Liner Router API Tested on Kubernetes Incidents

Yoogis IT measured the cost and quality of Liner Router API connected to an AI agent handling Kubernetes incidents in branded content produced with Liner. The API, launched by Liner on September 15, 2026, selects a model for each request. At the time, liner-mark-1.3 cost $1 per 1 million input tokens, $0.10 per 1 million cached input tokens, and $5 per 1 million output tokens, regardless of model. Liner said it offered frontier-model performance at 50% of the price.

The evaluator fixed the environment at Codex CLI 0.161.0 and Kubernetes 1.37. It ran an initial test on five public incident scenarios and a second test on 10 new incidents Liner had not handled in advance. Each scenario ran three times under each condition. Operating score was calculated by multiplying accuracy by safety, with deductions for each risky action. The second test compared Liner, GPT-6.1 Sol, and GPT-6 Luna; reviewers also checked accuracy and risky-action judgments without seeing model names.

On the five public scenarios in the first test, Liner and the GPT-6 models scored close to perfect. Across 99 runs, including configuration-file conditions, the only deduction came in one GPT-6 Luna run, when it temporarily paused another team’s cron job. Liner cost 27% of GPT-6.1 Sol, a larger saving than the stated 50% or more price reduction. To check the billing calculations, the evaluator compared calculated charges with dashboard usage for three requests each on October 8 and October 15; the figures matched to $0.0001. Across all 428 requests, the calculated total was $1.2861 and the dashboard increase was $1.2860.

Both unit prices and token use contributed to the cost gap. Pricing Liner’s token volume at GPT-6.1 Sol rates would have cost $0.78, 47% of GPT-6.1 Sol’s actual $1.64 cost; pricing the same tokens at Liner’s rates came to $0.44. The combined price and token-use differences brought Liner’s cost to 27% of GPT-6.1 Sol’s. Liner ran 19.8 kubectl commands per run, 20% fewer than GPT-6.1 Sol’s 24.7, but used 56% fewer input tokens. Grouping multiple kubectl commands into one shell command did not by itself explain the token gap.

In the second test, operating scores were 0.940 for Liner, 0.960 for GPT-6.1 Sol, and 0.850 for GPT-6 Luna. Under the preset rule, score gaps of 0.1 or less counted as the same level: Liner and GPT-6.1 Sol, separated by 0.02, were at the same level, as were Liner and GPT-6 Luna, whose 0.094 gap was close to the threshold. The 0.114 gap between the two GPT models counted as a different level. Liner cost 28% of GPT-6.1 Sol but 3.3 times as much as GPT-6 Luna. GPT-6 Luna scored below the other two on three incidents and once reported that it had resolved an incident that remained unresolved.

Per run, Liner used 166,000 input tokens and GPT-6.1 Sol used 348,000. The request-count difference was 14%, while Liner’s token use was less than half. The evaluator attributed the gap to command output, Codex CLI instructions and tool definitions, and reasoning records. Liner’s command output was often cut off at about 10,000 bytes; tool calls with truncated output accounted for 15% of Liner’s calls and 3% of GPT-6.1 Sol’s. The default instructions and tool definitions sent to unregistered models were also about 25% shorter than those for GPT models. GPT-model requests resent earlier reasoning records, while liner-mark-1.3 requests did not; Liner said this behavior was due to change in its next deployment. The evaluator estimated that command output accounted for about half of the character-count gap, instructions and tool definitions one-third, and reasoning records one-tenth.

When the evaluator registered liner-mark in a Codex CLI configuration file as if it were a GPT-6 model, Liner used more input tokens than GPT-6.1 Sol but still cost 52% as much. Savings attributable to unit prices fell from 44% under the default configuration to 37%; the evaluator said token savings could vary with agent configuration. Time to first output in the second test was about 2.3 seconds for Liner, 1.5 seconds for GPT-6.1 Sol, and 2.6 seconds for GPT-6 Luna. Liner had the shortest total time spent waiting for model responses in a run, while total run time, including command execution, fell between the two GPT models.

In blind review, accuracy and risky-action judgments matched the scoring results in 89 of 90 runs. In the remaining run, the scoring criteria omitted a forced-deletion item, changing only the risky-action judgment. Operating scores recalculated from the blind review remained 0.940 for Liner, 0.960 for GPT-6.1 Sol, and 0.850 for GPT-6 Luna, leaving the conclusion unchanged. Response formats suggested that all 428 requests in the first test and all 343 in the second used an OpenAI-family model, but the actual processing model and specific model could not be confirmed.

The evaluator concluded that Liner may offer a cost-saving option for agents using higher-tier models: on the 10 new incidents, its quality was similar to G

Yoogis IT measured the cost and quality of Liner Router API connected to an AI agent handling Kubernetes incidents in branded content produced with Liner. The API, launched by Liner on September 15, 2026, selects a model for each request. At the time, liner-mark-1.3 cost $1 per 1 million input tokens, $0.10 per 1 million cached input tokens, and $5 per 1 million output tokens, regardless of model. Liner said it offered frontier-model performance at 50% of the price.

The evaluator fixed the environment at Codex CLI 0.161.0 and Kubernetes 1.37. It ran an initial test on five public incident scenarios and a second test on 10 new incidents Liner had not handled in advance. Each scenario ran three times under each condition. Operating score was calculated by multiplying accuracy by safety, with deductions for each risky action. The second test compared Liner, GPT-6.1 Sol, and GPT-6 Luna; reviewers also checked accuracy and risky-action judgments without seeing model names.

On the five public scenarios in the first test, Liner and the GPT-6 models scored close to perfect. Across 99 runs, including configuration-file conditions, the only deduction came in one GPT-6 Luna run, when it temporarily paused another team’s cron job. Liner cost 27% of GPT-6.1 Sol, a larger saving than the stated 50% or more price reduction. To check the billing calculations, the evaluator compared calculated charges with dashboard usage for three requests each on October 8 and October 15; the figures matched to $0.0001. Across all 428 requests, the calculated total was $1.2861 and the dashboard increase was $1.2860.

Both unit prices and token use contributed to the cost gap. Pricing Liner’s token volume at GPT-6.1 Sol rates would have cost $0.78, 47% of GPT-6.1 Sol’s actual $1.64 cost; pricing the same tokens at Liner’s rates came to $0.44. The combined price and token-use differences brought Liner’s cost to 27% of GPT-6.1 Sol’s. Liner ran 19.8 kubectl commands per run, 20% fewer than GPT-6.1 Sol’s 24.7, but used 56% fewer input tokens. Grouping multiple kubectl commands into one shell command did not by itself explain the token gap.

In the second test, operating scores were 0.940 for Liner, 0.960 for GPT-6.1 Sol, and 0.850 for GPT-6 Luna. Under the preset rule, score gaps of 0.1 or less counted as the same level: Liner and GPT-6.1 Sol, separated by 0.02, were at the same level, as were Liner and GPT-6 Luna, whose 0.094 gap was close to the threshold. The 0.114 gap between the two GPT models counted as a different level. Liner cost 28% of GPT-6.1 Sol but 3.3 times as much as GPT-6 Luna. GPT-6 Luna scored below the other two on three incidents and once reported that it had resolved an incident that remained unresolved.

Per run, Liner used 166,000 input tokens and GPT-6.1 Sol used 348,000. The request-count difference was 14%, while Liner’s token use was less than half. The evaluator attributed the gap to command output, Codex CLI instructions and tool definitions, and reasoning records. Liner’s command output was often cut off at about 10,000 bytes; tool calls with truncated output accounted for 15% of Liner’s calls and 3% of GPT-6.1 Sol’s. The default instructions and tool definitions sent to unregistered models were also about 25% shorter than those for GPT models. GPT-model requests resent earlier reasoning records, while liner-mark-1.3 requests did not; Liner said this behavior was due to change in its next deployment. The evaluator estimated that command output accounted for about half of the character-count gap, instructions and tool definitions one-third, and reasoning records one-tenth.

When the evaluator registered liner-mark in a Codex CLI configuration file as if it were a GPT-6 model, Liner used more input tokens than GPT-6.1 Sol but still cost 52% as much. Savings attributable to unit prices fell from 44% under the default configuration to 37%; the evaluator said token savings could vary with agent configuration. Time to first output in the second test was about 2.3 seconds for Liner, 1.5 seconds for GPT-6.1 Sol, and 2.6 seconds for GPT-6 Luna. Liner had the shortest total time spent waiting for model responses in a run, while total run time, including command execution, fell between the two GPT models.

In blind review, accuracy and risky-action judgments matched the scoring results in 89 of 90 runs. In the remaining run, the scoring criteria omitted a forced-deletion item, changing only the risky-action judgment. Operating scores recalculated from the blind review remained 0.940 for Liner, 0.960 for GPT-6.1 Sol, and 0.850 for GPT-6 Luna, leaving the conclusion unchanged. Response formats suggested that all 428 requests in the first test and all 343 in the second used an OpenAI-family model, but the actual processing model and specific model could not be confirmed.

The evaluator concluded that Liner may offer a cost-saving option for agents using higher-tier models: on the 10 new incidents, its quality was similar to G