GitHub Cuts Copilot's Token Costs by Low Single Digits

GitHub said four changes shipped to Copilot CLI cut model-inference cost by 2 to 6 percent apiece, according to a Sept. 2, 2026 engineering post from staff software engineer Erik Krogh Kristensen and software engineer Napalys Klicius. None of the fixes made the underlying model cheaper or smarter. They removed formatting, prompt text and retrieval steps the model never needed. GitHub said every candidate change was first screened offline against its own agentic coding benchmarks, and only changes that held up there moved into controlled online experiments on real Copilot CLI traffic before shipping.

The post's most specific finding cuts against a popular assumption in the field: that shortening each tool's output automatically lowers an agent's bill. GitHub tested RTK, an open-source utility called Rust Token Killer that trims shell output before an agent reads it, against its own agentic coding benchmarks. When RTK omitted text the model needed, GitHub said the model sometimes reopened the original output or reran the command to recover it, adding turns and carrying more context forward. The tool response was shorter. The completed task, on average, used more tokens and took longer.

What each fix actually saved

GitHub ran four independent A/B experiments and reported the AI-credit savings from each, noting the effects are not strictly additive when combined. Two changes moved from an offline benchmark result into a measured online effect on real Copilot CLI users.

Change	Measured savings
Remove line-number prefixes from file reads	5% offline; 3% off daily per-user inference cost online
Selective output compaction (install, build, test noise)	5.5% in offline A/B testing
Compact task-tool prompt (cut roughly in half)	1.8% fewer prompt tokens per session; 2.9% lower cost per active hour
Batch background-task notifications	2.3% lower token-related usage, measured in AI Credits

The prompt-compaction change carried the clearest warning about doing this work carelessly. GitHub used a meta-prompting loop, in which Copilot rewrote its own instructions, to shrink the guidance its task tool sends on every turn by about half. The first online test caught a regression the offline evaluation missed: the shortened prompt had turned permissive guidance about running sub-agents in parallel into a rule that forced them to run one at a time. GitHub said it stopped the experiment, wrote a regression test for the exact behavior that broke, then replaced an explicit allowlist and denylist with one sentence: independent agents can run in parallel, consider side effects. That one-sentence fix shipped roughly 1,300 fewer tokens per turn without reintroducing the serialization bug.

GitHub was explicit that a saving in one product does not transfer automatically. A tighter file-tool instruction set that helped Copilot code review increased cost when tested in Copilot CLI, and GitHub did not ship it there. Removing line-number prefixes and selectively compressing output each cut review-task prompt tokens by about 5 percent independently, and combined with an earlier June 25, 2026 migration to shared file tools plus instruction tuning, GitHub said total code-review cost fell by about 20 percent.

The bigger lever the post skips

Every figure in GitHub's post describes savings on top of whichever model Copilot is already running. It says nothing about what model choice itself costs, and that gap is the more consequential number, the same variable that let Cognition price its SWE-2 agent below rivals on some benchmarks while trailing on others. A separate analysis published by cost-management firm Vantage in April 2026 priced a representative 50-turn agentic coding session, about 1 million input tokens and 40,000 output tokens, across four models and found roughly a tenfold cost spread: $6.00 on Claude Opus 4.6, $3.10 on GPT-5.4, $1.80 on Composer 2 Fast and $0.60 on Composer 2 Standard for comparable work.

1. Claude Opus 4.66 $
2. GPT-5.43.1 $
3. Composer 2 Fast1.8 $
4. Composer 2 Standard0.6 $

Source: Vantage, accessed 2026-09-11

Scaled to a 25-person engineering team running roughly 1,000 agentic sessions a month, Vantage put the annual gap between the cheapest and most expensive model choice at $7,200 versus $72,000, a $64,800 swing from one dropdown setting. Vantage also found input tokens, mostly repeated context re-sent on every turn rather than the code an agent writes, account for about 85 percent of a session's cost, which is exactly the category GitHub's fixes target. The fixes are real. They are also working on the smaller half of the bill.

What's missing from GitHub's numbers

GitHub did not publish a baseline dollar figure or AI Credit count for Copilot CLI before these changes, so none of the percentages in its post convert into an absolute cost readers can check. It also did not say which model or models the benchmarks ran on, information that matters given how much Vantage's data shows that variable alone can move the bill. What would change this read is GitHub publishing that baseline. Until then, the honest description of the Sept. 2 post is a set of real, individually modest savings on the fixed costs of running an agent, not evidence about the much larger lever of which agent to run in the first place.