Gemini 3 Pro Safety Evaluations and Red-Team Results

Published 2026-04-14 | Updated 2026-05-05

Google DeepMind published a 38-page Gemini 3 Pro safety report alongside the model — here's what is actually new.

This briefing is written with builders in London, UK in mind — local procurement, latency from regional Google Cloud / AWS / Azure regions, and time-zone-friendly support windows shape the practical recommendations.

What Shipped and Why It Matters

Google's April 2026 cadence around the Gemini 3 family, Antigravity, and the AgentSpace surface is the most coherent product narrative the company has put together in years. The pieces fit: a frontier model (Gemini 3 Pro), a fast variant (Gemini 3 Flash), an on-device tier (Gemini Nano), an IDE (Antigravity), an agent runtime (Vertex Reasoning Engine), an agent catalog (Agent Garden), an enterprise hub (AgentSpace), and a consumer notebook (NotebookLM Pro). For builders, the practical impact is that you can pick a Google story for almost any agent shape and have a credible delivery path from prototype to production.

For London, UK teams, the practical near-term move is to set up an evaluation harness against your top 3 production prompts before committing to a model swap.

Benchmarks That Actually Matter

On SWE-bench Verified, Gemini 3 Pro scores 71.8% — within striking distance of Claude Opus 4.7's 72.9% and ahead of GPT-5.5's 69.4%. On tau-bench retail, the new model lands at 95.1%, a meaningful jump from Gemini 2.5's 88.6%. MMMU sits at 84.0%. The numbers matter less than the spread: for the first time, the three frontier labs are within 3 percentage points of each other on most benchmarks that builders cite.

Pricing and Total Cost of Ownership

Gemini 3 Pro is priced at $1.25 / $10.00 per million input/output tokens up to 200K context; long-context (>200K) tier kicks in at $2.50 / $15.00. With prompt caching at a 75% discount and a 50% Batch API discount on async workloads, the realized cost for many production agents lands closer to $0.80 per million blended tokens. Compared to Claude Opus 4.7 ($15/$75) and GPT-5.5 ($10/$30), Gemini 3 Pro is positioned as the price-aggressive frontier option.

This is the short version; the full vendor documentation has more nuance, particularly on rate limits and regional availability.

Deployment Path: AI Studio to Vertex

The recommended path is prototype in AI Studio, then promote to Vertex AI for production. Vertex provides regional availability (12 regions globally, including europe-west4 and asia-southeast1), VPC-SC, CMEK, audit logging, and the new Reasoning Engine managed runtime. AI Studio's prompt IDE got a major refresh — versioned prompts, side-by-side eval, and one-click deployment to Vertex are now first-class.

Agent Stack: A2A, MCP, and the Garden

Google's open-protocol bet is real: A2A 1.0 ships as an open spec for agent-to-agent communication, complementing MCP 1.0 for tool integration. Vertex AI Agent Builder ships first-class A2A support, and Agent Garden's 80+ pre-built agents all advertise A2A endpoints. For builders, this means a Google-built sales agent can hand off to a third-party fulfillment agent (running on AWS or self-hosted) without custom integration glue.

For London, UK teams, the practical near-term move is to set up an evaluation harness against your top 3 production prompts before committing to a model swap.

Practical Builder Checklist

If you are evaluating this release for a 2026 deployment, work through the following checklist before signing a contract:

1. Confirm Vertex AI region availability for your data residency requirements (europe-west4 and asia-southeast1 are the two most-asked-for in 2026).
2. Run your top 3 production prompts against Gemini 3 Pro AND Gemini 3 Flash; the cost-quality crossover is workload-specific.
3. Validate prompt caching savings on your real traffic shape — 75% discount is a marketing maximum, realized savings vary.
4. Test A2A interop with at least one third-party agent before betting your architecture on it.
5. Stress-test long-context recall at 800K+ tokens; degradation past 1M is workload-dependent.

FAQ

Q: Is Gemini 3 Pro available in my region?

A: Gemini 3 Pro is generally available in 12 Vertex AI regions as of May 2026, including us-central1, europe-west4, asia-southeast1, and asia-northeast1. Check the Vertex AI region availability docs for the latest list.

Q: How does Gemini 3 Pro pricing compare on a real workload?

A: Headline price is $1.25 / $10.00 per million tokens up to 200K context. With 75% prompt cache discount and 50% Batch API discount, realized blended cost on long-running agent workloads typically lands at $0.80-$1.20 per million tokens.

Q: Can I use Antigravity with Claude or GPT-5.5?

A: Yes. Antigravity is unusually open — Claude Opus 4.7, GPT-5.5, and Gemini 3 Pro are all first-class providers in the IDE settings.

Q: What is the difference between A2A and MCP?

A: MCP is the agent-to-tool protocol; A2A is the agent-to-agent protocol. They are complementary, not competitive — most production agent stacks will use both.

Sources

• https://www.bloomberg.com/news/articles/2026-04-google-ai-strategy
• https://www.reuters.com/technology/google-deepmind-releases-2026/
• https://www.techcrunch.com/2026/04/google-gemini-3-pro-launch/
• https://www.theverge.com/2026/04/google-antigravity-ide

---

Last reviewed 2026-05-05. Pricing and benchmarks change frequently — check primary sources before relying on numbers in this article.