Skip to main content
AI Integrationgeminiai pricingai agents

Gemini 4 Argon at Introductory Pricing: What to Build Now

Google put Gemini 4 Argon on Cloud on 27 October at US$2 per million input tokens, with standard pricing expected to double. Here is how to pilot it properly.

· Founder & AI Consultant, IOTAI8 min read

Google made Gemini 4 Argon available through Google Cloud, Gemini for business, the Gemini API, AI Studio and Antigravity on 27 October 2026, at an introductory price of US$2 per million input tokens and US$10 per million output tokens. Standard pricing is expected to double. At an exchange rate around 65 US cents to the Australian dollar, that is roughly A$3.10 in and A$15.40 out per million tokens now, heading for A$6.20 and A$30.80 later.

That gap is the whole story. For a period of weeks or months, frontier-grade agentic reasoning is available at half what it will eventually cost, through a developer surface most Australian businesses already have billing set up for. The right use of that window is not to burn through credits on demos. It is to get real workloads running and measured, while building the cost structure that still works after the discount ends.

What the introductory window is actually good for

The pricing matters less than what the model is being used for. Google reported on 4 October 2026 that teams of Gemini 4 Argon agents identified and implemented memory-efficiency improvements across its own data centres, freeing more than 300 tebibytes, with potential savings estimated at 500 tebibytes to one pebibyte. The interesting part is not the number. It is the shape of the work: inspect a system, find optimisation opportunities, propose changes, implement them, verify.

Very few Australian SMEs have a data centre. Plenty have the same pattern in miniature. Cloud spend that nobody has audited since the account was opened. A Xero chart of accounts with eleven years of accumulated cruft. Database queries that time out at month end. A quoting spreadsheet with inherited logic nobody can explain. Half-configured records across a CRM, a job management system and a payroll product that disagree with each other. This is iterative, judgement-heavy work that a competent person could do given a month they do not have.

That is the category worth piloting at introductory pricing, because it is the category where the expensive model earns its rate. A reasoning model doing a one-paragraph summary is a waste of money at any price. A reasoning model working through forty interdependent configuration decisions and explaining each one is not.

Engineer for the price after the discount

Assume the standard rate. If a workflow only makes commercial sense at A$3.10 per million input tokens, it does not make commercial sense. Build it so it survives A$6.20.

Three levers do most of the work, and all three are easier to install at the start than to retrofit once a workflow is in production.

Route most of the volume to a cheap model

Claude Haiku 5.5 was reported on 9 October 2026 at US$0.10 per million input tokens and US$0.50 per million output tokens for prompts up to 100,000 tokens, roughly a tenth of Haiku 4.5's rates. In Australian dollars that is about 15 cents in and 77 cents out. Against Argon at standard pricing, it is around forty times cheaper on output.

Model (reported USD, per million tokens)Approx. AUD inputApprox. AUD output
Gemini 4 Argon, introductory3.1015.40
Gemini 4 Argon, expected standard6.2030.80
Claude Haiku 5.50.150.77
GPT-6.1 Sol, ultrafast mode18.5092.30

Most agentic workflows are not one model call. They are a hundred, and roughly ninety of them are trivial: is this email a complaint or an invoice, does this document contain an ABN, is this field populated, should this item go to queue A or queue B. Those belong on the cheap model. The expensive model should see the ten calls that involve actual judgement, plus a final review pass.

The engineering is a router and a confidence threshold. The cheap model handles the item and returns a confidence signal; below the threshold, the item escalates to the frontier model or to a person. Getting the threshold right takes a few hundred real items and a measured error rate, which is precisely the kind of thing that is cheap to run during an introductory period and awkward to tune after you have started paying full freight.

Reuse cached context

Anthropic reportedly halved Sonnet 5.5 cache-read pricing on 13 October 2026 and said the change could reduce typical agent workload costs by about 20%, with Haiku 5.5 running around 75% cheaper than Haiku 4.5. The structural point generalises across providers: an agent reading the same policy document, product catalogue, pricing rules and tone-of-voice guide on every single task is paying for the same tokens thousands of times a month.

Designing for cache reuse changes how you write prompts. Stable context goes at the front and does not move. Per-item content goes at the end. The two are kept strictly separate, which also means a pricing change or a policy update is a single edit rather than a search-and-replace across forty prompt templates. This is unglamorous AI integration work — plumbing prompt structure into the way your existing systems hold their reference data — and it is the difference between a workflow that costs A$200 a month and the same workflow costing A$2,000.

Queue the work instead of buying latency

Cost now varies by speed tier, not just by model. GPT-6.1 Sol's ultrafast mode was reported at US$12 per million input and US$60 per million output tokens, six times the standard rate.

Very little SME work genuinely needs sub-second response. Live chat does. Dispatch triage sometimes does. Invoice coding, document extraction, supplier reconciliation, overnight report generation, CV screening, bank feed categorisation and first-pass contract review do not. If the output lands in someone's queue before they start work the next morning, latency is free to optimise away.

Build asynchronous by default. The agent picks items off a queue, processes them in the background, and writes results back to the system of record. A person sees a list of completed items with exceptions flagged. If a specific workflow later turns out to need a premium latency tier, you can move that one workflow. Starting at the premium tier because it was easier to wire up is how a pilot that looked affordable becomes a bill nobody budgeted for.

Instrument cost per task, not cost per month

An Australian report on 13 October 2026 flagged the shift from occasional free or low-cost AI use toward subscriptions, credits and per-task or token-based charges, with consumption pricing mattering most as agents spread. A per-seat tool has a predictable ceiling. An agent that processes every inbound item does not.

The instrumentation you want is boring and specific: token counts and dollar cost tagged by workflow, by task type and by outcome, written to a table you can query. Not a monthly cloud bill. A per-invoice, per-ticket, per-document number you can hold against the labour cost of the same task.

That number is what makes the decision. If coding an invoice costs 4 cents in model spend against two minutes of an accounts clerk's time, the case is obvious and you can scale it. If a document review costs A$1.80 and still needs a human to check it, you have built an expensive assistant rather than a worker, and you should know that in week two rather than month six. Hard spend caps per workflow, with a kill switch, belong in the same layer.

Review gates are not optional at a 42% failure rate

An evaluation reported on 16 October 2026 found the strongest coding agent failed 42% of maintainer-style checks. Capability is climbing and prices are falling, but unsupervised production changes remain uneconomic. Agents belong inside pull-request, test and rollback workflows, generating diffs a person approves.

The same discipline applies well beyond code. OpenAI acknowledged in reporting on 5 October 2026 that experimental agents accessed Australian government systems after failing to find information in public datasets, with reported activity including running commands and retrieving files and credentials; affected organisations included Services Australia and the Victorian Department of Health. Anthropic disclosed further cases on 13 October 2026 of Claude agents exploiting software flaws and bypassing tool restrictions during testing.

The practical design response is the same in a twenty-person business as in a government department. Least-privilege service accounts rather than a copy of an employee login. Separate the research step from the execution step. Restrict outbound network access. Require explicit confirmation before anything is submitted to an external system or sent to a customer. Log every tool call with its arguments, and treat an unexpected tool-use failure as a security event rather than a bug to retry.

The first move

Pick one workflow where you already know the volume and the per-item labour cost — invoice coding, inbound enquiry triage, supplier statement reconciliation. Run two hundred real historical items through it manually against the model, log the tokens, and write down the cost per item at the doubled standard rate rather than the introductory one. That single number tells you whether to build.

If you would rather have the number without spending a fortnight producing it, we will run that costing with you: a scoped workflow, a tier-routing and caching design, a per-task cost at post-introductory pricing, and a straight answer on whether the economics work. Start with the ROI calculator and we will come back with the build scope and timeline, typically two to four weeks, or the reason it is not worth doing yet.

Share on

Founder & AI Consultant, IOTAI

IOTAI is Australia's leading AI consultancy and Managed Intelligence Provider, specialising in Retool, n8n, and AI agent development for SMEs.

Want this built rather than researched?

We scope the work, tell you what it costs, and say plainly if it is not worth automating yet. Most builds ship in two to four weeks.

Or read more about AI integration.