Gemini 4 Neon is sold on speed. The independent harness times it as the slow one
Google sells Gemini 4 Neon on fast end-to-end task completion. Artificial Analysis clocks 21.5 minutes and $6.44 per task, near the bottom of 94 models.

Google released Gemini 4 Neon on 18 November, the third new Gemini 4 model in nine days, and Koray Kavukcuoglu's launch post leads with the promise: frontier-level intelligence with fast end-to-end task completion. The price is $3 per million input tokens, $0.15 cached, $15 output, with 1M tokens of context and four thinking levels.
The intelligence claim survives an independent check. The speed claim does not.
Nine days, three elements, nothing resembling a ladder
Carbon arrived on 10 November, Hydrogen on 12 November, Neon on 18 November: three tiers of one family inside nine days. Per 9to5Google, Neon does not sit alongside Argon in the consumer app, it replaces it. Gemini 3.1 Pro is retired, 3.5 Flash-Lite becomes Auto, and the picker collapses to four rows: Auto, Everyday help as 4 Hydrogen, Deeper analysis as 4 Neon, Code and complex work as 4 Carbon.
Carbon itself is eight days old, and Google retires 3.1 Pro in the same release. The Hacker News thread collected a thousand points and 752 comments, most about vocabulary rather than capability. kristjansson: "So now we've got to get a periodic table out to understand Gemini releases?" Julian, top comment at 9to5Google with 27 likes: "Honestly it's getting difficult to keep up with all these names."
There is a defensible reading. WarmWash: "Neon is the long awaited successor to 3.1 pro." The elements are job descriptions rather than a size ladder, and Neon's job is the one Pro held.
Most of that scoreboard belongs to other people
The benchmark table is unusually well behaved for a launch, because most rows belong to outsiders:
| Benchmark | Google's figure | Operator |
|---|---|---|
| Artificial Analysis Intelligence Index, xhigh | 61, "#3 overall" | Artificial Analysis |
| Vals Index | 70.9%, second | Vals |
| Vals Legal Research, Tax Agent Bench | first on both | Vals |
| Zapier AutomationBench | 56.3%, first | Zapier |
| Arena Agent Arena Index | 1336, second over 4,200+ sessions | Arena |
| Vending-Bench 2 | $16,843 | Andon Labs |
| Terminal-Bench 4.1 and Science | 65.2% and 70.5% | third-party harnesses |
| Agents' Last Exam, HLE-Diamond, Chartography, RiemannBench | 75.8%, 71.3%, 73.5%, 84.0% | third-party harnesses |
Two rows matter more than the rest. Leading Zapier's automation bench, and taking second on an index built from 4,200 recorded agent sessions, are claims about finishing work rather than answering questions.
The 61 is real and the independent page agrees exactly. The rank is where the documents part: Google says third overall, the Artificial Analysis profile shows sixth of ninety-five, behind GPT-6.2 Astra at 62. The claim that travels nowhere is the internal one, a "lowest fabrication rate yet" measured in Google Antigravity by Google.
Twenty-one minutes, six dollars, 283 million tokens
The same page prints what the post omits. Time to first answer token: 56.51 seconds. Output speed: 110 tokens per second. Time per Intelligence Index task: 21.5 minutes, 72nd of 94, against a 12.4-minute median. Cost per task at list prices: $6.44, 78th of 94. Tokens generated to finish the index: 283M, 92nd of 94 on verbosity, against a 65.4M median.
Artificial Analysis rarely editorialises, but it does here: among the leading models in intelligence, "somewhat expensive when comparing to other models of similar price. It's also notably slow per task and very verbose." For $3 and $15 per million you buy a model that thinks for the better part of a minute, then writes four times as much as its peers at 110 tokens a second.
The comparison that matters sits on the same leaderboard. GPT-6.2 Astra scores 62 and costs $4.00 per task. Neon scores 61 at $6.44. One point less intelligence, 61% more money.
The 32% saving, and the 17% one
Google's argument is not that Neon is cheap. It is that Neon is cheap against Carbon, and the Antigravity blog puts a figure on it: at default High, a 32% lower cost per task than Carbon.
Half of that holds, and the arithmetic shows why. Carbon is listed at $4 input and $20 output, Neon at $3 and $15, a quarter less per token. Carbon runs 37.5 minutes and $7.76 per task at high effort; Neon runs 21.5 minutes and $6.44 at xhigh. That is a 17% saving, not 32%, and it conceals a detail: Neon emits 283M tokens where Carbon emits 205M. Neon is cheaper because it is charged less per word, not because it is more concise.
Which weakens the framing 9to5Google borrowed. Carbon quality at Flash speeds, its headline says. Quality, near enough. Flash speeds belong to Gemini 4 Hydrogen, the model Google labels the Flash replacement and still parks in the everyday slot.
Four dials, one honest default
Neon offers Low, Medium, High and Extra High reasoning, defaulting to High. Artificial Analysis scores each level: 44, 55, 60, 61. The default is within a point of the maximum, so the extra level buys a rounding error at the deepest token burn the model offers. Teams that leave the dial alone get most of the model and least of the invoice.
The caching discount deserves the same reading. Cached input at $0.15 against $3 is the advertised 95% cut, and the developer pricing page adds a line the blog does not: for Neon, caching is implicit only, with no explicit cache creation. A real discount for apps with stable system prompts, nothing for everyone else.
Who gets it, and what to check first
Neon ships in the Gemini API and AI Studio, in Gemini for business, in Antigravity for AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra. In Antigravity, Orchestrator can pair Neon as planner with Hydrogen as executor in one session, handing routine edits down and pulling the task back when Hydrogen hits friction. That pairing is the most practical thing in the release, and also an admission that one model does not suit a whole task.
SimpleBench cuts Google's way. Neon sits second at 94.4% over five runs, above Carbon's 92.6% and the 91.7% shared by Argon and Hydrogen, behind Claude Fable 5.2 at 94.9%. The human baseline is 83.7% and the best human 95.4%, so Neon is bumping the ceiling of a deliberately nasty test of everyday reasoning.
The honest summary is narrower than either framing. Neon is near-frontier at a mid-lineup price, cheaper than Carbon by 32% on Google's tasks and 17% on someone else's, and sluggish next to Hydrogen. For long analytical work where 21 free minutes per task is nothing, watch cost per task. For anything interactive, run twenty of your own requests through both endpoints and count tokens, not index points.


