# Gemini 4 Neon Makes Google’s AI Race About Finishing the Job

> Google’s new model pairs a million-token context window with a focus on long tasks, mid-course corrections and the cost of completed work.

By Sasha Steenkamp · November 19, 2026

Canonical: https://blockster.com/google-gemini-4-neon-price-task-completion

Google has launched Gemini 4 Neon with a pitch that goes beyond answering harder questions: finish complex work quickly, and stay useful when the user changes direction.

The [November 18 announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) describes a model for software engineering, analysis and scientific workflows. Neon is rolling out through the Gemini API and business offerings, and to Pro and Ultra subscribers in the Gemini app and Google Antigravity.

## What it costs—and what that number leaves out

API pricing is $3 per million input tokens and $15 per million output tokens, with cached input priced at $0.15 per million. The context window is one million tokens.

Those prices measure text and other information processed by the model, not the price of a finished assignment. A long task may require repeated attempts, tool calls and human review. Teams comparing models need to count those costs too.

That is the same distinction behind Blockster's [Qwen4-Max analysis](https://blockster.com/alibaba-qwen4-max-ai-pricing-cost-per-task): a cheap response is not always the cheapest route to a usable result.

## Better at taking a correction

Google emphasizes multi-step instruction following and steerability—the ability to adapt when a user interrupts or adjusts an assignment. It also reports a lower fabrication rate in its own Antigravity testing.

Those are relevant qualities for anyone delegating a spreadsheet, research task or software change. The real test is whether the assistant incorporates the correction throughout its work, rather than acknowledging it and continuing with an outdated plan.

> “frontier-level intelligence with fast end-to-end task completion”
>
> — Koray Kavukcuoglu, Google DeepMind, introducing Gemini 4 Neon

Google cites strong benchmark results and lower task costs than some competing models. Those comparisons depend on the evaluation, settings and workload; they should not be treated as a guaranteed saving on every business process.

For users, a useful trial is concrete: give Neon a familiar task, make one realistic correction midway, and inspect the delivered result. Measure time spent checking and repairing it as well as time spent waiting. That is where a faster model becomes a more useful colleague—or simply produces mistakes sooner.
