benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input (≤200K) | $2.00 |
| Standard input (>200K) | $4.00 |
| Standard output | $12.00 |
| Context caching | $0.20 |
| Context window | 1,000,000 tokens |
The 1M context window — practical applications
One million tokens is approximately 750 pages of text, 150–200 images, or 2–3 hours of video transcript. This enables workflows that are impractical or impossible with smaller context models: loading an entire 50K-line codebase into context for architectural review; processing a full legal contract package; analyzing a multi-hour video recording in a single call; synthesizing findings across 50 research papers simultaneously. For any task where chunking introduces coherence loss, Gemini 3.1 Pro's 1M window removes the constraint.
Note the tiered pricing: inputs over 200K tokens cost $4/1M rather than $2/1M. For workloads that regularly use 500K–1M tokens of context, the effective input cost per call is higher — budget accordingly.
GPQA Diamond: 94.3% on PhD-level science
GPQA Diamond tests PhD-level questions in biology, chemistry, and physics — questions that require domain expertise beyond pattern matching. Gemini 3.1 Pro's 94.3% score leads the benchmark among major commercial models, ahead of Claude Opus 4.8 at 93.6% and GPT-5.5. For research assistance, scientific analysis, medical reasoning, and technical document understanding, this performance level is directly relevant to task quality. For general-purpose coding or business tasks, GPQA is less discriminating than SWE-bench.
Native multimodal: video, audio, images, PDF
Gemini 3.1 Pro accepts text, images, video, audio, and PDF in the same API call. That makes it a candidate for workflows such as reviewing a technical presentation with slides and audio together. The feature list does not establish task quality; test the exact media mix, file limits, latency, and output requirements.
Cost scenarios
At 10M input + 2M output per month within 200K context tiers: $20 + $24 = $44/month. Claude Opus 4.8 at the same volume: $50 + $50 = $100/month — Gemini is 56% cheaper. For long-context use (averaging 500K tokens per call): 10M total input at $4/1M blended = $40 + $24 output = $64/month — still 36% below Opus 4.8 at the same output volume.
Use-case fit
Consider for: Long-document analysis approaching the 1M-token context maximum; scientific and technical tasks where GPQA is relevant; multimodal pipelines combining text, video, audio, and images; and Google Cloud or Vertex AI stacks.
Skip if: Your primary need is SWE-bench coding performance — Gemini 3.1 Pro's coding benchmark trails Claude Opus 4.8 and GPT-5.5. Also skip for simple volume tasks where the $2/1M base rate is higher than necessary — Gemini 3.5 Flash at $1.50/1M handles lighter workloads.
Decision checklist
Identify your actual context length requirements: if your p90 context is under 200K tokens, the 1M window is not a differentiator. Gemini 3.5 Flash at $1.50/1M covers most tasks at lower cost. Use Gemini 3.1 Pro specifically when the context ceiling or GPQA performance gap is the bottleneck.
For multimodal use: confirm whether your input modalities (video, audio) benefit from the native architecture. For simple image captioning or document OCR, lighter models are sufficient.