Google's Gemini 4 Neon tops a coding-agent ranking at $3 per million input tokens
Google DeepMind's new mid-sized model leads one coding ranking and long-document tests, but the benchmark scores come from Google itself.
The short version
- Neon sits between Gemini 4 Carbon and Gemini 4 Hydrogen in Google's new model family.
- It scores 61 on the AA Intelligence Index, ahead of Claude Opus 5.5 (59) and GPT-6.2 Sol (57).
- Price: $3 per 1M input tokens and $15 per 1M output tokens; the weights are closed.
- It trails Opus 5.5 on FrontierCode 2.0 (46.2% vs 47.3%) and on Terminal-bench 4.1.
- Most benchmark numbers are reported by Google and have not been independently reproduced.
Why it matters
If you work with long documents, big codebases or automated workflows, Neon is a new mid-priced option that Google says costs less per finished task than rivals. The claims deserve a test on your own material before you rely on them, since most scores come from Google.
What Google released
Google DeepMind, Google's AI research lab, has released Gemini 4 Neon. As marktechpost.com reported, it is a reasoning model, meaning one built to work through a problem in steps before it answers. It is aimed at coding agents, office-style knowledge work and science tasks.
Neon is the middle model of the Gemini 4 family. Gemini 4 Carbon takes on the hardest long-running projects, and Gemini 4 Hydrogen handles quick everyday jobs. Neon is meant to balance depth against cost. According to the model card, it was adapted from Carbon.
Google has not shared how large the model is or how it is built. It accepts text, images, audio and video, and its knowledge stops at March 2026.
How it handles long jobs
Neon's context window is 1M tokens. A context window is how much material a model can hold in view at once, and tokens are the small chunks of text it reads and bills by. Its default answer length is 64K tokens, and a background mode can stretch output up to 1M.
The model has several thinking levels, which control how much effort it spends reasoning. High is the default. The top setting, xhigh, is meant for the hardest engineering, analysis and science work.
For long agent runs, Neon can compress its own working history, a step called context compaction, and it avoids reading the same files again and again. An agent here means a model that takes actions, such as editing code, rather than only replying. In Google Antigravity, Google's coding tool, Neon can act as a manager that sends smaller tasks to several Gemini 4 Hydrogen helper agents running in parallel.
What the scores say
Google compared Neon with Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.2 Sol. The company reports these numbers itself; they were not independently reproduced. The exception is the AA scores for the two rivals, which match Artificial Analysis.
- Overall: 61 on the AA Intelligence Index, against 59 for Opus 5.5 and 57 for GPT-6.2 Sol.
- Coding agents: first place on the AA Coding Agent Index at 58.6.
- Long context: 95.2% on GraphWalks, which tests material from 256K to 1M tokens. Opus 5.5 got 66.8% and GPT-6.2 Sol got 80.1%.
- Office automation: 56.3% on Zapier AutomationBench, against 42.5% for Opus 5.5.
Neon does not win everywhere. Opus 5.5 edges it on Terminal-bench 4.1 (65.8% vs 65.2%) and on FrontierCode 2.0 (47.3% vs 46.2%). GPT-6.2 Sol leads on Chartography (75.1% vs 73.5%). Neon ranks second on the Vals Index (70.9%) and second on the Agent Arena index (1336).
Price and the fine print
Neon costs $3 per 1M input tokens and $15 per 1M output tokens, with a cached-input rate of $0.15. You pay for what you send and for what the model writes back, so feeding in huge piles of documents adds up.
Per token, that is more than GPT-6.2 Sol ($2 and $10) and less than Claude Opus 5.5 ($4 and $20). Carbon is $2 and $10 until November 30, then $4 and $20.
Google's pitch is about cost per finished task rather than per token. Its own figure is that, at the default high setting, a job on Neon costs 32% less than on Carbon. It also claims 50% to 80% lower cost per task than rivals on two benchmarks. As the source notes, those are task-level claims, not price-list comparisons.
Neon is closed and hosted only by Google. You cannot download it or run it on your own computer.
Safety figures and where to find it
Google also published safety numbers, again its own. It reports a 0.09% misaligned-action rate on its agent safety suite and a 99.2% pass rate on an honesty test for agents. On Gray Swan's prompt-injection benchmark at the highest thinking level, the attack success rate was 0.0%. A prompt injection hides instructions in content so that a model obeys them by mistake.
On a test of whether a model cheats on coding tasks by taking shortcuts, Neon did so 2.45% of the time. Google says that is fewer than any other Gemini 4 model.
Developers can reach Neon through the paid Gemini API in Google AI Studio and Android Studio, and through Gemini for business. Paid Google AI Pro and Ultra subscribers get it in Antigravity, and Pro and Ultra subscribers get it in the Gemini app.
What to watch
- Will independent testers reproduce Google's reported scores, especially the GraphWalks and SWE-sweep results?
- Will cost per task hold up against rivals in real work, given that Neon's per-token price is higher than GPT-6.2 Sol's?
- How will Opus 5.5 and GPT-6.2 Sol respond on price or coding performance?
Try it
If you have Gemini Pro or Ultra, paste a long document and ask about one detail you know is buried in the middle. A big context window means the model can accept the text, not that it catches every line.
Sources and links
First reported by marktechpost.com. 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.