Lilith.
⌕
Editorial illustration: Gemini 4 Neon targets frontier performance at $3 per million tokens
Lilith illustration · editorial remix

Google has launched Gemini 4 Neon for software development, analytical work and long-running agentic tasks. The model offers up to one million tokens of context and costs $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 per million tokens.

Neon pairs a million-token context with a price below Carbon

Neon is adapted from Gemini 4 Carbon, but Google positions it between the deeper Carbon and the faster Hydrogen. It scored 61 on the Artificial Analysis index. The model card also reports 65.2% on Terminal-Bench 4.1, 70.9% on the Vals Index and 95.2% on GraphWalks for contexts from 256 thousand to one million tokens.

Availability varies by channel. The paid API is live in Google AI Studio and the model is rolling out to Gemini for business. Google AI Pro and Ultra subscribers can select it in the Gemini app and Antigravity. The cyber variant is due to reach selected defenders in the Fairwind Program over the coming days.

Engineering teams will count the cost of a completed job

Google says Neon cuts cost per task by 32% against Carbon at the default high setting. That matters more for agentic workflows than the API price alone. A cheaper token does not help if the model rereads files, loses context or forces a full rerun.

The gap between two SWE-sweep runs is therefore revealing. The mini-SWE harness yields 12.8%, while Antigravity reaches 39.0%. The result shows how much performance comes from orchestration, tools and context management. Buyers are comparing complete agent systems, not isolated models.

The benchmark table mixes independent scores with home-field tests

Some results come from independent leaderboards, some from Google's internal evaluations and some depend heavily on a particular harness. The boldest safety claims, including a 0.09% misaligned action rate and a 99.2% pass rate on an honesty evaluation, are internal measurements. The model card also says the full safety report is still forthcoming.

Google acknowledges hallucinations and continuing work on jailbreak and indirect prompt injection resistance. The knowledge cutoff is March 2026. Strong tables are no substitute for evals on a team's own repositories, data and permissions.

Repeated runs and final invoices will settle the argument

Teams should watch three signals: completion without reruns, total cost per finished task and the number of human interventions. That is where the claimed 32% saving over Carbon either survives contact with real tools and failures or disappears.

Long agent runs will provide the second test. If Neon maintains permission boundaries and an auditable trail across dozens of steps, it earns a production role. If it does not, the million-token context merely lengthens the route to a more expensive repair.

Lilith's verdict

Google has sent Neon onto the track with a lower price tag, but the harness still holds the finish line. Anyone measuring tokens instead of completed jobs may search the invoice for savings that never arrived.