Google I/O 2026's first bombshell was not Gemini 3.5 Pro, but Flash.
Traditionally, Flash has been the cheaper, faster, and less capable sibling under the Pro line. This time, it has surpassed Pro on several benchmarks: Terminal-Bench 2.1 Flash scored 76.2% vs 3.1 Pro's 70.3%; MCP Atlas Flash hit 83.6% vs 73.9%; and Finance Agent v2 Flash achieved 57.9% vs 43.0% — a nearly 15-point gap.
Seeking Alpha bluntly declared: Gemini 3.5 Flash beats GPT-5.5 on agentic benchmarks.
A Flash tier that runs like a flagship
Pricing: $1.5 per million input tokens, $9 per million output tokens, with a 1M token context window. Such pricing for a model that can beat GPT-5.5 on agentic tasks is unprecedented.
Speed: 4x faster than comparable frontier models, and Google says an "optimized version" can triple that to 12x.
More importantly, Flash is no longer a "budget small model" — it is aimed directly at agent scenarios. On the GDPval-AA benchmark, 3.5 Flash scored 1656 Elo — less than 100 points behind Claude Opus 4.7's 1753, but at a fraction of the cost. Koray Kavukcuoglu, Google's head of Gemini models, said:
"3.5 Flash offers an incredible combination of quality and low latency."
Behind the scores: Antigravity 2.0
The benchmark numbers are powered by Antigravity 2.0, the upgraded agentic development platform launched alongside Flash. Flash is optimized for multi-agent parallel orchestration — it can spin up dozens of subagents for a single task and aggregate results. TechCrunch reported that Google demonstrated Flash building an entire operating system from scratch, taking several hours with occasional pauses to ask questions.
As Tulsee Doshi noted:
"will at times pause and ask for user input"The model actively stops at key decision points to wait for user input — not a fully autonomous drive, but an "agent that stops at red lights."
The real killer move: Flash as default
Gemini 3.5 Flash is now the default model for the Gemini app and AI Mode in Google Search, globally. Previously, users got the 2.5 or 3.1 Flash experience by default. Now Google has pushed Pro-level agentic capabilities into the default tier — users feel no price change, but get a generational upgrade in capability.
More critically, Spark — the $100/month 24/7 AI assistant — runs on the 3.5 Flash base. That means this "budget" model powers Google's most expensive subscription product at I/O.
The next price war
When OpenAI released GPT-5.5, the market assumed reasoning models would maintain a premium for months. Google has now undercut GPT-5.5's agentic scores with Flash-tier pricing. After DeepSeek V4-Pro slashed prices by 75% within 72 hours of launch, both open-source and closed-source models are racing to the bottom. This pricing puts pressure on Anthropic and OpenAI: can their flagship models justify being 5-8x more expensive than a Flash-tier model that matches or beats them?
| Benchmark | Gemini 3.5 Flash | Gemini 3.1 Pro | Claude Opus 4.7 |
|---|---|---|---|
| Terminal-Bench 2.1 | 76.2% | 70.3% | - |
| MCP Atlas | 83.6% | 73.9% | - |
| Finance Agent v2 | 57.9% | 43.0% | - |
| GDPval-AA Elo | 1656 | 1314 | 1753 |
And the kicker: Google sources have indicated that Gemini 3.5 Pro will launch next month.
Sources: With Gemini 3.5 Flash, Google bets its next AI wave on agents, not chatbots (TechCrunch); Gemini 3.5 Flash: Google's Fastest Agentic Model (DataCamp); Google releases Gemini 3.5 Flash; CocoLoop, surpasses GPT-5.5 in agentic benchmarks (Seeking Alpha)