Google DeepMind continues cooking with gas with yet another solid model release this week: Neon.
Exactly one year after Google introduced Gemini 3 Pro, it released a model that makes a pretty good argument for spending less on Google’s own flagship. At its default reasoning setting, Gemini 4 Neon matched Gemini 4 Carbon on the Artificial Analysis Intelligence Index.
At standard API prices, Neon cost $3.74 per benchmark task. Carbon cost $7.76. Google’s Gemini 3 announcement and Artificial Analysis’s head-to-head comparison.
If you spent much of the past year waiting for Google DeepMind to make the AI race interesting again, here’s your week. On Tuesday, OpenAI released GPT-6.2 Astra and Sol. On Wednesday, Google released Neon. By the end of launch day, Artificial Analysis had ranked Neon third overall at its highest setting, behind Anthropic’s Claude Fable 5.2 and OpenAI’s brand-new Astra.
Three labs had a model on the podium. Google’s entry was the one it had just made available to people on its Pro plan. OpenAI’s release notes, Google’s Neon announcement and Artificial Analysis’s November 18 evaluation.
What does one more benchmark point buy you?
Neon lets developers choose how much effort it puts into an answer. On the default high setting, it scored 60 on the Intelligence Index, level with Carbon. Turn Neon up to xhigh and the score rose to 61. The average cost per task rose from $3.74 to $6.44. That last point cost another $2.70 per run.
For a team running the same workflow a thousand times, that’s a $2,700 reason to check whether the extra thinking actually improves the result. Artificial Analysis’s breakdown by reasoning level.
There’s a wrinkle if you’re paying the bill this month. Artificial Analysis used Carbon’s standard price of $4 per million input tokens and $20 per million output tokens. Through November 30, Google is charging half that. Apply the promotional rate to the same Carbon run and its estimated cost falls to $3.88. Suddenly, default Neon at $3.74 and Carbon at $3.88 are a close call. December changes that math. Google’s Carbon pricing terms.
The comparison with Claude gives Google a different reason to celebrate. On Artificial Analysis’s separate Coding Agent Index, Neon running in Google’s Antigravity CLI scored 59. Claude Fable 5.2 running in Claude Code scored 58. I’d call that a close finish.
The bills were much farther apart: $6.86 per task for Google’s setup and $15.47 for Anthropic’s. If your team pays agents to work through a backlog, that’s a good reason to give both the same stubborn bug and see which fix you’d actually ship. Artificial Analysis’s coding-agent results.
Google’s seven-week comeback
Neon is the fourth Gemini 4 model in seven weeks. Argon started with a limited rollout on September 30. Carbon opened to paying developers on November 10 for demanding work like debugging complex software. Hydrogen followed on November 12 for faster, everyday jobs. Free Gemini users can get Hydrogen automatically when the app selects it for their prompt.
Neon arrived November 18, giving Pro and Ultra subscribers a way to bring a spreadsheet, a PDF and a meeting recording into the same conversation and ask where the numbers disagree. The Neuron’s look at Argon, Google’s Hydrogen release and Neon’s launch details.
Google also shipped something genuinely different this week. On the same day Neon launched, Genie 4 opened to eligible Ultra subscribers. It generates interactive worlds where you can walk around, change the weather with a prompt and hear the scene change around you.
Developers got API access the following day. Google also updated Genie 3 with three-minute sessions and the ability to start from a 360-degree panorama. Google’s Project Genie update and Genie 4’s API release.
Why the rest of us win
OpenAI, Anthropic and Google all have a credible claim on your next hard assignment. Meta and xAI give buyers more ways to shop around, too. Meta’s Muse Ember matched Claude Haiku 5.5 on the Intelligence Index at roughly a third of the cost per task. And xAI’s Grok 4.7 moved its lab into the top four on Artificial Analysis’s September evaluation. Different buyers can look at those results and make different, perfectly sensible choices. Artificial Analysis on Meta’s Muse Ember and its evaluation of Grok 4.7.
For anyone paying for AI out of pocket, the simplest win is already here. If you have Google AI Pro, you can try Neon on the work that made you consider a second subscription in the first place.
Give it last month’s messy budget and ask it to find the discrepancy you already know about. See how long it takes, whether it shows its work and how much checking you have to do afterward. Google is back in the race. Whether it gets your money is a much more interesting contest.
Related reading
- Gemini 4 Argon Can Work Overnight. What Gets Finished? — Our look at the model that started Google’s seven-week run asks how to judge an AI by the work it hands back.
- OpenAI Unveils GPT-6.2 Astra and Sol — OpenAI’s releases from the day before Neon show how the same price-versus-performance decision plays out for ChatGPT and Codex users.