Gemini 3.5 Flash Lands at Intelligence Index 55, #7 Globally — Google's Agent Play Has 1M Context at $1.50/M
Google released Gemini 3.5 Flash on May 19, and the positioning is explicit: this is a model built for agents, not conversations. TechCrunch’s framing from I/O 2026 — “Google bets its next AI wave on agents, not chatbots” — tracks with the benchmarks and pricing.
The Numbers
Gemini 3.5 Flash scores 55 on the Artificial Analysis Intelligence Index v4.0, placing it #7 out of 147 tracked models. The AA index incorporates 10 evaluations including GDPval-AA, τ²-Bench Telecom, Terminal-Bench Hard, SciCode, Humanity’s Last Exam, and GPQA Diamond. A score of 55 puts it well above the average of 36 for comparable models, and above Gemini 3.1 Flash Lite (released last month at $0.25/M) by a significant margin.
The model was added to the Chatbot Arena Text and Code leaderboards on the same day as release.
Pricing
| Metric | Gemini 3.5 Flash |
|---|---|
| Input | $1.50/M tokens |
| Output | $9.00/M tokens |
| Context | 1M tokens |
| Modalities | Text, image, speech, video |
For comparison: Claude Sonnet 4.6 runs $3/$15 per million; GPT-5.4 runs $10/$30. At $1.50 input, Gemini 3.5 Flash undercuts Sonnet by 50% on input cost while outscoring Sonnet’s AA Intelligence Index by several points. GPT-5.4 mini (sub-frontier tier) runs around $0.15/$0.60 — Gemini 3.5 Flash is more expensive but significantly more capable.
Speed
Google claims 4x faster throughput than rival frontier models at similar quality levels. The AA leaderboard confirms it as notably fast, though verbose — it generated 73M tokens during Intelligence Index evaluation versus an average of 36M for models in its class. Token verbosity is a cost multiplier in agent loops; developers building tight pipelines will want to account for it.
The Agent Framing
Google’s I/O pitch for Gemini 3.5 Flash was agent-first. The model is the compute substrate for Antigravity 2.0’s managed agents, for Gemini API Managed Execution, and for the new Android CLI developer tools announced at the same event. Rather than pitching raw benchmark leadership, Google is positioning this as the throughput engine for multi-step agentic loops where latency matters more than maximal reasoning depth.
The τ²-Bench Telecom component in the AA index — which tests autonomous workflow completion — is one of 10 weighted evaluations. Google has not published a standalone SWE-bench Verified score for Gemini 3.5 Flash at launch.
Competitive Position
At Intelligence Index 55, Gemini 3.5 Flash sits between Gemini 3.1 Pro Preview (55-57 tier) and the frontier ceiling held by GPT-5.5 xhigh (60). It outscores Kimi K2.6 (54), MiniMax M2.7 (57, per GDPval-AA), and every Chinese open-weight model currently tracked.
For workloads prioritizing throughput and cost over maximum reasoning capability, Gemini 3.5 Flash now sits at the top of its price tier. The question is whether the 73M-token verbosity issue in evaluation translates to production use — if it does, the $9/M output cost will erode the input-side savings on long-running agent tasks.
Key Specifications
- Intelligence Index: 55 (AA v4.0), #7 of 147 models
- Input cost: $1.50/M tokens
- Output cost: $9.00/M tokens
- Context window: 1M tokens
- Multimodal inputs: text, image, speech, video
- Released: May 19, 2026
- Arena: Added to Text and Code leaderboards day-one