benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $1.50 |
| Standard output | $9.00 |
| Cached input | $0.15 |
| Free tier | Available (rate-limited) |
| Context window | 1,048,576 tokens |
| Max output | 65,536 tokens |
289 tokens per second: what it means in practice
At 289 tok/s, Gemini 3.5 Flash generates a 1,000-token response in approximately 3.5 seconds. GPT-5 at 90 tok/s takes 11 seconds for the same output. Claude Opus 4.8 at 68 tok/s takes 15 seconds. The first-token latency is equally decisive: 195ms for Flash vs 520ms for GPT-5 and 700ms for Opus 4.8.
In a parallel agent framework, this throughput advantage compounds. You can run three Gemini 3.5 Flash agents within the same wall-clock time as one GPT-5 agent, at comparable cost per task. For pipelines that run many short tasks in sequence or execute multi-step reasoning loops, Flash's speed advantage directly reduces total pipeline latency.
Caching: the cheapest cached input rate
Cached input costs $0.15/1M — 90% off the standard rate, matching Anthropic's discount. For an agent with a 100K-token system prompt cached at 90% hit rate: effective input cost drops from $1.50/1M to $0.15 × 0.9 + $1.50 × 0.1 = $0.135 + $0.15 = $0.285 effective per million. At that rate, Gemini 3.5 Flash becomes the cheapest option among the major paid APIs for cache-heavy agentic workloads.
Vision: the highest multimodal score
Gemini 3.5 Flash has a vision capability score of 92 — highest of any model tracked here. It natively supports image, video frame, audio, and PDF document inputs via the Gemini API. For pipelines that process screenshots, product images, scanned documents, or video content, this is the practical default. DeepSeek V4-Pro has no vision. GPT-5 and Opus 4.8 support images but score lower (85–86).
Google Search grounding: native real-time web access
Gemini 3.5 Flash is the only model in this comparison with native Google Search grounding as an API tool. When grounding is enabled, the model retrieves live web results and incorporates them into the response — reducing hallucination on factual queries and eliminating the need for a separate RAG pipeline for current-data questions. This is a unique capability: not approximate web access via function calling, but actual Search integration at the API level.
The free tier
Gemini 3.5 Flash offers a rate-limited free tier: 15 requests per minute, 1 million tokens per minute, 1,500 requests per day. For prototyping, internal tools, and low-volume applications, the free tier eliminates infrastructure cost entirely during development. Most production applications exceed these limits; paid pricing applies from that point.
Cost scenarios
At 10M input + 3M output per month, standard pricing: $15 + $27 = $42/month — essentially the same as GPT-5 ($42.50/month). With 90% cache hit on input: $1.35 + $1.50 = $2.85 input + $27 output = $29.85/month. With caching, Gemini 3.5 Flash is meaningfully cheaper than GPT-5 at equivalent volume — with 3× the output speed and native vision. The output cap (65K vs 128K for other models) is the one constraint to check for your specific workload.
Use-case fit
Best for: High-throughput agent pipelines where latency is the bottleneck; multimodal tasks requiring image, video, or document input; any workload that benefits from Google Search grounding; prototyping via the free tier; parallel agent execution where 3× throughput is a meaningful advantage.
Skip if: You need the deepest single-call coding reasoning — Claude Opus 4.8 (88.6% SWE-bench) is substantially stronger; your pipeline produces outputs longer than 65K tokens (output cap is below the 128K standard); you require strict data residency outside Google's infrastructure.
Decision checklist
Before routing to Gemini 3.5 Flash: verify the 65K max output token limit against your longest expected outputs. For most tasks this is fine; for code generation or long documents, it can be a hard constraint. Measure first-token latency from your deployment region — Google's regional distribution is excellent but varies.
If you're evaluating against GPT-5 on cost: run the comparison with caching enabled. Gemini 3.5 Flash's $0.15/1M cached rate makes it materially cheaper than uncached GPT-5 ($1.25/1M) for cache-heavy workloads.