Skip to content
The AI-First Web

Google's Gemini 4 Neon shifts the AI buying question to cost per task

Per token, Neon looks mid-priced. Google says per finished task it is cheaper, citing third-party benchmarks.

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Google's Gemini 4 Neon shifts the AI buying question to cost per task
In brief
  • Google says Gemini 4 Neon costs 32% less per task than its Carbon model at its default thinking level. These are vendor-reported figures.
  • The rate card is $3 per million input tokens and $15 per million output tokens, so per-token comparisons with rivals can mislead.
  • Test cost per finished task on your own workloads, decide who sets the thinking level, and limit what agents can read and do.

The price on the rate card is not the price you pay

Google's new Gemini 4 Neon model charges $3 for every million input tokens and $15 for every million output tokens. A token is a small chunk of text that a model reads or writes. VentureBeat reports that these rates are 50% higher than those of Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.2 Sol. It also reports they are 25% below Anthropic's flagship Opus 5.5, which it lists at $4 and $20.

Per token, Neon looks mid-priced. Google instead points to cost per task.

Google says that at its default high thinking level, Neon costs 32% less per task than its own Gemini 4 Carbon. At that same level, it says Neon costs roughly 50% to 80% less per task than top competing frontier models. At xhigh, Google says Neon still costs 36% or more less per task than rival models at their highest settings. Google says these claims rest on third-party benchmarks. WebPulse has not tested them.

The dial moves a cost decision to a new desk

The lesson here is that AI is now bought by the finished job, not by the word. A model that thinks harder uses more tokens. A model that thinks less is cheaper. Google says Neon has four thinking levels so teams can "dial intelligence and token spend" to the workload.

That changes who makes the spending decision. A procurement team can negotiate a rate card. Rate-card negotiation does not control which thinking level an engineer picks for a routine task. The cost now sits in how each workflow is configured.

How the model is built to run long jobs

Google describes Neon as built for long-running agent work, where a model uses tools and takes many steps. It can act alone or as an orchestrator that directs parallel subagents in Google Antigravity, Google's coding platform. Google says it cuts redundant file reads and handles context compaction reliably. Compaction means summarising earlier work so a long session still fits in the model's memory.

Google lists a 1 million-token context window. It also prices cached input at $0.15 per million tokens, a 95% discount. For work that rereads the same documents, that matters more than the headline rate.

Most of the evidence comes from the vendor

Google ranks Neon third overall on the Artificial Analysis Intelligence Index, with a score of 61 at xhigh. It places second on the Vals Index at 70.9%. Its other cited results include 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0, at the default high setting. VentureBeat adds that Neon scored within about a percentage point of Opus 5.5 on two engineering benchmarks.

Speed is thinner. VentureBeat notes that Google gives no absolute completion times for a matched Neon and Carbon workload. The nearest thing is a customer quote. Glean's AI product lead, Thai Tran, said early tests found Neon matched or beat Carbon on overall quality, at triple the speed and 60% of the cost. That is one customer's early result.

The same agent reads the open web

Cheaper agents are agents that run more often, and they read web pages, emails and documents. That exposes them to indirect prompt injection, where an attacker plants instructions in content the agent reads. On Gray Swan's test for this attack, Google says Neon withstood every attempt when real-time safeguards were on and thinking was set to its highest level. That is a 0.0% attack success rate on one benchmark. It does not show the same result in a live deployment.

Google also reports cyber results. Neon scored 51.26% on WebExploitBench v1.2 and 90.5% on SEC-bench Pro, which tests finding code flaws. Google plans to give a cyber-specific edition to vetted defenders in the Fairwind Program, starting within days. The company says it added real-time safeguards against cyber misuse. On Terminal-Bench 4.0, it says the model declined just 1.98% of tasks.

Questions to put to your team

Ask what a finished task costs on your own workloads, not on a benchmark. Run the same job at two thinking levels and compare cost and quality.

Ask who owns the thinking-level setting, and whether anyone reviews it. Ask what each agent is allowed to read and do, and what happens if a web page tells it to do something else.

VentureBeat makes a useful point: the strongest case for Neon is on work where people can check the result. Start there.

A lower price per token is easy to read. A lower price per finished task has to be measured.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Google.