Skip to main content

Google's Gemini 4 Neon costs 25% less per token than Claude Opus 5.5

Ilustrační obrázek
Google's Gemini 4 Neon costs $3.00 per million input tokens and $15.00 per million output tokens. That is 25% below Claude Opus 5.5 at $4.00 / $20.00, and Neon now sits first on the Artificial Analysis Coding Agent Index with a score of 58.6. The price is the more interesting number. An agent loop that runs all day spends most of its money on tokens, not on launch-day demos.

Where Neon sits in the Gemini 4 family

Google DeepMind split the Gemini 4 generation into three named models. Gemini 4 Carbon targets long-horizon reasoning and deep work. Gemini 4 Hydrogen handles fast, cheap everyday execution. Gemini 4 Neon, launched on November 18, sits between them and replaces the earlier Gemini 3.1 Pro and Argon tiers in the lineup.

Neon's stated job is high-frequency agentic coding and complex enterprise workflows. That description matches a specific workload pattern: an assistant that calls tools, reads terminal output, edits files and repeats the loop dozens of times per task. Those loops multiply token counts quickly, which is why the pricing tier matters more here than raw benchmark position.

Pricing and context window

Input tokens cost $3.00 per million, output tokens $15.00 per million, and reasoning tokens are billed as output. Cached input drops to $0.15 per million.

The model carries a 1 million token unified context window with 64K default output tokens. Google lists throughput at roughly 110 output tokens per second, close to double Gemini 4 Carbon's ~57 tokens per second. For comparison against the rest of the field: GPT-6.2 Astra sits at $8.00 / $40.00, Claude Opus 5.5 and Gemini 4 Carbon both at $4.00 / $20.00, Grok 4.7 at $2.00 / $6.00, Mistral Large 4 at $0.70 / $2.10, GLM-5.3 FlashX at $0.37 / $1.25 and DeepSeek V4.1 Flash at $0.15 / $0.60 during off-peak hours.

Cost per task, calculated

List prices are hard to compare in your head. A single mid-sized coding request that takes 50,000 input tokens and produces 10,000 output tokens works out like this:

ModelInput costOutput costTotal per task
Gemini 4 Neon$0.150$0.150$0.30
Gemini 4 Carbon$0.200$0.200$0.40
Claude Opus 5.5$0.200$0.200$0.40
GPT-6.2 Sol$0.100$0.100$0.20
Grok 4.7$0.100$0.060$0.16
Muse Spark Open (hosted)$0.0625$0.0425$0.105
Mistral Large 4$0.035$0.021$0.056
DeepSeek V4.1 Flash (off-peak)$0.0075$0.006$0.0135

The table says two things. Neon shaves exactly a quarter off the same task on Carbon or Opus 5.5, which is the claim Google is making. It also shows that a $0.30 task is still 20 times more expensive than the same task on DeepSeek V4.1 Flash off-peak. Price cuts at the frontier compress the gap between premium models. They do not close the gap to open-weight and Chinese-hosted options.

Caching changes the arithmetic further on the Gemini side. If 40,000 of those 50,000 input tokens are cached and only 10,000 are fresh, input cost falls to $0.036 and the whole task lands at about $0.186, roughly 38% below the uncached figure. Agent loops with a stable system prompt and tool schema are the natural fit for that.

Benchmark results

On the Artificial Analysis Intelligence Index v5.0.1, Neon scores 61 at its Extra High thinking level. That puts it third overall, ahead of Gemini 4 Carbon on 60 and Claude Opus 5.5 on 59. On the Coding Agent Index it takes first place with 58.6.

Other published numbers: 65.2% on Terminal-Bench 4.1, 56.3% on Zapier AutomationBench, and 95.2% on GraphWalks at context lengths between 256K and 1M tokens. Artificial Analysis also reports a 32% lower cost per task than Gemini 4 Carbon at its default High reasoning level.

Zapier's AutomationBench figure is worth reading slowly. 56.3% means more than four in ten automation tasks were not completed in that evaluation. Benchmark scores describe relative position, not production reliability.

Speed and long agent loops

Throughput matters as much as price once an agent chains tool calls. At ~110 output tokens per second, Neon produces roughly twice the text per second that Carbon does. A twenty-step coding session that takes ten minutes of generation time on Carbon would finish in about five on Neon, assuming the tool-call overhead stays constant. Wall-clock time also translates into money when your orchestration layer bills by the minute.

Closed API models cannot run on our local AI Arena rig, so the RTX 5060 Ti numbers we publish for local Ollama models have no direct counterpart here. Our figures in this article come from list pricing and third-party benchmarks, not from our own inference runs.

Availability in Europe

Gemini 4 Neon ships through the Gemini API in Google AI Studio and Vertex AI, inside Google Antigravity for multi-agent orchestration, and on paid Google AI Pro and Ultra subscriptions. Nothing in the launch material suggests an EU exclusion, and Vertex AI has historically offered European regions, but region-by-region model availability changes without much notice. Check the Vertex model card for your specific region before you write a residency promise into a client contract.

Two EU points carry practical weight. Google bills API usage in US dollars, so European teams carry the exchange rate and add VAT on top of the listed price. And since August 2, 2026, Article 50 of the EU AI Act requires machine-readable marking of AI-generated content and clear notices when users interact with an AI system. If Neon ends up writing customer-facing copy or generating media, that obligation sits with the deployer, not with Google.

What the launch material does not cover

No independent reproduction of the Coding Agent Index score has been published yet. The index is run by Artificial Analysis rather than by Google, which helps, but a single evaluation run on a newly released model is thin evidence for weeks of production traffic.

There is also no word on whether the $3.00 / $15.00 list price holds through the rest of the year. Frontier models have had their launch pricing adjusted before, and an agent pipeline budgeted at 100 million tokens per month is exposed to that. At list price that pipeline costs $1,800 a month on Neon, against $2,400 on Opus 5.5 or Carbon.

Can I run Gemini 4 Neon on my own hardware?

No. Neon is a closed API model with no published weights, so it cannot be served through Ollama or vLLM. If self-hosting is a hard requirement, the closest options in this generation are Mistral Large 4, DeepSeek V4.1 Flash under MIT and GLM-5.3 FlashX, all of which can run under local control at far lower token prices but with weaker benchmark positions.

Is the 25% discount a temporary promotion?

Google lists it as standard API pricing, not a promotional rate, and the cached input tier at $0.15 per million is part of the same price sheet. That said, launch pricing at the frontier has been revised before, so treat the number as current rather than permanent.

Does the 1M context window work at full length in practice?

The GraphWalks result of 95.2% covers retrieval between 256K and 1M tokens, which is the closest published evidence. Long-context accuracy above that range and the cost of filling a 1M token prompt are separate questions the launch material does not address.

Discussion

No comments yet — be the first to share your thoughts.
X

Don't miss out!

Subscribe for the latest news and updates.