Two models, one budget, completely different machines. This is the rare comparison where you can ignore the rate card — a few cents per million tokens separates them in either direction depending on your input/output ratio. Everything that matters here is in the engineering columns.
Side-by-side specs
| Dimension | GPT-5 | Gemini 3.5 Flash |
|---|---|---|
| Input / 1M | $1.25 | $1.50 |
| Output / 1M | $10.00 | $9.00 |
| Cached input / 1M | $0.125 | $0.15 |
| Free API tier | No | Yes |
| Context window | 400,000 | 1,048,576 |
| Max output | 128,000 | 65,536 |
Latency needs a deployment test
Neither provider publishes one universal, directly comparable tokens-per-second or time-to-first-token guarantee for these two models, and benchr has not run a controlled cross-region study. The undocumented values formerly shown here were removed. Benchmark the same prompts, output lengths, settings, and concurrency in your target deployment and compare p50, p95, and p99 latency.
Reasons to evaluate GPT-5
GPT-5's listed 128K max output is larger than Flash's listed 65,536-token cap, which can matter for long code generations or documents. For coding, OpenAI publishes a SWE-bench figure for GPT-5, while the checked Google material does not publish a directly comparable SWE-bench Verified figure for Gemini 3.5 Flash. Run both on the same repository tasks and settings rather than filling the gap with an estimate.
Free-tier availability
Flash has a free API tier; GPT-5 doesn't. For prototyping, internal tools, and low-volume side projects, that's not a rounding error — it's the entire bill. Plenty of teams should prototype on Flash's free tier, measure, and only then decide where paid traffic goes. The free-AI roundup maps where the free tier taps out.