Pricing depends on the model and the task
According to Tai Tran, head of AI products at Glean, early tests found Neon matched or surpassed Gemini 4 Carbon in overall quality while running three times faster and costing 60% as much. Google has not given exact runtimes for Neon and Carbon on identical tasks; the clearest speed comparison comes from a customer.
Neon’s rates per million tokens — $3 for input and $15 for output — are 50% higher than the standard prices for Anthropic’s mid-sized Claude Sonnet 5.5 and OpenAI’s GPT-6.2 Sol. Google says, however, that at the highest reasoning level Neon outperforms both models on the Artificial Analysis Intelligence Index. Compared with Anthropic’s larger flagship model, Opus 5.5, Neon costs 25% less: Opus’s standard rates are $4 per million input tokens and $20 per million output tokens. The models’ results on two engineering benchmarks differ by about one percentage point.
Google positions Neon between Hydrogen, a smaller and faster model, and Carbon, which it recommends for the most difficult, long-running projects. But the results do not fully match that hierarchy: at the highest reasoning level, Neon slightly outperforms Carbon on the overall Intelligence Index, while the three models’ results on professional Vals tasks differ by no more than 2.5 percentage points.
Google published Neon’s Vals result. Artificial Analysis confirmed on X that the model ranked third on its Intelligence Index. Competitors’ results come from the Artificial Analysis table and the Vals index; the published snapshot of the latter predates Neon’s release.
The Artificial Analysis table lists rates for input, cached input and output tokens. For Carbon, they are $2, $0.10 and $10 until November 30, after which they are set to change to $4, $0.20 and $20. The reasoning levels refer to the Artificial Analysis column. Companies set them differently, so two models at their highest level may spend different amounts of time and money on the same question.
The comparison also changes depending on the date and prompt length. Until the end of November, Carbon’s introductory rate makes it cheaper than Neon per token. From December 1, Neon’s input and output tokens will be 25% cheaper than Carbon’s planned rates. OpenAI has separate pricing for long prompts: if a request to Sol exceeds 272,000 input tokens, the entire request costs $4 per million input tokens and $15 per million output tokens. At that length, Sol loses its advantage on the standard output-token price. Cursor lists requests of more than 256,000 input tokens and up to 500,000 for Grok 4.7 Fast.
Professional tasks show a different picture
Vals combines tasks in finance, software, law and tax, weighting industries according to their contribution to the US economy. Google says Neon ranked second overall, with 70.9%, and first in legal research and tasks for tax agents.
A law firm cares more about performance on legal tasks than the overall score. And for a buyer choosing a model for several departments, the published scores offer little reason to distinguish Neon from Sonnet on overall accuracy alone.
Sol offers a different trade-off. The Vals index lists the cost of one test at $3.18 for Sol and $16.18 for Sonnet. Sol trails Sonnet by almost five percentage points, even though the two models have the same standard prices for short prompts. This shows how much cost can depend on a model’s behavior while working. Google does not provide a comparable Vals test cost for Neon, leaving an important part of the three-model comparison unresolved.
Generation speed and waiting time
According to benchmark tracker Priors, on launch day Neon generated about 110 tokens per second at xhigh. Artificial Analysis lists Carbon at 59 tokens per second at high, its highest level. By those figures, Neon is about 51 tokens per second ahead of Carbon, or roughly 1.9 times faster.
Neon’s generation speed is close to Gemini 3.8 Flash, which produces 116 tokens per second at high. Hydrogen is faster: 167 tokens per second at high and 171 at xhigh.
Fast generation does not mean the response starts arriving immediately. Priors lists Neon’s time to first token at xhigh as about 57 seconds. If a developer is waiting for a fix in a repository, that time is followed by the agent reading files, running tests and refining its changes.
Google separately says that at the standard high level, Neon’s task cost is 32% lower than Carbon’s, according to the Artificial Analysis index. At xhigh, where Neon scored 61, its cost advantage over Carbon at high falls to about 17%.
In programming, the environment around the model matters
The Cognition FrontierCode 2.0 benchmark evaluates code changes against criteria set by reviewers. Google says Neon scored 46.2% at the standard high level. In Google’s comparison, Opus scored 47.3%, 1.1 percentage points higher.
For comparison, in Artificial Analysis results compiled by The Aggregate, Sol scored 44.38% and Sonnet 41.7%; both models were tested at their highest level. These scores show the general range of results, but the configurations differ. Google’s own data illustrates how much settings and the software environment can change the outcome.
In a separate programming test, SWE-sweep, Google reports that Neon scored 12.8% with the mini-SWE agent software and 39.0% with Antigravity. The model itself did not change.
In Antigravity, Neon can split a project among several Hydrogen agents working in parallel. For an engineering team using its own agent software, a threefold difference in these results is a reason to test Neon in Antigravity rather than rely only on its benchmark ranking.
What the first reactions showed
On launch day, Varun Mohan wrote on X that, based on his experience using Neon internally, it was cheaper, faster and behaved better than Carbon. Omar Sanseviero, head of developer relations at Google DeepMind, separately highlighted the shorter wait for the first token — a difference developers notice before an agent has finished working.
Google DeepMind technical team member Logan Kilpatrick acknowledged that the company had got carried away with model names and shared a practical clarification for subscribers: Pro users can select Neon in the Gemini app, while Carbon is available only to Ultra subscribers.
The @synthwavedd account pointed to Neon’s advantage over Opus 5.5 and Sol in most rows of Google’s table. The post also called the model cheaper than both alternatives, but that needs a qualification for Sol: its standard token prices are lower. The table compares performance, while cost depends on the task the agent is doing.
A table shared by DeepMind co-founder Demis Hassabis makes Neon’s advantage in certain tests especially clear. In Zapier’s AutomationBench, the model scored 56.3%, compared with 42.5% for Opus and 41.3% for Sol. In Terminal-Bench Science, which a Google representative highlighted in an email to VentureBeat, Neon scored 70.5%, Sol 68.0% and Opus 63.3%.
The ranking shifts on other tasks. On Terminal-Bench 4.1, Opus narrowly beat Neon, 65.8% to 65.2%. In Chartography, which evaluates reading complex charts, Sol scored 75.1% and Neon 73.5%.
Google returns to the frontier-model market
Seven weeks ago, Google introduced Gemini 4 Argon. According to results published by the company, the model beat GPT-6 Astra and Claude Opus 5.5 in 12 of 18 tests Google selected. Most developers could only look at the scores at the time: access initially went only to approved cybersecurity partners.
Then, on October 28, Carbon launched. With access similarly restricted, it took first place on the updated Artificial Analysis Intelligence Index. By the time Google opened Carbon to paying customers on November 10, Anthropic’s Claude Fable 5.2 had already overtaken it.
Google now offers models for three kinds of tasks and budgets: Hydrogen for high-volume work, Neon for more complex recurring tasks, and Carbon for long-running projects. On Neon’s launch day, developers gained access to another Google model rather than just the results of closed tests.
Neon’s release coincided with developers’ interest in OpenAI’s new GPT-6.2 Astra and Sol. As VentureBeat reported the day before the launch, Every CEO Dan Shipper found that Astra sometimes finished end-to-end builds prematurely, although it performed at the level of its predecessor in most coding-team tests. Ethan Mollick of the Wharton School of Business managed to get Astra to create a space elevator game, but noted that the model was difficult to direct.
Those observations apply specifically to Astra. Sol remains a cheaper alternative to Neon. Early feedback suggests that all three models can be compared by how much intervention it takes to turn a promising first attempt into finished work.
What tasks might justify the higher price
One difference with Neon is its support for mixed inputs. Google says its one-million-token context window can take video, audio, charts, spreadsheets and PDFs. For example, an analyst examining a billing dispute could give the model a call recording, a contract and an accounting ledger.
Sol accepts text and images through its API and supports a 1.05-million-token context window, but a team would need another tool to process a recording. Sonnet also has a one-million-token window, so capacity alone is not enough to distinguish these three models.
Google is preparing a separate access process for security teams. A specialized cyber version of Neon is expected to become available to eligible Fairwind Program participants in the coming days. Google says it scored 90.5% on a vulnerability discovery test and 70.7% on a vulnerability remediation test. Companies can request access through their Google Cloud representative.
With the same ratio of input to output tokens, Neon would need to use about a third fewer tokens than Sol or Sonnet to offset its 50% premium over their standard short-prompt prices. A workflow that finishes sooner or requires less human review could also justify the difference. Glean reported this kind of advantage for Neon over Carbon. Whether Neon is more cost-effective than Sol or Sonnet depends on what a team spends to get a finished result.
Dataist’s take: Neon should be compared with competitors by the cost of completing a task, not just by token prices or benchmark rankings. SWE-sweep results vary threefold depending on the software environment, and Google has not disclosed a comparable Vals cost for Neon. Access is another limitation: Carbon requires an Ultra subscription, while Neon is available to Pro subscribers too.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X