Antigravity CLI research review
Four sessions · agy 1.5.11

Where Gemini
stopped looking

Three Gemini models delivered answers in 15–20 seconds. Each searched the web once. None checked both a downloadable checkpoint’s size and the hardware’s available memory. Claude Opus checked both on the same prompt.

Tool calls per session
Carbon6
Argon3
Hydrogen4
Opus24
Research task · paraphrased

Could an existing DGX Spark cluster run Mistral Large 4, also known as Le Chonk?

What each model answered

Each card quotes the model’s final reply, followed by a summary of the work recorded in its session.

Gemini 4 Carbon High

6 tool calls · 20s · 2,681 output tokens
you would likely need to look at cloud-based instances with high-bandwidth interconnects

Carbon found an inventory entry and raised a networking objection. A reference in its own results pointed to a more authoritative hardware guide. It finished without opening that guide or checking the size of a downloadable checkpoint.

Gemini 4 Argon High

3 tool calls · 15s · 1,773 output tokens
running large language model inference directly inside Spark tasks is often inefficient for high-throughput, low-latency needs.

Argon selected a cloud capacity document and interpreted “Spark” as the Apache data-processing framework. Its answer explained how to serve a model to data-processing jobs. The prompt concerned NVIDIA DGX Spark hardware.

Gemini 4 Hydrogen High

4 tool calls · 15s · 1,934 output tokens
Verify the exact GPU model and count per unit (e.g., how many H100s or B200s per GB10 unit)

Hydrogen identified the GB10 hardware. A broad search for specifications returned mostly unrelated material. It then handed the hardware lookup back to the user, even though the identified product had a published specification.

Claude Opus 5.5 High

24 tool calls · 3m 33s · 15,740 output tokens
Yes, you can run it, but only as a compressed version. The official checkpoints don't fit in the cluster's memory.

Opus followed the hardware documentation, retrieved published weight sizes and retried a failed read-only capacity check. Its answer named compressed candidates and labeled its speed estimate as unmeasured. It did not test inference.

Tool calls, timing and token usage

The three Gemini sessions made 13 tool calls combined. Opus made 24 in one session, including a second attempt at a hardware check and a separate report write.

MetricCarbonArgonHydrogenOpus
Tool activity
Tool calls63424
Shell commands21316
Web searches1114
File reads3103
File writes0001
Responses and timing
Elapsed time20s15s15s3m 33s
Model responses74518
Responses calling a tool63417
Tool-using responses with a reasoning field2 of 61 of 32 of 414 of 17
Reported token counts
Output tokens2,6811,7731,93415,740
Input tokens45,19853,35740,933136,457
Cache-read tokens165,88855,29698,3041,318,132

Elapsed time runs from the successful prompt to the end of the reply. Hydrogen took 25 seconds including a rejected XHigh request and resubmission at High. A shell call can perform several checks. File reads and writes count the dedicated file tools.

What the comparison supports

The missing checks affected the conclusions.

Carbon read a summary and left its network diagnosis unverified. Argon answered for another platform. Hydrogen identified the hardware and then asked the user for a specification it could have looked up. Opus’s extra work produced a concrete candidate and a clear remaining test.

One task, one completed answer per model.

All four successful sessions used the same prompt, CLI version, working directory and High effort. The counts were derived from individual transcript events. Output totals include tool-generating responses and report writing. A recorded reasoning field indicates that the CLI received metadata. Providers expose it differently.

These sessions establish how four answers were researched. They cannot establish a model-wide regression, its cause, or whether the proposed deployment would work.