Models

Gemini 4 Neon runs at 110 tokens per second, undercuts Claude

Google DeepMind's Gemini 4 Neon reaches 61 points on Artificial Analysis, runs at 110 tokens per second and costs $6.86 per coding task, undercutting Claude Fable 5.2.

Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks ago
Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks agoShutter Speed / unsplash
By Rebecca Stone5 min read

Why it matters

  • Google DeepMind released Gemini 4 Neon, its fourth model in seven weeks.
  • Neon scored 61 on Artificial Analysis at xhigh, ranking third behind GPT-6.2 Astra and Claude Fable 5.2.
  • Artificial Analysis measured Neon at about 110 tokens per second, nearly double Carbon's 57.
  • Neon costs $6.44 per task at xhigh, about 17 percent below Carbon's undiscounted $7.76.
  • An average coding task costs $6.86 with Neon, versus $12.61 for Claude Opus 5.5 and $15.47 for Claude Fable 5.2.

Gemini 4 Neon generates about 110 output tokens per second, nearly double Gemini 4 Carbon's 57 tokens per second, according to Artificial Analysis.

Google DeepMind released Neon as its fourth model in seven weeks. The model supports a one-million-token context window for text, images, audio, and video, and Google positions it between Carbon and Hydrogen. The Decoder reports that Neon beats Carbon on most benchmarks and costs less overall.

For teams buying model access or building coding agents, Neon combines higher throughput with lower per-task charges than Carbon and lower coding-task costs than Anthropic's Claude models. The figures sharpen the commercial contest over the cost of each completed task, not just the sticker price per token.

Where does Neon rank on intelligence?

Artificial Analysis scored Neon 61 points on its Intelligence Index at the highest reasoning level, "xhigh." Carbon scored 60, leaving Neon third in the ranking behind OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2.

Google omitted GPT-6.2 Astra and Claude Fable 5.2 from the comparison table for its own release, The Decoder said. Their absence does not change the third-place result in Artificial Analysis's ranking, but it narrows the set of rivals shown directly against Neon in Google's material.

Neon improves on Carbon by one point at "xhigh," while Google's hierarchy places it below Hydrogen. The result shows a modest internal gain, not an outright lead over OpenAI or Anthropic in the independent ranking.

Why is Neon cheaper than Carbon?

The Decoder makes the pricing distinction explicit: "The Cost Advantage Comes from List Price, Not Token Efficiency."

Google AI Studio and Gemini for Business list Neon's current prices as follows:

  • $3 per one million input tokens.
  • $0.15 per one million cached input tokens.
  • $15 per one million output tokens.

Those prices are 25 percent below Carbon's regular rates of $4 for input and $20 for output. They are also 50 percent above Carbon's promotional pricing, which Google says will remain in effect through the end of November.

Artificial Analysis put Neon's per-task cost at $6.44 at "xhigh," about 17 percent below Carbon's undiscounted $7.76. Neon used roughly 15 percent more output tokens per task than Carbon. The lower task cost therefore comes from the lower per-token list price, not from fewer generated tokens.

At the second-highest reasoning level, "high," Neon matched Carbon's intelligence score while halving per-task cost, according to Artificial Analysis. The benchmark run also shortened from about 36 minutes to roughly 21 minutes. Neon generated about 110 tokens per second during that run, compared with Carbon's 57 tokens per second.

The pricing structure gives buyers two different comparisons. Neon wins on undiscounted rates today, while Carbon's promotion remains cheaper on a per-token basis through the end of November. The 17 percent task-cost saving therefore applies against Carbon's regular rate, not its current promotional rate.

What does Neon cost on coding-agent work?

Inside Google Antigravity, Neon at "xhigh" scored 59 on the Coding Agent Index. Carbon scored 55, while Claude Fable 5.2 and Claude Opus 5.5 each scored 58.

The average coding-task costs were:

  • Gemini 4 Neon: $6.86.
  • Claude Opus 5.5: $12.61.
  • Claude Fable 5.2: $15.47.

The price gap is wider than the score gap. Neon leads both Anthropic models by one point, but its average task cost is less than half Fable's and roughly 46 percent below Opus's. For teams running repeated agent loops, the measured task price carries more weight than the one-point intelligence advantage.

Google said the Antigravity environment tripled Neon's hit rate to 39.0 percent on the SWE-sweep test. That result is tied to the named coding environment, making Antigravity part of the measured setup rather than a neutral test label.

How strong are Neon's visual and science results?

Artificial Analysis found that Neon at "high" achieved the top score on its GDP.pdf document test, at 33.4 percent. That result placed Neon five points ahead of Carbon. At "xhigh," Neon's score fell to 28.0 percent. The reversal warns buyers not to treat the highest reasoning setting as uniformly better across workloads.

On AA-Briefcase, Artificial Analysis placed Neon first for presentation quality with 1,771 Elo points, ahead of Claude Opus 5.5. The result supports Neon's visual-processing standing, although the source did not provide a separate task price for the AA-Briefcase comparison.

Google's own evaluations put Neon at the top of individual categories on the Vals Index. Google also reported 56.3 percent on Zapier's AutomationBench and 84.0 percent on the RiemannBench science test. These scores extend the case beyond coding and document analysis, though they remain Google's measurements rather than independent results.

What does Google claim on cyber safety?

Google is launching a cyber variant of Neon through the Fairwind Program. The company says the variant earned top scores on two security benchmarks at half the price of competing models:

  • SEC-bench Pro: 90.5 percent.
  • WebExploitBench v1.2: 51.26 percent.

Neon's brief Model Card says the model "blocks 100 percent of indirect prompt-injection attacks on the Gray-Swan benchmark." The claim covers one class of indirect prompt injection on one named benchmark. For security buyers, that narrow scope matters more than any blanket claim about general cyber safety, and the Model Card provides no broader safety score in the source.

Neon's next test is whether its benchmark advantage holds in other coding workflows and document tests. Artificial Analysis already records one reversal, with "xhigh" trailing "high" on GDP.pdf, while Carbon's discount expires at the end of November and changes the direct price comparison.

Original: the-decoder.de

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

802 articles

Related articles

  1. Gemini 4 Neon cuts task cost while intelligence gain stays modest
  2. Google Ships Gemini 4 Neon, a Cheaper Near-Carbon Model at $3 Per Million Tokens
  3. Google Opens Gemini 4 Neon to All Developers at $3/$15 per Million Tokens
  4. Google's Gemini 4 Carbon Posts Top AI Benchmark Score at 59
  5. Google Opens Gemini 4 Carbon to All Paying Users at $2 Per Million Tokens

« Previous articleNext article »