Arena has claimed that Google DeepMind's Gemini 4 Carbon becomes relatively cheaper than competing frontier models as tasks grow more demanding, citing GPT-6 Astra from OpenAI as costing 3.4 times as much for a typical task.
Cost scaling · Arena claim
Gemini 4 Carbon may get relatively cheaper as tasks get harder
Arena says the cost gap between Gemini 4 Carbon and rival frontier models widens as workloads grow more complex. Its headline figure: OpenAI's GPT-6 Astra costs 3.4 times as much per typical task.
3.4×
GPT-6 Astra cost per typical task, relative to Gemini 4 Carbon
1×
Gemini 4 Carbon baseline for a typical task (the reference point)
↑
Claimed gap widens as task complexity rises (no curve published)
Relative cost per typical task
Each block = 1× Gemini 4 Carbon cost. Column height is proportional to the multiplier.
Top partial block shows the 0.4 fraction of the 3.4× figure. Claude Opus 5.5 is referenced in the claim, but its multiplier was not available and is not plotted.
What the claim leaves out
The 3.4× figure cannot be reproduced without the test conditions behind it.
→ Benchmark tasks→ Token counts→ Pricing tiers→ Hardware conditions
If the claim holds
Long, multi-step jobs could yield larger savings than with alternatives.
Per-task cost is central to whether enterprise AI agent and inference projects are viable.
Caveats
Figures are a company's assertion and not yet independently verified.
Not checked against published pricing from Google, OpenAI or Anthropic.
Methodology must be published before the result can be treated as established.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...