Gemini 4 Argon (High): Intelligence, Performance and Price Analysis
TL;DR Highlight
Gemini 4 Argon (High) scores 53 on the AA Intelligence Index but is slow, with a 48s time to first token
Who Should Read
A backend developer on LLM APIs weighing whether to switch models based on cost, speed, and quality
Core Mechanics
- Intelligence Index 53, about double the median
- It scores 53 on the Artificial Analysis Intelligence Index versus a median of 27 for similar-price models.
- Overall rank is shown as 8 of 217.
- Price matches the median
- Input is $2.00 and output is $10.00 per 1M tokens, both equal to the medians.
- Cache discount is 95%.
- $1.99 per Intelligence Index task
- That is the average cost to run one task of the index.
- Cost rank is 83 of 114.
- Output speed slightly below average
- It generates 84.2 tokens/s on Google's API versus a median of 85.4.
- Speed rank is 97 of 186.
- Time to first token is 47.97s
- TTFT is 47.97s on Google's API.
- The median for similar-price models is 3.81s.
Community Reactions
- Opinion that Google will avoid a price war
- One commenter said Google could sell at cost but everyone earning healthy margins is better game theory.
- They noted Google also sells compute and services.
- Opinion that the lower hallucination rate stands out
- A commenter called it the most impressive jump given how weak earlier Gemini models were here.
- View that it trails Opus 5.5 by 5 points
- Commenters called it an unimpressive release at 5 points behind Opus 5.5 on AA.
- Another said it trades blows with Mimo v2.6 and 6.1-Sol, which are very cheap.
- Experience of being underwhelmed in daily use
- One commenter said their Gemini subscription disappointed them so they downgraded the tier.
- Another complained that a paid plan still offers Gemini 3.1 Pro on gemini.google.com.
How to Apply
- Build a side-by-side table with your current model
- Today, list the intelligence index, input/output price, and TTFT for your model and Argon.
- Copy the article's numbers as-is (53, $2/$10, 47.97s).
- Keep it off latency-sensitive paths
- Mark real-time paths like chatbots as exclusion candidates given the 47.97s TTFT.
- Try batch and async jobs first.
- Enable prompt caching and recompute cost
- If you have a long shared prompt, recalculate monthly cost assuming the 95% cache discount.
- Validate on your own eval set
- Don't trust benchmark scores alone; run 50 of your own cases and compare accuracy and token usage.
Try It
# Paste into Claude/ChatGPT (build a model-switch comparison table)
Below are the candidate models.
- Model A: {{current model}}, index {{score}}, input ${{price}}/1M, output ${{price}}/1M, TTFT {{sec}}
- Model B: Gemini 4 Argon (High), index 53, input $2.00/1M, output $10.00/1M, cache discount 95%, speed 84.2 tok/s, TTFT 47.97s, context 1M
Our service: {{service description}}, {{requests}} requests per day, avg input {{tokens}} tokens, avg output {{tokens}} tokens.
Build a comparison table of monthly cost and latency, and judge whether B fits the real-time path or the batch path, with reasons. Mark unknown numbers as 'unknown' instead of guessing.
Next: run 50 batch-path cases through the real API and compareTerminology
Related Papers
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length
Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware