Google introduced Gemini 4 Neon on November 18 and began rolling it out through its API and paid products. Announced API rates are $3 per million input tokens and $15 per million output tokens. The launch, also covered by Kaldata, emphasizes extended tasks combining tools and different kinds of source material.
What one million actually means
The model card describes a combined input-output context of up to one million tokens. Default synchronous output is up to 64K, while longer background generation is a separate mode. A large context therefore does not mean an arbitrarily long immediate response. The document acknowledges hallucinations and gives March 2026 as the training knowledge cutoff.
For users, the distinction is between capacity and reliability. Accepting more material allows more dependencies to be included in a task. It does not establish that every table, note or contradiction will be handled correctly. Output still needs verification, especially when it will change a document, software project or business decision.
The cost of finished work
Our editorial interpretation is to compare the complete route to a checked result. For an illustrative task using 200,000 paid input tokens and 20,000 output tokens, the announced rates give $0.90 for those tokens. This is our arithmetic example, excluding caching, additional tools and retries, rather than an invoice for a particular customer workflow.
If an initial answer requires several further attempts, the final cost and review time change. We would therefore record the reasoning level, input, number of steps and human corrections. A comparison based only on the pricing table can favor a mode that fails to finish the required work.
The execution environment matters too
Kaldata highlights differences between results obtained in different execution environments. The same question belongs in a local pilot: models need the same tools, permissions and completion criteria. Otherwise the comparison measures different systems as well as different models.
For a team handling large projects, a useful acceptance example would include known contradictions and check whether the final deliverable addresses them. Separately verify that claimed files exist and applicable checks were actually run. This is an editorial evaluation proposal, not a benchmark performed by SIExpo.
Neon expands the choices for these workflows. Its practical value will depend on the quality of completed work under the relevant budget, rather than context-window size alone.



