Using Google Gemini for customer support: the cost-leader on long-context tickets

Gemini's lineup, ranked for support

Why Gemini is the OpenAgent default

OpenAgent ships with Gemini configured by default for two reasons:

• Lowest unit economics. At ~$0.0011/ticket on Flash, AI support at any volume is a rounding error on most budgets. Other providers are within an order of magnitude; Gemini is the cheapest by ~30%.
• First-party embeddings. The same provider handles both retrieval (embeddings) and generation. One key, one bill, one data residency conversation.

The long-context advantage

Gemini 2.5 Pro has a 2M-token context window. Flash sits at 1M. Most support tickets don't need anywhere near that much context, but two situations make the headroom genuinely useful:

1. Returning visitors. A B2B SaaS visitor with 80 past conversations? Gemini reads the entire history in one shot. GPT-4o would truncate or summarise.
2. Cross-document reasoning. “If a customer on the Pro plan is in Germany and bought before March 2025, do they qualify for the legacy 5% discount?” Gemini can consume the pricing page, the eligibility doc, and the regional policy in one prompt.

Picking the right Gemini model

For most teams: Flash everywhere. The cost difference between Flash and Pro is ~7x; the quality difference on grounded support questions is much smaller. Pro's advantage shows up on multi-step reasoning, and most support traffic isn't multi-step.

For workspaces with a refunds team: route the “Refunds” agent to Pro. The cost delta on the 5-10% of traffic refunds represent is small in absolute terms, and the answers are noticeably more nuanced.

Wiring your Gemini key into OpenAgent

1. Get a key. Two paths:
   • Individual / sandbox: aistudio.google.com/apikey gives you a key in two clicks.
   • Enterprise: create a service account on Google Cloud, enable Vertex AI, generate a JSON credential. Paste the service account JSON in OpenAgent's Vertex AI provider config.
2. In OpenAgent: Settings → Providers → Add provider → Gemini. Paste the key. Encrypted on submit.
3. For each agent, the model dropdown defaults to gemini-2.5-flash. Leave it alone unless you have a reason to upgrade.

Embeddings, configured automatically

OpenAgent uses gemini-embedding-001 at 768 dimensions for the RAG knowledge base. The configuration is set during workspace provisioning; you don't need to touch it.

One nuance: if you switch your generation model away from Gemini later (say, to Claude), the embeddings stay. The embedding model and the generation model are independent, so you can mix and match. Many teams keep Gemini for embeddings (cheapest, fastest) and run generation on a different provider.

Watch-outs specific to Gemini

• Free tier limits matter for testing. AI Studio free keys throttle at ~15 requests per minute. Fine for development, instantly insufficient for production. Move to a paid key before you ship.
• Safety filters can over-trigger. Gemini's default safety thresholds occasionally refuse benign support questions (e.g. anything mentioning “legal”). In OpenAgent's provider config, you can relax the safety thresholds for production agents; we recommend BLOCK_ONLY_HIGH for HARASSMENT/HATE_SPEECH/SEXUALLY_EXPLICIT, BLOCK_NONE for DANGEROUS_CONTENT (which over-triggers most).
• Regional latency. AI Studio routes through US west by default; Vertex AI in your region cuts 100-300ms of latency. Worth the migration if your visitors are geographically concentrated.

Bottom line

Gemini 2.5 Flash is the cheapest credible model for customer support, with the best price-to-quality ratio as of mid-2026. Pair it with first-party Gemini embeddings and you have a single-provider stack that costs roughly a dollar a month even at moderate scale. It's the OpenAgent default for a reason.