Gemini 4 Neon vs Kimi K3.1 Pro: Cost per Task

Model: Gemini 4 Neon / Kimi K3.1 Pro

In Artificial Analysis' current coding-agent comparison, the Gemini 4 Neon setup wins on overall score and speed. The Kimi K3.1 Pro setup costs 57% less per task and scores slightly higher on Terminal-Bench 4.1.

That is a comparison of two complete agent setups, not two isolated models. Neon ran in Antigravity CLI at Xhigh effort; Kimi ran in Kimi Code CLI. The harness, prompts, tools and effort settings can all move the result.

Disclosure: Wallaby sells API access to Kimi K3.1 Pro, not Neon. We did not run Neon for this article. The benchmark figures come from Artificial Analysis, accessed 20 November 2026.

The short answer

Metric Antigravity CLI + Neon Xhigh Kimi Code CLI + K3.1 Pro
Coding Agent Index 59 49
FrontierCode-AA 2.0 43 22
Terminal-Bench 4.1 65 67
SWE-Atlas-QnA-AA 68 57
Cost per task $6.86 $2.94
Time per task 30.5 min 41.8 min

The tested Kimi setup cost $3.92 less per task, while the Neon setup finished 11.3 minutes faster and led the composite index by 10 points. Those differences describe these runs. They do not predict the gap on your repository.

Google introduced Neon on 18 November with a 1M-token context. Its announcement cites a Coding Agent Index of 58.6 and the top position at publication time; the comparison page above displays the rounded score of 59.

API price is not task price

Here is the separate rate-card comparison. Prices are per million tokens and exclude taxes, retries and subscription fees.

API rate Gemini 4 Neon Kimi K3.1 Pro official Wallaby K3.1 Pro
Fresh input $3.00 $2.00 $1.80
Cached input $0.15 $0.20 $0.18
Output $15.00 $10.00 $9.00

Kimi has the lower fresh-input and output list prices. Neon has the lowest cached-input price in this table, including against Wallaby. That matters for agents that repeatedly send a long, stable prefix.

The rate card still cannot tell you the cost of a completed task. In the Artificial Analysis runs, the Neon setup used 13.8M tokens and 177.1 turns per task on average; the Kimi setup used 8.89M tokens and 185.2 turns. Cache-hit rates were 95% and 96%, respectively. More turns do not automatically mean more tokens or a higher bill.

Our K3.1 Pro rate card breaks down fresh input, cached input, output and reasoning tokens if you want to reproduce the arithmetic.

Why the harness belongs in the comparison

We ran into the same measurement problem with three coding clients on K3.1 Pro. Codex CLI, OpenCode and Crush each fixed the same small bug, and all five unit tests passed after every run. Only the client changed.

Client Requests Time Bill
Codex CLI 0.154.0 3 47 seconds $0.034
OpenCode 1.19.5 6 76 seconds $0.070
Crush 0.98.1 7 68 seconds $0.097

The highest bill was 2.9 times the lowest for the same model, task and passing-test outcome. The full test and gateway receipts include the token counts and limitations.

This was one run per client on a toy repository, so it is not a client ranking. It also cannot explain Artificial Analysis' Neon-Kimi gap. It does show why a model-only comparison can miss a large part of the invoice.

Where Neon is the better fit

Choose the Neon setup to evaluate first if the higher composite score is your priority. Its clearest advantage in this comparison is FrontierCode-AA 2.0: 43 versus 22. It also leads SWE-Atlas-QnA-AA by 11 points and has the faster observed completion time.

Neon's $0.15 cached-input rate is another concrete advantage for long-context workflows with reusable prefixes. Do not let the lower total task cost in one Kimi comparison obscure that rate-card difference.

Antigravity CLI is part of the result. If that is already your workflow, this table is more relevant to you than a model benchmark run in an unfamiliar harness.

Where Kimi K3.1 Pro is the better fit

Put K3.1 Pro on the shortlist if cost per accepted task carries more weight than the composite score. The tested Kimi setup cost $2.94 per task and edged Terminal-Bench 4.1 by 2 points.

Kimi's official fresh-input and output rates are also one-third below Neon's list rates. Your token mix can still produce a higher bill, but these rates are a useful starting point for output-heavy tasks.

If you are moving from K3 rather than choosing a new provider, our K3 vs K3.1 Pro migration guide covers the model ID, context window and price changes.

Run a comparison you can buy on

Use the same repository tasks, tests and definition of done for both setups. Where both models are available in one client, hold the client fixed too. Otherwise, name the full model-and-harness combination in every result.

The primary metric is:

Cost per accepted task = total API spend / tasks passing your tests without human repair

Include failed runs, retries and follow-up prompts in total API spend. A cheap first attempt that needs human repair is not a cheap completed task.

A practical pilot is 10 representative tasks with two runs per setup. Keep a table with pass count, total spend, elapsed time and human interventions. Separate regression tests you wrote before the run from fixes made during it. Ten tasks will still leave uncertainty, but the result will describe your work instead of a public leaderboard.

The buying decision

Pick Neon for the higher observed composite score and speed. Pick Kimi for the lower observed task cost and Terminal-Bench result. Do not pick either from this table without accounting for the harness.

For your own workload, the deciding number is not dollars per million tokens. It is the total bill divided by tasks you would actually accept.

If Kimi is on your shortlist, Wallaby offers $0.50 in trial credit with no card required. Create a free account, set a small spending cap, and run the task set above on a branch you can discard.

Sources: Google's Gemini 4 Neon announcement, 18 November 2026; Artificial Analysis' Antigravity CLI + Neon Xhigh vs Kimi Code CLI + K3.1 Pro comparison, accessed 20 November 2026; Moonshot and Wallaby K3.1 Pro rate cards; Wallaby gateway logs for the three-client pilot, 16 November 2026. Prices and benchmark positions can change.

Like the Dispatch? Add Wallaby as a preferred source on Google, or subscribe via RSS.
Did this help?