Google just published hard pertoken bills for gemini-omni-1.1-flash $1.50 per 1M input tokens, $9.00 per 1M text output tokens and a jawdropping $17.50 per 1M video output tokens and paired that with clearer tiered pricing across Flash, FlashLite, Pro and Omni on the Agent Platform and Gemini Developer API. At the same time Cloud Run now has "Cloud Run instances": alwayson instances with a reference price (~$5.70/month for 1 vCPU + 1 GiB running 30 days). Those two moves together change the calculus of how you architect latencysensitive, multimodal AI services on GCP.
Let's be blunt: explicit pertoken prices for multimodal outputs force engineering tradeoffs into the open. The math is simple and unforgiving if your pipeline frequently emits video frames or long generated text, output tokens will dominate your bill. The $17.50/1M videooutput token line item makes it obvious where to push work back into cheaper systems: do preprocessing, local summarization, sparse retrieval, or serverside rendering outside the model before you pay for expensive model outputs.
Three immediate, practical consequences:
- Cache and reuse context aggressively. Google documents discounted rates for cached input tokens and distinct batch pricing bands; investing in local or shared context caches (and validating cache hit rates) directly translates to savings. Make context content immutable or strongly versioned so cached tokens remain valid across sessions.
- Batch and async inference become firstclass cost levers. Batch input pricing is explicitly cheaper in several bands move nonlatencysensitive requests to batch pipelines, and architect fallbacks for when realtime model calls are too expensive.
- Split model responsibilities. Use smaller Flash or FlashLite models for retrieval, filtering, or embeddings and reserve Omni/Pro for heavy multimodal outputs. The pricing pages now codify that split as a measurable optimization, not a hope.
Cloud Run's alwayson capability is the other quiet leverage point. Min-instances and the "CPU always allocated" option let you keep serverless containers warm; a continuously allocated 1 vCPU + 1 GiB instance costs on the order of $5–6/month depending on region and configuration. That provides a predictable, lowoverhead option to host warmed agent runtimes, local caches, or lowlatency inference routers and avoid repeated context reingestion. This is a straightforward way to reduce output token churn and therefore bill shock.
There are risks. Alwayson instances are an expanded operational surface that needs governance; cheap permonth instances multiplied across many services will still add up and create shadow spend if teams treat them like free. Pricing fragmentation including introductory discounts for Flash models and different context thresholds across model families will tempt teams into fragile optimizations that break when discounts end. If your architecture banks on introductory pricing, you're building on sand.
This is the right call from Google. Transparent, modellevel pricing removes guesswork and makes cost an engineering constraint you can reason about and code for. It also signals where platform teams should invest: token budgeting, cache layers, batch pipelines, and usageaware routing. Ignore it, and youll be surprised by bills; internalize it, and youll find straightforward levers to push down costs without sacrificing capability.
One practical note: Vertex AI release notes have adjusted some agent runtime pricing and clarified that features such as sessions, persistent memory, and code execution can be billable, so the incremental costs of stateful, agentic features should be included in longterm forecasting. Treat sessions and persistent memories as billable resources the same way you treat a database.
This isn't just a pricing update it's a nudge toward a new engineering pattern: treat tokens and context like capacity. Build caches, prefer batch, split models by job, and use lowcost alwayson frontends where latency matters. If you don't, the next monthly invoice will do that for you. For background on the pricing rhythm Google has been following with Flash models, see my earlier writeup on Gemini 3.7 Flash GA.
Prediction: within 12 months we'll see opensource libraries and internal platform primitives that expose token budgets per request, autoscale caches based on hitrates, and router layers that choose FlashLite vs Omni based on a cost�latency policy. If your platform team isn't allocating time to build those primitives now, your SREs will be doing expensive retrofits later.