Gemini 3.5 Flash vs GPT-5.5 mini
Two fast, low-cost models aimed at high-volume work. We ran both through the same speed, cost, and quality rigs and scored each round on measured results, not on which is "smarter" in the abstract.
Gemini 3.5 Flash takes the overall by a two-point margin, won on throughput and price. GPT-5.5 mini is the better pick when output quality matters more than cost, particularly on reasoning-heavy prompts where it holds an accuracy edge that the speed advantage does not overcome. For chat-volume workloads where latency and per-token price dominate, Flash is the more defensible default.
Both of these models are sold for the same job: fast, cheap responses at high volume. So we scored them the way a buyer for that job would, weighting measured speed, measured price, and measured quality rather than which model is more capable at its respective top end.
Each round below names the exact procedure behind it. Speed and price rounds are pure measurement; the quality rounds are scored on fixed prompts for first-attempt correctness, with re-runs whenever two passes disagree.
| Test category | Winner | Result & method |
|---|---|---|
| Latency (time to first token) | Gemini 3.5 Flash | Flash posted a lower median time-to-first-token across our prompt set. The gap was consistent rather than driven by outliers. How we measured it: Median time-to-first-token measured over 500 requests per model, issued from the same region at the same time of day to keep network and load conditions comparable. |
| Throughput (tokens / second) | Gemini 3.5 Flash | Flash sustained higher output tokens-per-second on long generations. GPT-5.5 mini was steadier on short responses but fell behind once outputs exceeded roughly 800 tokens. How we measured it: Sustained output tokens-per-second measured on long generations (1,000+ token outputs), averaged across the same 500-request set. |
| Price (per million output tokens) | Gemini 3.5 Flash | At list pricing, Flash is cheaper per million output tokens, which compounds in favor of Flash for high-volume deployments. How we measured it: List price per million output tokens from each vendor's published pricing page, normalized to the cost of our observed request mix. |
| Reasoning accuracy | GPT-5.5 mini | On our multi-step reasoning subset, GPT-5.5 mini scored higher on first-attempt correctness. The margin was small but stable across re-runs. How we measured it: A multi-step reasoning subset of fixed prompts scored on first-attempt correctness against a known answer key, re-run when two passes disagreed. |
| Instruction following | GPT-5.5 mini | GPT-5.5 mini adhered more closely to formatting and constraint instructions, with fewer dropped requirements on prompts that stacked several constraints. How we measured it: Prompts that stack several formatting and constraint requirements, scored on the share of stated constraints satisfied in a single response. |
| Long-context recall | Gemini 3.5 Flash | Flash retrieved facts placed deep in long inputs more reliably in our needle-in-context probe, consistent with its larger usable context window. How we measured it: A needle-in-context probe placing facts deep in long inputs, scored on retrieval rate at increasing context lengths. |
Both of these models are sold for the same job: fast, cheap responses at high volume. So we scored them the way a buyer for that job would, weighting measured speed, measured price, and measured quality rather than which model is more capable at its respective top end.
Reading the result
The overall margin is two points, which is narrow enough that the round breakdown matters more than the headline. If your workload is dominated by token volume and you care about cost and latency, Flash wins three of the rounds that bear directly on that. If your workload is reasoning-heavy and a wrong answer is expensive, GPT-5.5 mini’s accuracy and instruction-following edge is the more relevant signal, and the speed gap is unlikely to change the decision.
Devon Mizrahi measures what a model costs to run and how fast it answers. He maintains the price-per-token tables and the latency rigs, and he is the reason the Tracker reports tokens-per-second next to every quality score.