Close the laptop and the job keeps running. That is the practical promise behind the new tasks switch in Google’s Gemini app: instead of staying in a conversation, a user can hand over work that continues while they are away.
Google’s October 13 rollout brings Gemini 4 Argon and tasks to Google AI Pro and Ultra subscribers. It follows the model’s September 30 announcement. Developers now have paid API access, while Google is also putting its Antigravity coding interface into the Gemini macOS app.
The interesting change is not simply a model that can produce more text. Google is packaging longer work in several different ways. A task that continues outside the chat, an extended model response and a useful finished result are not the same thing. One public code change makes that distinction unusually clear.
A small bottleneck, a real code change
Google’s Agent Development Kit, or ADK, supplies plumbing for software agents. Its in-memory session service was copying an entire conversation history before selecting the events it actually needed. As a conversation grew, that made routine bookkeeping more expensive.
An October 7 commit attributed to Argon changes that behavior: select the needed events before copying them. It also improves deep copies and delays some module imports until they are needed. The proposed benefit is intelligible without trusting a demo: avoid work the program does not need to do.
The project reports that, after 150 conversation turns, the wait before the first model request fell from 200 to 122 milliseconds. Across its entire 158-turn test, runtime fell from 65.6 to 56.1 seconds, about 14%. But a newly started process took essentially the same time to reach its first model request: 3.07 seconds before, 3.09 afterward.
Those measurements isolate ADK overhead with a no-op model and instantly returning tools; garbage collection was disabled for the first-request tests. Vastkind has not reproduced them. The code is inspectable, and the claimed improvement is specific. It is not evidence that Argon makes every agent 40% faster.
That distinction matters as more AI work moves beyond a single answer. As Arm’s attempt to replay agent workloads illustrates, a working agent depends on much more than the language model.
A million tokens is one shared workspace
For developers, Argon’s larger generation budget gives a difficult response more room to continue. Its 1,048,576-token window is shared by the input, internal reasoning and output. Tokens are the pieces of text and other content the model processes. This is not a million tokens of input plus another million for the answer.
The default generation limit is 65,536 tokens, including reasoning. In the Interactions API, going beyond that requires background execution and an explicitly higher limit. Turning on background execution alone does not raise the cap. Google’s servers continue the response in chunks of up to 64K tokens.
This can help when one difficult calculation or code response needs more room. It does not mean every agent needs a huge response. Google’s long-decode guide says agents that make many short tool calls generally do not need it. A long workflow can consist of many small steps, rather than one enormous answer.
Background API interactions can run for up to 24 hours. That ceiling is neither an app-task duration promise nor a completion guarantee. More room permits more work; it does not establish that the work is correct.
Access, time and cost remain separate
In the app, the change is a switch from chat to tasks, with connected services and progress updates. In the API, developers control requests, budgets and how their software collects results. A Pro subscription’s app access should not be confused with paid API access or an included API allowance.
Google’s standard API rates are $2 per million input tokens and $10 per million output tokens, including reasoning, through November 30, 2026. They rise to $4 and $20 on December 1. These are token rates, not app subscription prices or a fixed fee per completed job.
Long responses also reuse previous context as input at each continuation. Cached input is discounted, but a long job is not priced by its final visible answer alone. Processing already performed still costs money if the interaction fails or is canceled.
What counts as finished?
Google’s model card still lists hallucinations, occasional slowness and timeouts. Its coding results vary across tests rather than showing a clean sweep. These are documented capabilities and product announcements, not a Vastkind hands-on reliability test.
The ADK example offers a useful standard: a concrete change, a checkable mechanism and measurements tied to the part of the system that changed. That is more informative than how long a model can stay busy. Our explanation of AI task-length measurements shows why duration alone is an uneven yardstick.
Argon gives users more ways to leave a job running. The meaningful handoff comes when they can return to an inspectable result, not merely a longer transcript.
