Google has released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, two models pitched squarely at the running cost of enterprise AI agents. The headline claim is roughly 17 percent fewer output tokens than the previous Flash release, with synthetic coding benchmarks showing reductions of up to 65 percent, alongside better scores on multi-step reasoning tests. For anyone running background agents at volume, that is a meaningful improvement in the price of a unit of work.
It is also an admission about how the market prices intelligence. The reason token efficiency is a headline feature is that the economics of a production agent are dominated by a number most vendors prefer not to lead with: every extra token the model generates while reasoning adds cost and latency to a workflow that might run thousands of times an hour. Making the model chattier is expensive, so making it terser is a selling point. What has not changed is who owns the meter.
Aphelion AI is a private enterprise AI platform built to deploy your AI agent inside infrastructure you own and control, with a flat per user fee rather than per token billing, and with model agnostic selection so you are never tied to one vendor's roadmap or price list. That combination is the reason a 17 percent efficiency gain, welcome as it is, does not change the underlying calculation for a business planning three years ahead.
What Google Actually Shipped
The release is worth understanding properly before arguing with it, because the engineering is genuinely good. Google split the trade off between reasoning quality and throughput across three models:
- Gemini 3.6 Flash for coding and multimodal reasoning, priced at $1.50 per million input tokens and $7.50 per million output tokens, with a reported jump from 37 to 49 percent success on the DeepSWE benchmark.
- Gemini 3.5 Flash-Lite for high volume, low latency work at $0.30 per million input tokens and $2.50 per million output tokens, measured at 350 output tokens per second.
- Gemini 3.5 Flash Cyber, a restricted variant for validating and remediating code vulnerabilities, distributed only to governments and vetted partners through a pilot programme.
Google also folded a client side computer use tool directly into its API and enterprise agent platform, removing intermediary software that engineering teams previously had to build themselves. The direction of travel is clear enough: hosted agents are being made cheaper and faster so that more of your workload lives on the provider's platform.
A 17 percent reduction in output tokens lowers the price of each unit of work. It does not change the shape of the curve. On metered pricing, the bill still rises every time the agent is used more, which means efficiency gains are absorbed by growth rather than banked as savings.
Efficiency Gains That Never Reach the Budget
There is a well documented pattern in resource economics: when something becomes cheaper to use, consumption rises to fill the gap and total spend often increases rather than falls. AI agents are a textbook case. The teams most excited by a 17 percent token reduction are precisely the teams who will respond by giving the agent more work, deeper reasoning chains, longer context and more subagent calls. The per unit price falls, the unit count climbs, and the invoice grows.
That is not a criticism of the model. It is a description of what happens when the cost of your core capability is variable, external and indexed to your own success. Three specific consequences follow, and each of them lands on the finance team rather than the engineering team:
- Forecasting becomes guesswork. Budgeting for a metered agent means predicting adoption, prompt length and reasoning depth twelve months out, which nobody can do accurately.
- Optimisation competes with usefulness. Engineering effort goes into trimming tokens and routing requests to cheaper tiers, work that adds no business value and exists only to manage the meter.
- The price is not yours to set. Rates, tiers and deprecation schedules are the provider's to change, and your operations depend on the outcome.
A private deployment inverts all three. When the model runs on compute you own, the marginal cost of an additional reasoning step is close to zero, so the incentive to hold back disappears. Feeding the agent more context through data enrichment becomes a straightforward improvement rather than a line item to justify, because the value you add does not arrive attached to a proportional increase in the monthly invoice.
"Every efficiency gain on a metered platform is a discount on a bill you still do not control. Ownership is the only version of cheaper that survives contact with growth."
Stuart Smith, CEO, Aphelion AI
Model Agnostic by Design
The other lesson in this release is how fast the ground moves. Gemini 3.5 Pro is still in partner testing, pre-training for the next architecture is already underway, and competing labs are shipping on similar cycles. Any business that has wired its knowledge, prompts, integrations and governance directly into one vendor's API has quietly made a bet on that vendor staying ahead, staying available and staying reasonably priced.
Aphelion is built on the opposite assumption. The platform separates the knowledge layer, a business focused retrieval system populated with your own documents, from whichever model does the reasoning. Build the knowledge layer once, then point it wherever you like. If a new Flash-class model is genuinely the best fit for high volume document processing, use it. If a private open weight model on your own hardware is the right answer for sensitive work, use that instead, for the same knowledge base and the same governance.
That architecture is what makes model agnostic selection a practical option rather than a slogan. Because system integration is handled by the platform rather than hard coded against one provider's interface, swapping the underlying model does not mean rebuilding your CRM connectors, your document pipelines or your audit trail. The model becomes a component you choose on merit, not a dependency you inherit.
What Ownership Changes, Line by Line
Set the two approaches side by side and the difference is structural rather than a matter of a few percentage points. The figures below are indicative and exclude data enrichment and integrations, which Aphelion scopes and prices to each customer's needs.
| Consideration | Aphelion private AI | Metered hosted model |
|---|---|---|
| How you are billed | Flat fee per user, per week | Per million input and output tokens |
| Cost as usage grows | Flat, tracks headcount only | Rises with every interaction |
| Effect of an efficiency gain | Free capability improvement | Absorbed by higher usage |
| Who sets the price | You, at contract | The provider, at any time |
| Model choice | Agnostic, swap without rebuild | Locked to one vendor's range |
| Where your data goes | Stays in your environment | Transmitted to third party servers |
| Compliance evidence | Your own logs and controls | Vendor attestations you cannot inspect |
| Who manages it day to day | Your team, with Aphelion support | Opaque, changes on the provider's schedule |
None of this argues that hosted models are a bad product. For low volume, occasional or exploratory work they are the sensible choice, and the engineering behind a release like Gemini 3.6 Flash is exactly what the field needs. The argument is narrower and harder to dismiss: for an agent embedded in daily operations, the cost structure of renting intelligence works against you at precisely the moment the deployment starts to succeed.
Aphelion sets up your agents, applies default guardrails and then hands you the controls, so the capability is yours to manage and extend. Your knowledge layer is built once from your own documents, your models are chosen on merit rather than contract, and your spend is a fixed number you set rather than a variable someone else adjusts.
Frequently Asked Questions
What are AI token costs and why do they matter for enterprise agents?
A token is a fragment of text that a language model reads or writes, and hosted providers bill per million tokens in and per million tokens out. For a chat assistant used a few times a day that is trivial. For a background agent that reasons through multi-step tasks thousands of times an hour, every additional token of internal reasoning multiplies across the whole workload, so token efficiency becomes the single largest driver of running cost. The important consequence is structural rather than arithmetic: on a metered model your bill grows in direct proportion to how useful the agent has become.
How does Aphelion AI reduce the cost of running AI agents?
Aphelion deploys a private AI agent inside infrastructure you own or exclusively control and charges a flat fee per user per week rather than billing per token. Because the compute is yours, the marginal cost of an extra document, an extra reasoning step or an extra thousand queries is effectively zero, so unit economics improve as usage rises instead of degrading. Data enrichment and system integrations are scoped and priced separately, which means the running cost of the platform itself stays a fixed, forecastable line in the budget.
Is a private AI deployment more secure and compliant than a metered API?
It removes an entire category of risk rather than mitigating it. With a private deployment every prompt, document and output stays inside your governed environment, so sensitive material is never transmitted to a shared external platform or retained under a third party's policies. That makes GDPR, HIPAA and ISO work a matter of reviewing logs and controls you own, and it means a change to a provider's data handling terms cannot silently alter your compliance position. Default guardrails are applied at deployment and then handed to you to manage and audit.
Can a model agnostic AI platform integrate with existing business systems?
Yes, and that is the point of separating the knowledge layer from the model. Aphelion builds a business focused retrieval layer from your own documents and connects it to the CRMs, ERPs, databases and document stores you already run, then points that layer at whichever model you choose. Because the integration work sits with the platform rather than with any one model vendor, swapping the underlying model later does not mean rebuilding your connections, your prompts or your governance.
Gemini 3.6 Flash versus a private Aphelion deployment: which is cheaper at scale?
A cheaper, more efficient hosted model lowers the price of each unit of work but leaves the cost curve pointing the same way, so a busy agent still gets progressively more expensive and the price of that access remains the provider's to change. A private deployment carries a higher initial commitment and then flattens, because spend tracks headcount rather than interaction volume. For low volume, occasional use the hosted route is perfectly rational. For agents embedded in daily operations at sustained volume, ownership wins on three year total cost while also removing vendor dependency and data exposure. You can read more about the team behind the platform on our About page.
Reading the Efficiency Story Correctly
Faster, terser, cheaper models are good news, and they will keep arriving. The mistake is treating each one as a reason to postpone the decision about where your AI capability actually lives. Efficiency improves the terms of the rental. It does not convert the rental into ownership, and it does not move a single byte of your data back inside your walls.
The businesses that will be comfortable in three years are the ones that treat the model as replaceable and the knowledge layer as the asset. Build that layer once, on infrastructure you control, and every improvement the labs ship becomes an upgrade you can adopt on your own terms rather than a repricing you have to absorb. That is what Aphelion is for, and it is why owning your AI still beats renting it, however cheap the tokens get.