SIAgentNews.com
← All news

Google releases Gemini 4 Neon with faster speeds and lower costs

The new model offers a slight intelligence boost over its predecessor while cutting task costs by 17% and completion times by 40%.

· First reported by artificialanalysis.ai

40%reduction in task time compared to Carbon

The short version

  • Neon scores 61 on the Intelligence Index, placing it 3rd overall behind GPT-6.2 Astra and Claude Fable 5.2.
  • The model completes tasks ~40% faster than Gemini 4 Carbon, taking about 21 minutes per task.
  • Neon costs $6.44 per task, which is roughly 17% cheaper than the standard pricing for the Carbon model.
  • It achieves a 1771 Elo score on the AA-Briefcase, leading in Presentation Quality over Claude Sonnet 5.5.
  • The model supports a 1M token context window and handles text, image, video, and speech inputs.

Why it matters

For developers and enterprises, Neon provides a way to achieve high-level intelligence with significantly lower latency and overhead. It positions Google as a leader in 'agentic' work, where models must perform complex multi-step reasoning.

A new balance of speed and cost

Google DeepMind has introduced Gemini 4 Neon, marking its fourth model release in a span of only seven weeks. This rapid release cadence suggests a push to provide users with more specialized options for different types of workloads. Neon is designed to offer a small intelligence gain over the previous Gemini 4 Carbon model while optimizing for efficiency.

The model achieves a score of 61 on the Artificial Analysis Intelligence Index. This puts it just one point ahead of Gemini 4 Carbon (high) and ranks it third overall in the current landscape. It sits behind only GPT-6.2 Astra (max) and Claude Fable 5.2 (max) in terms of raw intelligence metrics.

One of the primary goals for Neon appears to be making high-level reasoning more accessible. By balancing performance with lower operational costs, Google is targeting users who need reliable output without the heavy price tag of the largest models. It provides a middle ground for complex tasks that require more than basic processing but don't necessitate the highest tier of compute.

Technical performance and pricing

Neon delivers a significant jump in speed and a reduction in cost per task. It takes approximately 21 minutes to complete a task, which is about 40% faster than the 36 minutes required by the Carbon model. This speed increase is driven by a higher output rate of roughly 110 tokens per second, compared to 57 tokens per second for Carbon.

The cost structure for Neon is also more favorable for high-volume users. It costs $6.44 per task, which is about 17% less than the standard $7.76 price for Carbon. This savings comes from a 25% lower token price of $3 for input and $15 for output per 1M tokens, rather than the $4 and $20 standard for Carbon.

While the cost is lower, the model actually uses about 15% more output tokens per task than its predecessor. This means the efficiency gains are a result of optimized pricing and faster generation speeds rather than a reduction in the amount of data produced. Users can also benefit from a 95% discount on cached input tokens, which currently cost $0.15.

Advancements in visual reasoning

Google DeepMind has highlighted specific improvements in how Neon handles visual data and presentation. In the GDP.pdf evaluation, Neon (high) scored 33.4%, which is 5 points higher than Carbon (high). This puts it in line with the highest scores currently measured for visual reasoning tasks.

The model also shows strength in presentation quality. On the AA-Briefcase benchmark, Neon (xhigh) earned a 1700 Elo score overall. It specifically led the field in Presentation Quality with a score of 1771 Elo, beating the previous leader, Claude Sonnet 5.5 (max), by 45 points.

These improvements are particularly relevant for multimodal knowledge work. The Multimodal Knowledge Agent (MKA) benchmark tests how models handle complex visual files. Neon reached a Climb WSR of 56% on this test, showing a 2-point improvement over Carbon and performing on par with the current leader, Claude Fable 5.2 (max).

Coding and agentic capabilities

Neon is also proving to be a competitive tool for developers building coding agents. When tested with the Antigravity CLI, Neon (xhigh) scored 59 on the Artificial Analysis Coding Agent Index. This is 3 points higher than the score of 55 achieved by Gemini 4 Carbon (high).

When compared to other industry leaders, Neon holds its own in coding tasks. It scored in line with Claude Code, which received a 58 from Claude Fable 5.2 (xhigh), and Claude Opus 5.5 (max), which also scored 58. These results suggest that Neon can handle complex programming logic effectively.

Furthermore, Neon offers a significant cost advantage for these coding workflows. In the Antigravity environment, Neon has an average cost of $6.86 per task. This is considerably cheaper than the $12.61 cost for Claude Opus 5.5 (max) and the $15.47 cost for Claude Fable 5.2 (xhigh). This makes it a more economical choice for high-frequency automated coding tasks.

Availability and effort levels

Google is making Gemini 4 Neon available through several different channels at launch. It is publicly accessible via Google’s API and the Gemini Enterprise Agent Platform. Additionally, Pro and Ultra subscribers can access the model within the Gemini app and the Antigravity CLI.

A unique feature of the Neon release is the inclusion of four different effort levels: low, medium, high, and xhigh. The xhigh level was recently introduced to the Gemini family during the Hydrogen release. This allows users to toggle the amount of computational power and reasoning time allocated to a specific prompt.

The model supports a context window of 1M tokens, allowing it to process large amounts of information in a single session. It is fully multimodal, accepting text, image, video, and speech as inputs while providing text as the output. This versatility ensures it can be integrated into a wide variety of professional and creative workflows.

What to watch

  • How the 17% cost reduction affects enterprise adoption of Gemini 4.
  • Whether future updates to Gemini 4 Carbon will close the speed gap with Neon.
  • How other providers respond to Google's rapid seven-week release cadence.

Try it

Check the Artificial Analysis Intelligence Index to see how Neon compares to other models like GPT-6.2 Astra.

Sources and links

First reported by artificialanalysis.ai. 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.

More AI news