Historical snapshot, clarified September 9, 2026: This page preserves a September 2025 Gemini 2.0/2.5 comparison. We could not trace the original benchmark sources or test methodology, so its figures and quality rankings are unverified historical claims, not current measurements or deployment recommendations.
Retired model: Google shut down gemini-2.0-flash on June 1, 2026. Do not use it for new work. Check current availability at https://ai.google.dev/gemini-api/docs/deprecations.
For a different model generation, see our Gemini 3.1 Pro, Gemini 3.1 Flash-Lite, and Gemini 3 Flash comparison. That guide retains an April 2026 comparison with a September model-ID correction; it is not a complete ranking of the latest models.
The original comparison considered speed, quality, and cost. Use it to understand those trade-offs, not to select a current model from unverified old numbers.
What is the quick answer for which Gemini model to use?
This is the original ranking, qualified as historical rather than a current recommendation. Throughput and time to first token are different measures of speed.
Fastest, smartest, and best value at a glance
- Highest listed throughput: 2.5 Flash Lite → 2.0 Flash → 2.5 Flash → 2.5 Pro. The 2.0 model is retired.
- Original quality ranking, unverified: 2.5 Pro → 2.5 Flash → 2.0 Flash → 2.5 Flash Lite.
- Original value pick: 2.5 Flash for most apps, not a current best-value finding.
What this means in practice: In the retained table, Flash Lite has higher throughput, but 2.0 Flash has the shorter time to first token (0.53s versus 0.59s). Neither establishes today's fastest model, and 2.0 Flash is no longer available.
How do Gemini models compare on speed, cost, and limits?
This historical table puts latency, throughput, and cost side by side. All numbers, including token limits and prices, are retained as originally published and have not been revalidated as current specifications.
How to read the comparison table
- Latency (seconds) tells you how long users wait for the first token.
- Throughput (tokens per second) matters most for batch jobs and high volume workloads.
- Per-1k token costs make it easy to estimate monthly spend from your expected token usage.
| Model | Latency (s) | Throughput (tps) | Max Output Tokens | Input Cost per 1M | Output Cost per 1M | Input Cost per 1k | Output Cost per 1k |
|---|---|---|---|---|---|---|---|
| Gemini 2.5 Flash Lite | 0.59 | 178.70 | 66,000 | $0.10 | $0.40 | $0.0001 | $0.0004 |
| Gemini 2.0 Flash | 0.53 | 153.10 | 8,000 | $0.15 | $0.60 | $0.00015 | $0.0006 |
| Gemini 2.5 Flash | 0.65 | 97.39 | 66,000 | $0.30 | $2.50 | $0.0003 | $0.0025 |
| Gemini 2.5 Pro | 2.22 | 85.37 | 66,000 | $1.25 to $2.50 | $10 to $15 | $0.00125 to $0.0025 | $0.01 to $0.015 |
What this means in practice: If you run many small requests, focus on latency and per-1k cost. For long, complex tasks, you may accept higher cost and latency from 2.5 Pro to get better answers.
Which Gemini model should I use for my use case?
These are the original September 2025 use-case suggestions, not a current shortlist. The retired 2.0 Flash suggestion is withdrawn.
Customer chat and live tools
Best model: 2.5 Flash Lite.
Use when you need ultra-fast replies for interactive chat or tools.
High volume content drafting
Best model: 2.5 Flash.
Good balance of quality and cost for bulk content and back office workflows.
Snappy UI helpers with short replies
Original suggestion, withdrawn: 2.0 Flash.
Shut down June 1, 2026. Its old latency figure is retained for historical context only; do not use this model.
Advanced analysis and long form reasoning
Best model: 2.5 Pro.
Pick this when answer quality matters more than token cost.
| Use Case | Historical Model Suggestion | Why |
|---|---|---|
| Customer chat and live tools | 2.5 Flash Lite | Ultra-fast response times, cost-effective for high volume |
| High volume content drafting | 2.5 Flash | Good balance of quality and cost for bulk content |
| Snappy UI helpers with short replies | 2.0 Flash (retired) | Original suggestion withdrawn after shutdown; historical latency only |
| Advanced analysis and long form reasoning | 2.5 Pro | Best intelligence for complex tasks |
What this means in practice: Start with the row that looks most like your product. If you later discover a new bottleneck, switch models rather than trying to force one model to fit every task.
Frequently Asked Questions
Which model is the overall value pick?
The original September 2025 article picked Gemini 2.5 Flash. That judgment has not been revalidated and is not a current best-value ranking.
Why does 2.0 Flash still matter?
Only as a historical comparison here. Google shut down gemini-2.0-flash on June 1, 2026, so the old recommendation is withdrawn. Do not use it for new work.
When should I justify 2.5 Pro?
Use 2.5 Pro when accuracy matters more than price. Good examples are finance work, legal-style drafting, or complex reasoning with many steps.
How do I calculate my monthly AI costs?
Divide the token count by 1,000, then multiply by the per-1k rate. Using the historical Flash input rate shown here: 1,000,000 / 1,000 × $0.0003 = $0.30. This is input cost only; calculate output separately and check current pricing before budgeting.
Can I mix different models in the same application?
Yes. Use Flash Lite for simple queries, Flash for standard tasks, and Pro for complex reasoning.
What's the difference between latency and throughput?
Latency = time to first token (UX). Throughput = tokens/sec (batch speed).
How should I choose between Gemini models?
Use three questions to narrow down your choice: Is speed your bottleneck, is answer quality critical, or is cost the main constraint?
A simple rule of thumb for choosing models
- The old table lists 2.5 Flash Lite first for throughput, not for the lowest time to first token.
- Pick 2.5 Flash by default for a balance of cost and quality.
- Reserve 2.5 Pro for workloads where mistakes are very expensive.
What this means in practice: Choose based on your bottleneck: time to first token, output throughput, answer quality, or cost. The historical rankings above do not establish today's fastest, smartest, or best-value model.