Google launched two fast Gemini models this week. They arrived with similar names, different jobs and one useful question hanging over both:
What does “fast” mean when the clock is actually running?
Gemini 3.6 Flash is Google's new general-purpose workhorse. Gemini 3.5 Flash-Lite is the smaller, cheaper model intended for high-volume, latency-sensitive work. On Google's benchmarks, both improve substantially over their predecessors.
On BulletBench, the difference between them is easier to see.
Gemini 3.6 Flash is the stronger model when it has a full minute to work. Gemini 3.5 Flash-Lite is the cleaner fit when every response needs to land in about a second.
The result in one table
| Model | Clock | Ladder Elo | Median move | p90 move | Time losses |
|---|---|---|---|---|---|
| Gemini 3.6 Flash (minimal) | Bullet, 60s | 808 | 1.178s | 1.397s | 21/96 |
| Gemini 3.5 Flash (minimal) | Bullet, 60s | 808 | 1.179s | 1.481s | 26/96 |
| Gemini 3.5 Flash-Lite (minimal) | Lightning, 10+1s | 776 | 0.958s | 1.051s | 0/96 |
| Gemini 3.5 Flash-Lite (low) | Lightning, 10+1s | 749 | 0.966s | 1.112s | 0/96 |
Gemini 3.6 Flash and Gemini 3.5 Flash finish level on rating and score at Bullet: 808 Elo and 55.7%. The new model takes first place under BulletBench's published tie-break order because it loses five fewer games on time, then edges the latency comparison by one millisecond at the median. Its p90 is also 84ms faster.
That is not a sweeping generational victory. It is a narrow but useful one. Under an actual deadline, slightly more of Gemini 3.6 Flash's capability reaches the board in time.
Flash-Lite tells a different story. At minimal reasoning it finishes second at Lightning, only five Elo behind Gemini 3.5 Flash, while answering in a median 0.958 seconds and losing zero of 96 games on time. Its p90 is just 1.051 seconds.
That tail-latency result may matter more than the five-point rating gap.
Gemini 3.6 Flash is a one-minute model
Google describes Gemini 3.6 Flash as its new workhorse for coding, knowledge, multimodal understanding and agentic tasks. The company reports a 49% result on DeepSWE, 63.9% on MLE-Bench and 83% on OSWorld. Google also says it uses 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Intelligence Index, with reductions as high as 65% on some tasks.
The commercial proposition is straightforward: more capability, less meandering and a lower output price. Gemini 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens. The previous Gemini 3.5 Flash output price was $9.
BulletBench supports the “workhorse” positioning, with a qualification.
At Bullet, Gemini 3.6 Flash on minimal reasoning sits at the top of the full 59-configuration roster. It scores 55.7%, reaches 808 ladder Elo and keeps its p90 move time below 1.4 seconds. That is an unusually strong balance of decision quality and response time.
At Lightning, it falls to fifth. The shorter clock magnifies every slow move, and 26 of its 96 games end on time. Its 694 rating trails both minimal and low Flash-Lite, the previous Gemini 3.5 Flash and Gemini 3.1 Flash Lite.
So Gemini 3.6 Flash is fast, but not universally latency-proof. Give it a minute and it is the current BulletBench leader. Ask it to sustain a decision almost every second and the lighter models catch it.
Flash-Lite's real feature is predictability
Google calls Gemini 3.5 Flash-Lite the fastest model in the Gemini 3.5 family and cites roughly 350 output tokens per second. It costs $0.30 per million input tokens and $2.50 per million output tokens - one fifth and one third of Gemini 3.6 Flash's respective prices.
Those economics make it an obvious candidate for search agents, document pipelines, classification, routing and other workloads where one request becomes thousands.
The BulletBench result adds something token throughput alone cannot show: sequential predictability.
At Lightning, minimal Flash-Lite posts:
- 0.958-second median move latency
- 1.051-second p90 latency
- zero time losses in 96 games
- 776 ladder Elo, second overall
Low reasoning is only slightly slower: a 0.966-second median, 1.112-second p90 and again zero time losses. It gives up 27 rating points, but remains third on the Lightning board.
This is the shape product teams want in an interactive dependency. The model is not merely fast on average. Its slow tail stays close to its median, so one late response is less likely to stall the next step in a chain.
At Bullet, however, minimal Flash-Lite drops to 672 Elo. Low reasoning does better at 716, but still trails Gemini 3.6 Flash by 92 points. Flash-Lite is the sharper latency tool; it is not a smaller substitute for every Flash workload.
More reasoning made every new model worse
The launch documentation gives both models adjustable thinking levels. Google's current model guide says Gemini 3.6 Flash defaults to medium thinking, while Gemini 3.5 Flash-Lite defaults to minimal. For complex agentic work, Google recommends increasing Flash-Lite to medium or high.
That may be sensible when the task has time to wait. It is disastrous under these clocks.
Gemini 3.6 Flash falls from 808 Bullet Elo at minimal reasoning to 694 at low and 513 at both medium and high. At Lightning, medium and high reasoning lose 95 of 96 games on time.
Flash-Lite behaves similarly. Its minimal configuration leads its family at Lightning. Medium loses 95 of 96 Lightning games on time; high loses all 96.
This does not mean additional reasoning makes the models less intelligent. It means the extra thinking consumes the resource BulletBench makes scarce. A stronger move that arrives after the clock expires scores exactly like any other loss.
The practical lesson is to treat thinking level as part of the latency budget, not only as a quality control. Google's defaults are general-purpose defaults. A real-time route should be benchmarked at the actual deadline and with the actual reasoning setting it will use in production.
Which model should you choose?
For a hard sub-second or one-second dependency, start with Gemini 3.5 Flash-Lite at minimal reasoning. It gives up some decision quality, but its latency distribution is exceptionally tight and its price supports high-volume use.
For a workflow that can tolerate roughly 1.2 seconds per model turn and benefits from stronger judgement, Gemini 3.6 Flash at minimal reasoning is the more capable option. It currently leads BulletBench's 60-second format, and Google's lower output-token use may further improve cost and end-to-end latency on longer responses.
For complex coding or agent work without a hard interactive deadline, BulletBench cannot tell you whether medium or high thinking is worthwhile. Chess under a clock is deliberately narrow. Use the broader task benchmarks - and test on your own prompts - before making that decision.
The short version:
- Gemini 3.6 Flash: the better balance when capability comes first but latency still matters
- Gemini 3.5 Flash-Lite: the better infrastructure primitive when cost, volume and predictable response time come first
- Minimal reasoning: the clear fast-format winner for both families
Two launches, two definitions of fast
Model launches tend to compress speed into a single number. Time to first token, output tokens per second and median latency all describe something real, but none captures the complete product experience.
Gemini 3.6 Flash is fast because it delivers more useful capability inside a finite budget. Gemini 3.5 Flash-Lite is fast because it answers with very little variance, at a price that makes repetition practical.
BulletBench makes the distinction visible. One wins the minute. The other wins the second.
Explore the complete 11,328-game fast-format leaderboard, confidence intervals and latency results on BulletBench. For release details and API migration guidance, see Google's launch announcement and Gemini model documentation.
