Artificial Analysis's Coding Agent Index v1.5 defines an agent as a combination of a model and a test harness, using three agentic coding evaluations: DeepSWE v1.1, Terminal-Bench 4.0, and SWE-Atlas-QnA. The original post records Gemini 4 Argon (high) scoring 64 on Antigravity CLI, at a cost of $5.84 per task. This is a public index snapshot for a specific model–harness combination, not a comparison of the model alone, and it should not be conflated with the Intelligence Index results in review02.
Suitable judgments: Recording the index score and evaluation-task cost for Argon and other model–harness combinations in this Coding Agent Index v1.5 snapshot.
Unsuitable extrapolations: Inferring Argon's general coding ability, speed, production billing, or success rate across all software-engineering tasks outside Antigravity CLI; scores from different harnesses should not be treated as a ranking of the models alone.
Model version: Gemini 4 Argon (high); the original post does not provide an API model ID or more specific version identifier.
Test environment or client: Antigravity CLI; the Coding Agent Index explicitly includes the harness in the evaluated agent. The original post does not disclose the call endpoint, provider, inputs, sample size, or complete runtime environment.
Reasoning tier and parameters: high; parallel settings, prompts, context, output limit, retries, and other parameters are not disclosed.
The post says that Artificial Analysis's Coding Agent Index measures an agent—that is, a combination of a model and a harness—and covers three agentic coding evaluations. The attached image labels the version Coding Agent Index v1.5 and lists:
DeepSWE v1.1
Terminal-Bench 4.0
SWE-Atlas-QnA
The image is labeled Higher is better. The original post does not disclose per-evaluation inputs or outputs, sample size, run count, stopping conditions, failure handling, randomness controls, call endpoint, provider snapshot, or complete harness, so the individual experiments cannot be reconstructed from it.
These data differ from the Artificial Analysis Intelligence Index v4.3.2 in review02: this post discusses the Coding Agent Index, so its score of 64 must not be substituted for or combined with the Intelligence Index score of 53.
| Agent combination | Reasoning tier | Coding Agent Index score | Cost per task | Page's wording |
|---|---|---|---|---|
| Gemini 4 Argon + Antigravity CLI | high | 64 | $5.84 | Google promotional pricing; Argon was not yet publicly available at the time |
| Claude Sonnet 5.5 + Claude Code | max | 68 | $14.19 | The original post calls it the top-ranked model and the most expensive per task measured |
| GPT-6.1 Sol + Codex | xhigh | 63 | $1.04 | The original post says the cost was about one-sixth of Argon's |
The three rows are different model–harness combinations and can serve only as contextual comparisons within the same index snapshot. The original post says Argon's cost was less than half that of Sonnet 5.5, but provides no cost breakdown by evaluation, input/output tokens, provider, or billing item.
English text of the original post: Artificial Analysis says Claude Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon rank near the top of the Coding Agent Index, with different performance–cost trade-offs. It also says the index measures “the combination of a model and harness” through three agentic coding evaluations.
Argon entry: Argon's high tier scored 64 on Antigravity CLI at $5.84 per task; the original post also notes that this used Google's promotional pricing and that Argon was not yet publicly available at the time.
Attached image: The title is Artificial Analysis Coding Agent Index; the subtitle says v1.5 includes DeepSWE v1.1, Terminal-Bench 4.0, and SWE-Atlas-QnA, and is marked Higher is better. The Argon entry is labeled Antigravity CLI / Gemini 4 Argon with a score of 64; the cost chart labels it Antigravity CLI - Gemini 4 Argon (Google).
Replies in the thread: Artificial Analysis posted a follow-up about exploring the full results and comparing coding agents, with an Artificial Analysis link. The current original-post page shows no additional per-evaluation Argon data. Visible replies question whether the model was publicly available and ask whether this was early access; these are availability discussions, not independent reruns.
Engagement data on the original post: The page shows 96 replies, 105 reposts, 1,005 likes, 121 bookmarks, and 98,514 views. These engagement figures are not evaluation scores.
The score of 64 and cost of $5.84 per task apply only to the Gemini 4 Argon (high) + Antigravity CLI combination in Artificial Analysis Coding Agent Index v1.5. They cannot be separated into an independent score or general price for the Argon model itself.
The attached image clearly identifies the three evaluations, but the page does not disclose per-evaluation prompts, inputs, samples, complete harness, provider, call endpoint, or run logs. “Three evaluations” does not mean this record contains three independently reproducible test datasets.
high is the Argon tier listed in the original post. The other parameters are not public, so parallelism, context length, output limit, retry strategy, or token statistics must not be filled in.
The $5.84 uses Google's promotional pricing and must not be treated as a fixed production cost. The original post provides no billing period, caching, platform markup, input/output-token, or exchange-rate basis.
Sonnet 5.5, Argon, and GPT-6.1 Sol use Claude Code, Antigravity CLI, and Codex respectively. Different harnesses enter the same index, so the models cannot be rigidly ranked in isolation; 68, 64, and 63 alone cannot establish software-engineering quality or the actual range of tasks that can be completed.
The original post says Argon was not publicly available on October 2. The thread's early-access discussion belongs to that historical snapshot and does not describe the status after broader availability on October 13; the current entry point should be checked against the official model catalog.
This post's Coding Agent Index score of 64 must not be combined with or substituted for review02's Intelligence Index score of 53 or AA-Omniscience metrics; their evaluation compositions and definitions differ.
To reproduce this snapshot, fix Artificial Analysis Coding Agent Index v1.5, the three evaluations, Gemini 4 Argon (high), Antigravity CLI, provider, inputs and output limits, stopping conditions, failure handling, retries, token accounting, and the pricing snapshot. Also record the model tier and harness configuration separately for the Claude Code and Codex combinations. The original post does not disclose these inputs, endpoints, or runtime details, so only the public claims of a score of 64 and $5.84 per task can be checked; this article does not claim to have independently reproduced the individual evaluations.
Gemini 4 Argon
