Gemini 3.5 Flash, released May 20, 2026 at Google I/O, is the first Flash-tier model to beat Gemini 3.1 Pro on agentic and coding benchmarks — at roughly half the cost. It is the default production tier for developer applications and is rolling out to billions of users across Google Search, Maps, Gmail, Docs, YouTube, and Chrome.
Why It's in Trial
Gemini 3.5 Flash changes the Flash-vs-Pro tradeoff. On the benchmarks that matter for agentic engineering — multi-step execution, tool use, reasoning — it outperforms the current flagship Gemini 3.1 Pro:
- Terminal-Bench 2.1: 76.2% (vs Gemini 3.1 Pro's 70.3%)
- MCP Atlas: 83.6% (vs 78.2%)
- Finance Agent v2: 57.9% (vs 43.0%)
- GDPval-AA: 1656 Elo (vs 1314)
- CharXiv (multimodal): 84.2%
- SWE-bench Verified: 55.1% — trails Gemini 3.1 Pro's 76.2% on isolated code editing tasks
This is a notable inversion: for agentic pipelines, 3.5 Flash is now the stronger choice. Gemini 3.1 Pro retains an advantage on pure code editing tasks (SWE-bench). Gemini 3.5 Pro was expected in June 2026 but has not reached GA as of July 20, 2026; the July 17 limited AI Studio whitelist preview window passed without broader rollout; secondary sources report July 22–28 as the next broader GA target, unconfirmed by Google.
Benchmark sources: Terminal-Bench 2.1, MCP Atlas, GDPval-AA, Finance Agent v2, and CharXiv from the Google Cloud Blog primary source; SWE-bench Verified (55.1%) from multiple independent aggregators (llm-stats.com, Artificial Analysis).
Where It Stands Out
- Agentic benchmarks: Leads Gemini 3.1 Pro on all multi-step, tool-using evaluations
- Cost efficiency: Less than half the cost of comparable frontier models (Google Cloud Blog); query latency under 200ms for most workloads
- Universal deployment: Running in Google Search, Maps, Gmail, Docs, YouTube, and Chrome — rare at a Flash tier, and a strong signal of production readiness
- GitHub Copilot: Available as a model option from May 19, 2026
- Antigravity: Default model in Google's agentic coding CLI (Antigravity) as of I/O 2026
Tradeoffs
- SWE-bench trails 3.1 Pro: For complex code editing or refactoring, Gemini 3.1 Pro remains stronger
- No free tier: Requires paid Google AI plans (Google AI Plus/Pro, ~$20/mo); the free tier remains on Gemini 3 Flash
- New benchmark suite: Terminal-Bench 2.1, MCP Atlas, and Finance Agent v2 are Google-run evaluations — independent replication is limited at this stage
- 3.5 Pro still pending: If evaluating Google's model family as a whole, the Pro variant was expected in June 2026 but has not reached GA as of July 20, 2026 — Google scrapped the original base model for a full rebuild (per Geeky Gadgets, TechTimes); the July 17 AI Studio whitelist preview window passed without broader rollout; secondary sources now report July 22–28 for broader GA — Google has not confirmed any date, model card, or pricing; per Bloomberg (July 16), the original base model was scrapped and rebuilt due to hallucination rate and real-world reliability concerns; watch
cloud.google.com/blogfor the GA announcement
How to Access It
- Google AI Studio — paid tier
- Gemini API (
gemini-3.5-flashmodel ID) - Antigravity — Google's agentic coding CLI (announced at I/O 2026)
- Gemini Enterprise Agent Platform, Gemini Enterprise app
- GitHub Copilot — model selector option
Key Characteristics
| Property | Value |
|---|---|
| License | Proprietary |
| Context window | 1,000,000 tokens |
| Pricing | $1.50 input / $9.00 output per 1M tokens (secondary sources: llm-stats.com, Artificial Analysis; Google pricing docs returned 403 on direct fetch — verify at ai.google.dev) |
| Free tier | No (Gemini 3 Flash is the free tier) |
| Provider | Google DeepMind |
| API model ID | gemini-3.5-flash |