Google offers Gemini 4 Neon, a reasoning model designed for demanding tasks, from programming to searching for information across multiple files. In four tests conducted within the Gemini app, it succeeded in several complex exercises but made errors on simpler tasks, such as counting audio pulses.
Neon is adapted from Gemini 4 Carbon. Google positions it between Hydrogen, the fast option for daily use, and Carbon, reserved for the most difficult and lengthy projects. It thus targets intermediate workloads, such as analyzing technical documents or writing a testable utility.
Announced results across several rankings
At launch, Neon scored 61 points on the Artificial Analysis Intelligence Index in Xhigh reasoning mode, ranking it third overall. In High mode, it scores 60 points, the same as Carbon in the same mode. It also places second on the Vals Index with 70.9% and second for the ability to be guided on Agent Arena, a test measuring agents’ adaptation to changes mid-task.
The figures vary depending on the ranking and evaluation framework. On the SWE-sweep benchmark, Neon achieves 39.0% with the provider’s harness, compared to 12.8% in the common Mini-SWE environment. Claude Opus 5.5 is slightly ahead on Terminal-Bench 4.1, with 65.8% versus 65.2% for Neon. These are the scores published at launch, with different settings than those used in practical tests.
Practical tests: successes and errors
On a thirteen-second audio recording containing three types of alerts, Neon correctly identified the order of the patterns and the final descent. However, its timestamps were slightly early, and it counted too many pulses in the first two patterns.
For reading a vernier caliper, it obtained the measurement of 4.6 mm by explaining the origin of each value. However, the phrasing of its answer could have been clearer regarding the gap between the graduations.
The most notable test involved a Python series compression codec. The produced code, unmodified, passed 781 round-trip checks, split a 205-character series into 99, 99, and 7, and rejected all nine malformed inputs tested. Conversely, the ranking of a tournament after modifying a single tie-breaking criterion was not correctly updated in the app’s two responses.
Access and pricing
In the Gemini app, Google AI Pro and Ultra subscribers can select 4 Neon from the model selector, then choose the reasoning level. Via the Gemini API, the identifier is gemini-4-neon, on a paid project. The context window is 1,048,576 tokens, shared between input, model reasoning, and output, and the standard synchronous output is limited to 65,536 tokens.
The standard API price is $3 per million tokens for input and $15 per million tokens for output, reasoning included. According to Artificial Analysis, a test task costs an average of $3.74 with Neon in High mode, compared to $7.76 with Carbon in High mode, at Carbon’s non-discounted rates. The latter benefits from a temporary rate of $2 for input and $10 for output per million tokens until November 30.
