Near-Opus 5.5 Coding Power at 25% Lower Cost, Challenging OpenAI and Anthropic

Google has launched Gemini 4 Neon, a new mid-range AI model that promises flagship-level performance at a more competitive price. Released on November 18, 2026, Neon delivers coding capabilities close to Anthropic’s Claude Opus 5.5 while charging 25% less per API token.

The model costs $3 per million input tokens and $15 per million output tokens, positioning it between Google’s lightweight Gemini 4 Hydrogen and its more powerful Gemini 4 Carbon.

Early benchmarks suggest Neon is particularly strong in software engineering, business automation, legal research, and scientific workflows. It also offers a one-million-token context window that supports multiple information formats, including video, audio, spreadsheets, charts, and PDFs.

However, the competition remains fierce. OpenAI’s GPT-6.2 Sol and Anthropic’s Claude Sonnet 5.5 are both cheaper at standard token rates, and individual benchmarks reveal that Neon doesn’t outperform its competitors in every category.

The bigger story is Google’s expanding AI ecosystem: rather than relying on a single flagship model, the company now offers different models for different workloads, potentially making its platform more attractive to developers and enterprises.

Key Points

1. Flagship-level coding performance at a lower price

On Cognition’s FrontierCode 2.0 benchmark, Gemini 4 Neon scored 46.2%, just behind Claude Opus 5.5 at 47.3%. Despite the narrow performance difference, Neon costs 25% less per input and output token.

2. Neon leads several competitive benchmarks

Google reported a 56.3% score on Zapier’s AutomationBench, compared with 42.5% for Opus 5.5 and 41.3% for GPT-6.2 Sol.

On Terminal-Bench Science, Neon achieved 70.5%, ahead of Sol at 68.0% and Opus at 63.3%.

However, Opus narrowly won Terminal-Bench 4.1, while Sol performed better at interpreting complex professional charts.

3. Faster output than Gemini 4 Carbon

At its highest reasoning setting, Neon generated approximately 110 tokens per second, compared with Carbon’s 59 tokens per second.

That makes Neon roughly 1.9 times faster in output generation, although its reported 57-second delay before producing the first token suggests that total task completion time remains an important consideration.

4. Agent infrastructure can dramatically improve coding results

One of the most revealing findings concerns Google’s SWE-sweep evaluation.

Neon scored just 12.8% when paired with mini-SWE agent software but reached 39.0% when running inside Google’s Antigravity development environment.

The underlying model remained unchanged, highlighting how much AI performance depends on the tools and agent infrastructure surrounding it.

5. Strong potential for business and enterprise applications

Google demonstrated Neon handling practical assignments such as investigating disputed bills, analyzing software vulnerabilities, researching scientific hypotheses, and generating editable spreadsheets.

In one demonstration, Neon agents reconstructed an image using 100,000 spreadsheet cells and more than 500,000 formulas, without embedding the original image or using a script.

These examples highlight its potential to deliver inspectable, usable work products rather than simply generating text.

6. Competitive pricing comes with important trade-offs

AI modelInput per 1M tokensOutput per 1M tokens
Gemini 4 Neon$3$15
Claude Opus 5.5$4$20
GPT-6.2 Sol$2$10
Claude Sonnet 5.5$2$10
Gemini 4 Hydrogen$1.50$7.50

Neon is cheaper than Opus but 50% more expensive per token than Sol and Sonnet at their standard rates.

The article notes that Neon would need to consume approximately one-third fewer tokens than those cheaper rivals to eliminate its pricing disadvantage for comparable workloads.

7. Google’s three-model strategy is taking shape

Google now positions Hydrogen for high-volume operations, Neon for demanding recurring tasks, and Carbon for complex, extended projects.

This gives developers and enterprises multiple performance and pricing options within the same AI ecosystem.

8. Multimodal capabilities could become an enterprise advantage

Neon’s million-token context window can accommodate video, audio, spreadsheets, PDFs, and charts together.

This allows users to analyze multiple types of evidence without necessarily converting everything into text first.

Google is also preparing a specialized cybersecurity version for eligible Fairwind Program participants, reporting promising results in vulnerability detection and patching evaluations.

Key Quotes

On Neon’s speed, quality, and operating costs:

“In our early evaluations, it matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost.”

— Thai Tran, AI Product Lead at Glean

This early customer assessment suggests that Neon could offer significant efficiency improvements over Google’s existing higher-end model.

On Neon’s advantages over Carbon:

“It’s cheaper, faster, and better behaved than Carbon.”

— Varun Mohan, describing his internal experience with Neon

The comment reinforces the potential appeal of a mid-range model that provides strong performance without requiring flagship-level spending.

On Google’s increasingly complicated model names:

“a little too much fun with naming”

— Logan Kilpatrick, Google DeepMind

His comment accompanies a practical distinction: Gemini app Pro subscribers can select Neon, while Carbon requires Ultra access.

On early enthusiasm surrounding Google’s performance gains:

“I’m impressed DeepMind”

— @synthwavedd on X

The reaction followed comparisons showing Neon outperforming Opus 5.5 and Sol across several benchmarks, although some pricing claims accompanying that enthusiasm required qualification.

Implications

1. The distinction between mid-range and flagship AI models is becoming less clear.

Neon’s near-Opus coding performance indicates that developers may not always need the most expensive model for technically demanding projects. Model selection could increasingly depend on the specific task rather than the product’s position in a provider’s lineup.

2. Businesses may need to evaluate the cost of completed work, not just API pricing.

Although Neon costs more per token than Sol and Sonnet, faster execution, fewer corrections, or reduced human review could make it economical for certain workflows.

The article emphasizes that this advantage still needs to be demonstrated against those competitors in real-world assignments.

3. AI agent environments are becoming critical differentiators.

The jump from 12.8% to 39.0% on SWE-sweep demonstrates that the surrounding development environment can dramatically change a model’s practical effectiveness.

For engineering teams, testing the complete agent workflow may be more informative than comparing isolated benchmark scores.

4. Google’s integrated ecosystem could strengthen its competitive position.

By offering Hydrogen, Neon, and Carbon for different workloads, Google gives customers more opportunities to operate within a single AI platform.

This could matter even as competing providers exchange leadership positions on individual benchmarks.

5. Multimodal analysis may make AI more useful for complex professional work.

Neon’s ability to process recordings, documents, charts, and spreadsheets together creates possibilities for investigating financial discrepancies, examining technical problems, and delivering work that colleagues can verify.

The emphasis on inspectable outputs also reinforces the importance of human oversight.

6. No single model is the clear winner across every business scenario.

Neon leads in several automation and scientific benchmarks, Opus retains advantages on certain coding evaluations, and Sol offers lower standard pricing and stronger complex-chart interpretation.

For enterprise buyers, the practical decision comes down to which model delivers accurate, usable results with the least total expenditure and intervention.

Ultimately, Gemini 4 Neon’s launch illustrates a broader competitive shift described in the article: the AI race is no longer just about which model scores highest, but which combination of model, tools, speed, and pricing produces the most valuable finished work.

Source: https://venturebeat.com/technology/google-launches-gemini-4-neon-with-opus-5-5-level-coding-performance-at-25-lower-api-token-prices

Share This Article

Related Post

OpenAI Finally Builds a Teen-Specific ChatGPT

OpenAI Finally Builds a Teen-Specific ChatGPT

OpenAI Finally Builds a Teen-Specific ChatGPT OpenAI...

Could AI Handle 90% of Your Accounting Tasks?

San Francisco-based fintech startup Puzzle ha...

AISQ Progress Update #11 – Digital Week Hig

AISQ Progress Update #11 – Digital Week Highlights, a...

Leave a Comment

Prove your humanity: 0   +   4   =  

I'm a paid user

(I’ve purchased Next Level Marketing AI credits)

AISQ | Squirrly created this web Customer App for all of you who own licenses for AISQ’s Next Level Marketing AI, AISQBusiness, Squirrly SEO, Hide My WP Ghost and more.

Read More about Customer App by AISQ | Squirrly, on the Squirrly Company’s official website

I'm a free user

(I haven’t purchased any Next Level Marketing AI credits)