Google's Gemini 3.5 Flash Cuts LLM Output Costs to $9 Per Million Tokens

Google undercuts flagship pricing with production-speed model

Google launched Gemini 3.5 Flash in June 2026 at $1.50 per million input tokens and $9 per million output tokens — pricing that delivers near-flagship performance at a fraction of the cost of earlier-generation enterprise APIs. The model runs four times faster than previous Gemini versions while Google simultaneously retires the entire Gemini 2.0 family effective June 1, collapsing its SKU lineup and forcing customers onto the 3.x line.

For a production copilot processing 50 million input tokens and 10 million output tokens monthly, the model cost drops to roughly $165 per month before infrastructure and margin. That unit economics shift makes multi-workflow LLM deployment affordable at enterprise scale and directly targets contact centers, real-time copilots, and agentic automation where sub-second latency matters. Enterprises that standardized on Gemini 2.0 now face re-testing and model re-qualification costs, particularly in regulated environments where prompt behavior and output drift require validation.

OpenAI makes GPT-5.5 Instant default, launches deployment arm with 19 SIs

OpenAI made GPT-5.5 Instant the default ChatGPT model in June 2026 and launched the OpenAI Deployment Company alongside the model update. The new entity partners with 19 global investment firms, consultancies, and system integrators — including Bain, McKinsey, and Capgemini — to build AI into enterprise core operations. OpenAI reports that enterprise revenue now accounts for more than 40% of total revenue and is on track to reach parity with consumer revenue by year end.

The default model shift signals OpenAI's view that speed-optimized models are sufficient for most workloads, aligning with the broader industry move toward reason-when-needed architectures rather than always-on heavy reasoning modes. The Deployment Company bundles OpenAI models and tooling with SI implementation capacity, turning OpenAI from a pure API vendor into a full enterprise program partner. That directly competes with Microsoft's own AI consulting, EY, and cloud SIs that previously mediated OpenAI access.

Deployment economics now favor router architectures over single-model strategies

Gemini 3.5 Flash's speed and cost profile makes it viable as a router default model: enterprises can route most calls to Flash and selectively escalate to heavier models only when reasoning depth justifies the cost and latency penalty. That mirrors OpenAI's Instant versus Thinking model split and Kimi's reason-when-needed pattern. Google's Managed Agents, which let developers spin up a sandboxed Linux environment where an agent can reason, execute code, browse, and manage files via a single API call, extend the same principle to agentic workflows.

The forced Gemini 2.0 retirement introduces vendor lock-in risk. Enterprises with 2.0-tuned prompts and evaluation baselines face re-validation costs and potential behavior drift. Risk-averse buyers may hedge with a multi-model gateway to avoid being trapped in future SKU retirements, particularly as competitive pressure from Anthropic's Claude Sonnet 4.5 and Mistral's open-weight models for on-premises and VPC deployment increases.

SI-led deployment programs shift budgets from API subscriptions to transformation projects

With SIs embedded in OpenAI's Deployment Company, buying centers shift from pure software subscriptions to large transformation projects. CIOs and CTOs should expect multi-year, multi-million-dollar proposals rather than usage-based API bills. The competitive landscape now includes EY and Microsoft's $1 billion deployment initiative, which targets the same strategy and deployment budget. Anthropic's Claude Code, credited with driving $2.5 billion in annual recurring revenue nine months after launch, demonstrates that deployment-focused offerings generate materially higher revenue per customer than API access alone.

OpenAI's disclosure that more than 40% of revenue is now enterprise strengthens its negotiating position but also signals that large-enterprise references and price tiers exist. Buyers should expect volume commitments and enterprise licensing agreements rather than pure consumption pricing, particularly for deployments that involve SI partners.

What to watch

Imminent regulatory deadlines — including the Colorado AI Act and EU AI Act enforcement dates — change deployment risk calculus. Enterprises must validate that router architectures and multi-model gateways preserve audit trails and model lineage across escalation paths. The Gemini 2.0 retirement deadline of June 1 forces immediate re-qualification for any regulated workload currently running on that family. Buyers should also track whether Google's SKU consolidation triggers similar moves from OpenAI and Anthropic, as model retirement risk becomes a standard enterprise architecture consideration rather than an edge case.