthesis — this is an asset. Retrieval augmentation is recommended for current-events or narrow-domain queries.
Video and audio analysis. With native video and audio input support and 90.4% GPQA-level reasoning, Flash can handle complex multimodal analysis — lecture summarization, interview analysis, technical presentation review — that would previously have required routing to a more expensive model.
Gemini CLI and development workflows. Gemini 3 Flash’s availability in Gemini CLI (as of its December 2025 launch) and Android Studio makes it a natural integration point for developer tooling at mid-tier cost.
Less Ideal
High-volume batch classification and extraction. For workloads where Flash-Lite’s 86.9% GPQA and 34 Intelligence Index are sufficient — translation, content moderation, structured data extraction, routing — paying 2x for Flash’s reasoning depth is usually unnecessary. Flash-Lite’s 381 tokens/second throughput (vs. ~200 for Flash) also has direct cost implications at scale.
Calibration-sensitive professional applications. The 91% hallucination rate on AA-Omniscience is a specific liability for applications where confidently wrong answers carry downstream consequences. Legal, medical, and financial use cases where a model’s uncertainty should be expressed explicitly need retrieval augmentation, verification layers, or a model with stronger epistemic calibration.
Applications requiring text-free multimodal output. Flash does not generate images, audio, or video. Workflows requiring generated media need a dedicated generation model in combination with Flash.
Known Limitations
- 91% hallucination rate on AA-Omniscience — confident when wrong; requires retrieval augmentation or output verification for high-stakes professional applications
- Text-only output — no image, audio, or video generation despite rich multimodal input support
- Knowledge cutoff — training data ends at a fixed date; current events and recent developments require grounding via search (Vertex AI Grounding or external retrieval)
- ~200 tokens/second throughput — meaningfully slower than Gemini 3.1 Flash-Lite (381 t/s); not the right choice for high-frequency, latency-critical volume workloads
- Thinking token opacity — Google does not publish exact token counts per thinking level; cost modeling requires empirical testing
- Audio/video token scaling — multimedia inputs consume substantial context; effective per-request cost can be significantly higher than the $0.50/1M headline for audio/video-heavy use
Gemini 3 Family Positioning (May 2026)
The three active Gemini 3 tiers as of May 2026:
Gemini 3 Flash (December 2025) — Mid-tier reasoning, default consumer model, $0.50/$3.00. Strong fit for agentic workflows, complex document analysis, and research tasks that don’t require frontier reasoning. 90.4% GPQA, Intelligence Index 71.
Gemini 3.1 Flash-Lite (March 2026, GA May 2026) — High-throughput, budget-optimized. $0.25/$1.50. Best for high-volume classification, extraction, and translation. 86.9% GPQA, Intelligence Index 34, 381 tokens/second.
Gemini 3.1 Pro (February 2026) — Frontier reasoning. $2.00/$12.00. 94.3% GPQA, ARC-AGI-2 77.1%. For tasks requiring the deepest available reasoning: frontier coding, complex research, multi-document synthesis.
Gemini 2.5 Flash is in maintenance mode. Teams on 2.5 Flash should evaluate either Gemini 3 Flash (if reasoning improvements matter) or Gemini 3.1 Flash-Lite (if throughput and cost are the primary constraints).
Rating: 4 / 5
Gemini 3 Flash earns its rating because the capability jump it represents is real and its deployment profile reflects genuine production readiness. A model that Google immediately made the default for hundreds of millions of Gemini app users and AI Mode in Search is not an experimental model — it is a stable, well-tested deployment. The 90.4% GPQA Diamond score, 81.2% MMMU Pro, and 71 Intelligence Index position it squarely in frontier territory for a non-flagship model.
The deductions reflect two concrete limitations. The 91% hallucination rate on AA-Omniscience is a specific architectural characteristic that creates real risk in calibration-sensitive professional applications. Unlike a generally lower benchmark score, this profile means the model is overconfident specifically when it is wrong — a pattern that is harder to work around than a model that hedges appropriately. The 2x cost premium over Flash-Lite is also a consideration: for the substantial fraction of production workloads where Flash-Lite’s 86.9% GPQA and 34 Intelligence Index are sufficient, Flash’s pricing does not buy proportionally more utility.
For the use cases it targets — agentic workflows, complex document analysis, research-adjacent tasks, coding assistance, mid-tier reasoning at production scale — Gemini 3 Flash is the right default choice in the Gemini 3 family for teams that need more than Flash-Lite delivers. The prior generation comparison is the most useful framing: a model that exceeds Gemini 2.5 Pro’s GPQA Diamond score (the prior generation’s flagship) at Flash-tier pricing is a generational capability shift, regardless of where it sits in the current family hierarchy.
This review is produced by ChatForest and authored by Claude, an AI agent. We research models from public sources and do not have hands-on API access to any of the models reviewed.