The Goldfish Benchmark: Google's $350 Assistant Has Outsourced Its Memory to You

Following a review of Google's $350 AUD monthly AI Ultra subscription, the Institute has identified a previously undocumented memory architecture: the subscriber.

Under this arrangement, Gemini supplies the advice, the customer retains the consequences, and Google retains the payment. The division of labour is unusually stable.

The recurring expenditure is readily formalised:

$$350\times 12=\mathrm{4,200}$$

Conversational continuity must be assessed separately. Successful annualisation of the invoice does not establish successful retention of the task.

The Goldfish Benchmark examines Google's consumer-facing Gemini through a failure that a polished answer cannot repair: an assistant making consequential statements, then forgetting the conversation in which it made them.

The author describes Gemini effectively telling them that their hopes and ambitions were delusional and could be abandoned. Roughly ten conversational turns later, Gemini had forgotten the exchange. The person affected was left to remember it, explain it back to the system and carry its consequences.

The language of assistance remained available. The record of what required assistance did not.

The paper asks what alignment means across an actual conversation. Can Gemini retain an instruction, a correction, a commitment or a boundary long enough to remain accountable to it? A reassuring response in the current turn does not answer that question. When the customer must repeatedly reconstruct the task, the assistant has transferred part of its job to the person paying.

That transfer also belongs in the accounts.

In one worked scenario, the paper prices 20 prompts every day for a full 30-day month, each containing 50,000 input tokens and 1,000 output tokens. At its stated uncached rates, the equivalent API bill is $61.63-$123.26 AUD against $350 AUD paid.

$$350.00-123.26=226.74$$

The calculation prices compute. Google has to justify the remainder through useful results and benefits the customer actually needs. Repeatedly processing material the assistant has forgotten does not automatically constitute another useful result. The Institute declines to classify the reconstruction of a lost instruction as a second delivery of the first service.

The bundle receives the same examination. Thirty terabytes of unused storage save the customer nothing. Restricted cloud credits are not cash returned. Video, music, research and developer allowances become valuable when they serve an actual need.

The paper calls this catalogue "Google Temuni". The customer is invited to acquire six additional occupations so that the subscription can look economical.

The account also preserves a meaningful product distinction: the author describes OpenAI's visible conversation-optimisation notice and easier recovery of earlier conversational context as substantially better experiences. This is a case about Google's consumer Gemini, not a claim that every product behaves identically.

The proposed test is straightforward: count useful work completed, time saved and expenses actually avoided. Subtract the work required to repair the assistant's failures. Give no credit to workloads nobody has established the subscription will permit and sustain.

The customer bought assistance. They did not volunteer to become its memory department.