How much did removing line numbers save?

Removing line-number prefixes from file reads cut model-inference cost by about 5% in offline benchmarks. In an online experiment with Copilot CLI users, it reduced average daily model-inference cost per user by about 3%.

Did the prompt compression change affect behavior?

The first online experiment caused a regression, making independent agents run sequentially. After adding a regression test and a one-sentence fix, the shipped prompt reduced about 1,300 tokens per turn without quality regression.

AI Coding AssistantsLarge Language ModelsGitHub Blog (AI)4h agoSep 3, 2026, 04:01 JST2 min read

GitHub Copilot cuts AI coding costs without hurting quality

Send on LINE

Key takeaway

GitHub Copilot became more cost-efficient without hurting quality. The company removed unused formatting and compressed repetitive output.
These changes cut costs by a few percent in tests.
Efficiency must be measured across complete tasks, not single tool calls.

3 Key Points

What happened
GitHub shared how it made Copilot more cost-efficient through four changes, including removing unused line numbers and compressing repetitive output. One change cut model-inference cost by about 5% offline and about 3% online.
Why it matters
The company found that focusing on token count per tool call can backfire, increasing overall cost. Optimizing for the whole task, not individual calls, is key to balancing efficiency and quality.
What to watch
Other changes, like the prompt compression, save about 1,300 tokens per turn. GitHub notes that a change saving tokens in one workflow may increase costs in another, so each must be tested in its intended context.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

Context & Analysis

GitHub's post highlights a key principle: optimizing for the outcome rather than the tool call. The company demonstrated that shortening individual responses can lead to extra work and higher costs, as the agent may need to reread output or rerun commands. This was shown with RTK, a utility that shortens shell output, which actually increased overall token usage in tests.

The four changes—selective output compression, removing line-number prefixes, prompt compression via meta-prompting, and delivering background results directly—each required careful evaluation. For instance, the prompt compression initially backfired, causing agents to run sequentially instead of in parallel. Only after adding a regression test and simplifying the guidance did it become safe, saving about 1,300 tokens per turn.

The broader takeaway is that efficiency improvements must be measured in the specific workflow where they run. A change that helps code review may hurt CLI usage, and vice versa. GitHub also found that removing line numbers and compressing output each cut prompt tokens per review by about 5%, with no quality loss, highlighting the importance of context-specific validation.

FAQ

How much did removing line numbers save?: Removing line-number prefixes from file reads cut model-inference cost by about 5% in offline benchmarks. In an online experiment with Copilot CLI users, it reduced average daily model-inference cost per user by about 3%.
Did the prompt compression change affect behavior?: The first online experiment caused a regression, making independent agents run sequentially. After adding a regression test and a one-sentence fix, the shipped prompt reduced about 1,300 tokens per turn without quality regression.

GitHub Blog (AI)Read Original Article

Send on LINE