Google rebrands Bard as Gemini and launches Advanced plan

Google retired the Bard name on February 8, 2024, making Gemini the brand for its assistant, models and much of its AI offering. On the same day it launched Gemini Advanced, providing access to Ultra 1.0 through the Google One AI Premium plan, and began rolling out a mobile experience. The change simplified names but made it more important to identify what a customer buys: a model, interface, integrations, storage and service terms are different components.

That distinction applies to any AI subscription. “Advanced,” “Pro” and “Premium” do not describe a measurable improvement by themselves. Before paying, define tasks, verify which tier enables them and compare the result with the free option. This prevents a model’s reputation from being confused with the value of the complete product.

One name for several layers

The original Bard-to-Gemini announcement said the free experience already provided Pro 1.0 in more than 40 languages and over 230 countries and territories. Advanced added Ultra 1.0 and was initially available in English across more than 150 countries and territories. The plan cost $19.99 a month, included a two-month trial and carried Google One Premium benefits, including two terabytes of storage.

Google also said subscribers would soon use Gemini in Gmail, Docs, Slides, Sheets and other apps, replacing features previously grouped under Duet AI. “Soon” is not the same as available on purchase day. A rigorous inventory separates what is active, what has an announced date and what has no date. It also separates geographic availability, interface language, languages a model can process and the functions enabled in each country.

Sundar Pichai’s corporate explanation of the change showed the scale of the reorganisation: Bard became Gemini, Duet AI became Gemini for Workspace and Google One housed the consumer paid tier. This is a product strategy, not a technical property. The same name can identify a model family, chatbot, app and features embedded in other services.

The model is not the experience

Gemini 1.0 was divided into Nano, Pro and Ultra. The family’s technical report described multimodal models working across text, images, audio and video and reported results on dozens of benchmarks. Google presented Ultra as its most capable version and highlighted a 90% result on MMLU, obtained with a particular reasoning method, as an indicator of academic performance.

A benchmark score cannot predict how many correct emails, useful analyses or errors a subscription will produce. The outcome depends on the question set, prompt, scoring method and, in this case, an inference technique. It does not measure interface quality, usage limits, latency, conversation retrieval, file access or app integration either. A benchmark suggests what to test; it does not replace the test.

Google also said that external raters in blind evaluations preferred Advanced with Ultra 1.0 to leading alternatives. The announcement did not provide the complete conversation set, competitors, instructions or preference distribution, so the claim cannot become a percentage improvement for an individual. Treat it as a vendor hypothesis: Ultra may perform better on complex work, and a customer should identify which work.

The four-layer matrix

The first layer is access: declared model, languages, message or file limits, countries and exclusive features. The second is capability: tasks requiring reasoning, instruction following, multimodal analysis or long context. The third is surface: web, mobile app, email, documents and the places where work actually happens. The fourth is data governance: what information is sent, which account controls it, where output remains and what review or deletion options exist.

This matrix prevents incomplete comparisons. A somewhat better model inside an unused app may be worth less than a sufficient model embedded in a daily workflow. Two terabytes have value for someone already buying storage, but prove nothing about AI quality. An assistant inside email removes steps while bringing it closer to sensitive information. Keep each benefit in its own unit: minutes, errors, storage, exposure or money.

Mobile adds another operational layer. Google began rolling out an Android app that could be invoked as an assistant through the power button, a gesture or “Hey Google.” The company said many Google Assistant voice functions would be present and more would follow. On iOS, Gemini would live inside the Google app. Replacing the entry point does not guarantee every previous command survives; timers, calls and home control need testing alongside conversation and writing.

A test that fits the free trial

Before starting, choose ten to twenty real tasks rather than demonstrations designed to favour the tool. They might include summarising a known document, turning notes into an email, checking a formula, explaining an image and following instructions with several constraints. Define what success means for each, how long a person may spend reviewing and which failure is unacceptable.

Run every task on the free tier and Advanced with the same material and instruction. Record factual accuracy, compliance, time to a usable result, number of corrections and ease of moving the output to its final app. Do not score only the first response: an impressive answer whose every figure requires checking can cost more than a restrained, traceable one.

Test integrations separately. In email and documents, use non-sensitive or trial material, observe requested permissions and confirm where output is saved. On the phone, repeat ordinary commands and note missing features from the previous assistant. Future promises receive no score until they are active in the particular account and language.

The final calculation uses incremental cost. If a person already needs the included storage, subtract the expense it replaces from $19.99; if not, its value is zero. Then compare reviewed time savings, not raw generation speed. Record exit cost as well: documents, messages or workflows that depend so strongly on the platform that changing providers becomes difficult.

When to pay and when to wait

Advanced makes sense when Ultra repeatedly beats the free tier on valuable work, an integration removes real effort or the bundle replaces another expense. It does not when gains appear only on occasional questions, the important language is unavailable or review consumes the claimed saving. Cancelling after the trial is not failure; it is the correct result when evidence does not justify renewal.

The Bard rebrand tidied Google’s storefront, but customers still need to decompose it. Model, plan, application and integration answer different questions. The enduring skill is turning a brand into a matrix and a subscription into an experiment: fixed tasks, advance criteria, a free-tier comparison and incremental cost. That is how to buy demonstrated usefulness instead of the promise inside a new name.

Archive the final table with the date and tier tested, because aliases and bundles change. If a later renewal introduces another model or removes an integration, rerun only the affected tasks. The decision then remains grounded in comparable evidence rather than a memory of an earlier service version.