SIAgentNews.com
← All news

Google launches Gemini 4 Neon, a cheaper rival to Claude Opus 5.5

Google's new mid-tier model is live for users and developers on day one, priced at $3 per million input tokens, and it posts strong scores against Anthropic and OpenAI.

· First reported by Ifanr

61Neon's Artificial Analysis index score, versus 59 for Opus 5.5

The short version

  • Neon launched three weeks after Gemini 4 Carbon and is available the same day in the Gemini app, AI Studio, the API and Antigravity.
  • Price: $3 per million input tokens, $15 for output, with a 1 million token context window.
  • Artificial Analysis index: Neon 61, Opus 5.5 59, GPT-6.2 Sol 57. Opus 5.5 still leads on two coding tests.
  • In one blogger's test, Neon cost $3.5 for a task that cost $7.2 on Opus 5.5.
  • Most benchmark numbers come from Google's own chart, so independent checks still matter.

Why it matters

Most daily AI work does not need the single strongest model, just a very good one at a fair price. Neon is Google's attempt to be that model, and it pushes rivals toward cheaper pricing. If you pay per use, a side-by-side test on your own tasks is now worth the time.

A launch with no waiting list

Google has released Gemini 4 Neon, its new main model, only three weeks after Gemini 4 Carbon. As Ifanr reported, it came as a surprise, and it was available straight away rather than behind a sign-up queue.

You can use it in the Gemini app, in Google AI Studio (Google's developer playground), through the business API, and inside Google Antigravity, the company's coding environment. In the app, both Pro and Ultra subscribers get it.

Google also gave eligible Ultra subscribers access to Project Genie and a world model called Genie 4. Per Ifanr, it lets you generate a 3D world in your browser, walk around in it, and remix worlds shared by others.

Where Neon sits in the Gemini 4 lineup

Google's naming for this generation comes from the periodic table: Hydrogen, Argon, Carbon and now Neon. Ifanr says Google finally published a chart that explains who is who.

Hydrogen is the cheapest. It targets everyday coding and light agent work, scores 54 on the Artificial Analysis (AA) index, and costs $1.5 per million input tokens and $7.5 for output. A token is a small chunk of text, and model prices are quoted per million of them.

Carbon is for the longest and hardest projects. It scores 60 and has a standard price of $4 input and $20 output. Until November 30 it is discounted to $2 and $10. Neon sits in the middle, aimed at software engineering, knowledge work and scientific reasoning.

Neon is the odd one out. It has an extra top thinking level called xhigh, which pushes its score to 61, one point above Carbon. At $3 input and $15 output, it also costs a quarter less than Carbon's standard price.

The scoreboard against Opus 5.5 and GPT-6.2 Sol

Google's own chart puts Neon next to Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.2 Sol. According to Ifanr, the results are mixed. Opus 5.5 keeps a slight lead on Terminal-bench 4.1 and FrontierCode 2.0, two coding tests. GPT-6.2 Sol wins narrowly on reading dense charts.

Neon leads on the overall AA index, with 61 against 59 for Opus 5.5 and 57 for GPT-6.2 Sol. It also leads on agent-style coding, with 65.4% on CursorBench and 39.0% on SWE-sweep inside Antigravity. Ifanr adds math and science reasoning and security-flaw hunting to the list of Neon's stronger areas.

The widest gap is on GraphWalks, which tests retrieval across very long documents, from 256K to 1M tokens. Neon scored 95.2%, against 80.1% for GPT-6.2 Sol and 66.8% for Opus 5.5. The context window, meaning how much text the model can hold in view at once, is 1 million tokens.

What early users report on cost

The most useful evidence is about price per finished task, not price per token. A blogger named JAZII ran the same job on both models at their highest reasoning settings. Neon took 20 minutes and cost $3.5. Opus 5.5 took 18 minutes and cost $7.2.

JAZII's verdict, as relayed by Ifanr, was that Opus 5.5 still has a slightly higher ceiling, but Neon's output was the most polished of the Gemini 4 family so far, at less than half the cost. That is one person's test, not a benchmark.

Glean, a company that builds on these models, said Neon delivers results of the same quality as Carbon at three times the speed and 60% of the cost. It put the cost per task at 50% to 80% below top rivals. Ifanr also cites Google's safety report: a 0.0% cheating rate on Terminal-Bench 4.0, and shortcut-taking in internal coding tests at 2.45%, the lowest yet.

Another tester, Chubby, built 17 interactive demos in a day. In one, Neon added building types and demolition tools nobody had asked for.

Why the timing matters

Ifanr links the launch to a crowded moment. Anthropic is reportedly preparing an IPO, and both OpenAI and Google have lately put out lower-priced workhorse models aimed at developers and company workflows. The outlet suggests this is not entirely accidental, though that is its reading, not a stated strategy.

For readers, the practical point is simple. Our earlier note said to wait for published pricing, and Ifanr now reports it. Still, most scores come from Google's own chart, and the cost comparison rests on a single test. Run your three most common tasks on both models and compare what each finished answer costs.

What to watch

  • Do independent testers confirm Google's benchmark chart, especially the GraphWalks gap?
  • Does Anthropic or OpenAI respond with price cuts on its main models?
  • Does Carbon's promotional pricing end on November 30, and does Neon's price stay as announced?

Try it

Take one task you do often, run it on Neon and on your current model at similar settings, and compare quality and the total cost of each finished answer.

Sources and links

First reported by Ifanr. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news