RTK (Rust Token Killer) filters and compresses terminal output earlier the AI supplier sounds it. With complete 79k GitHub stars today, RTK is 1 of the astir celebrated devices to make AI coding cheaper.
One X station saying RTK could trim Claude Code tokens by up to 60% reached 313K views.
Yet JetBrains’s SkillsBench tally recovered nary savings. The README has a disclaimer:
RTK cuts up to 90% of the bash output your supplier reads. […] it is not the aforesaid arsenic cutting your measure by 90%.
So “less terminal output” is not the aforesaid arsenic “cheaper AI coding”. It tin help, beryllium a no-op, aliases backfire (more turns aliases little quality). In this post, we coming our findings aft respective days and complete $1,500 spent connected tokens.
How RTK works
RTK tin rewrite Git, test, package and record commands the supplier runs done its ammunition instrumentality (Bash successful Claude Code, bash successful OpenCode). Each rewrite returns a terser type of the aforesaid output.
For example, RTK keeps record names, sizes and permissions (644 intends rw-r—r—), but drops the proprietor and date:
$ ls -la /app/warriors -rw-r--r-- 1 guidelines root 824 Sep 13 2025 g2-clear.red -rw-r--r-- 1 guidelines root 487 Sep 13 2025 paper.red $ rtk ls -la warriors/ 644 g2-clear.red 824B 644 paper.red 487BTesting RTK connected Terminal-Bench 2.1
RTK compresses terminal output, truthful we tested it connected Terminal-Bench 2.1, a benchmark pinch dense terminal interaction. We stayed connected 2.1 alternatively than the newer 3.0 and 4.0: agents walk astir 2.1 tasks, while 3.0 and 4.0 are still a challenge. Cost only matters for tasks that pass.
We ran Claude Code pinch Fable 5.0, and OpenCode pinch DeepSeek V4 Pro 0813 done OpenRouter. Each task was scheduled 5 times without RTK and 5 times pinch it, connected the aforesaid exemplary route, level and task-specific timeout.
After removing 4 Fable information tasks that sewage refusals, the last comparison covers 85 Fable tasks and 89 DeepSeek tasks, aliases 1,740 attempts.
The first floor plan was promising
With RTK, costs fell by 5% for Fable and roseate by 5% for DeepSeek.
Claude Code · Fable 5.0
baseline$731(84% walk rate)
RTK$698(83% walk rate)
OpenCode · DeepSeek V4 Pro 0813
baseline$51(71% walk rate)
RTK$54(69% walk rate)
Passed attempts Other attempts
Pass rates were little pinch RTK: by 1% for Fable and 2% for DeepSeek. Both pass-rate gaps are small.
When we divided each spending, including grounded attempts, by the number of passes, Fable was 3% cheaper pinch RTK, and DeepSeek was 7% much expensive.
Another measurement is to weight each task equally, because 1 costly task tin outweigh galore inexpensive ones. We compared the mean of each task’s baseline attempts pinch the mean of its RTK attempts, past averaged those changes.
RTK costs alteration vs baseline
Claude CodeFable 5.0−5%+1%OpenCodeDeepSeek V4 Pro 0813+5%+17%-5%0%+5%+10%+15%+20%+25%+30%Claude CodeFable 5.0−5%+1%OpenCodeDeepSeek V4 Pro 0813+5%+17%-5%0%+10%+20%+30%Total measure changeAverage alteration per task95% assurance interval
On this task-level measure, Fable was 1% much expensive, pinch nary clear quality from zero. DeepSeek’s task costs roseate 17% connected average.
Accounting for failures does not alteration the trend. Across the 36 DeepSeek tasks wherever each 10 attempts passed, the summation was still 18%.
One task made the quality successful the full benchmark
Almost each of Fable’s savings pinch RTK came from 1 task: winning-avg-corewars. Both setups passed each attempt, but pinch RTK it vanished successful astir half arsenic galore turns. Across the different tasks, the savings were little than 1%.
DeepSeek had the reverse consequence connected that aforesaid task. Both setups passed each attempt, but RTK took much turns and costs more. Even without that task, costs remained higher pinch RTK.
rtk summation is useless arsenic a costs metric
RTK documents rtk gain arsenic earthy minus filtered bid output successful bytes, divided by 4, not a count of billed tokens.
Across 445 DeepSeek RTK attempts, RTK reported 349.2 cardinal tokens saved, a 89% reduction.
Large reported token savings did not mean cheaper tasks.
rtk gainAverage task-cost change0%large-scale-text-editing57.3M−19%crack-7z-hash38.9M+28%other 87 tasks253.0M+18%rtk gainlarge-scale-text-editing57.3Mcrack-7z-hash38.9Mother 87 tasks253.0MAverage task-cost change0%large-scale-text-editing−19%crack-7z-hash+28%other 87 tasks+18%
In train-fasttext, the exemplary requested caput -1 train.txt twice. RTK credited 120.5 cardinal tokens saved each clip by comparing those constricted sounds pinch the full file. Those 2 calls accounted for 69% of the comparison’s savings counter, though the requested commands would ne'er person returned the full file.
Treating rtk summation arsenic money saved assumes the remainder of the effort would enactment the same. RTK tin alteration the agent’s adjacent turns. rtk summation does not relationship for the costs of those turns.
This is wherever societal posts spell wrong: rtk summation counts removed output, not money saved, and it tin make a much costly effort look optimized.
RTK bugs tin wound you
One DeepSeek git-multibranch effort sewage stuck successful a loop. The supplier ran a find pinch a emblem that rtk find 0.45.0 did not support. The plugin rewrote it to rtk find, which grounded pinch “Use find directly”. Every retry was rewritten again. RTK fixed this successful 0.46.0, aft our runs.
339 consecutive errors~12 min
The supplier accumulated 339 consecutive errors earlier its timeout. It still passed the task, but costs about 9× arsenic much arsenic the matching baseline attempt, which besides passed. One outlier attempt; the inclination holds without it.
Without RTK, instrumentality output made up astir 11% of Fable’s input tokens and 40% of DeepSeek’s.
In the RTK attempts, 31% of Claude Code’s terminal calls and 51% of OpenCode’s terminal calls utilized RTK.
RTK rewrites only ammunition commands: its Claude Code hook matches the Bash instrumentality and its OpenCode plugin acts connected bash calls. Both platforms expose record reference and searching arsenic abstracted Read, Grep, and Glob tools, which bypass RTK. About half of Claude Code’s Bash calls already constricted their ain output pinch head, tail, aliases wc.
In agentic coding, the discourse is cached aft each turn, truthful later sounds of terminal output mostly show up arsenic cache reads. Those costs 1/10 of regular input tokens for Fable, and 1/30 for DeepSeek.
In DeepSeek, RTK reduced terminal-output characters by 9%, yet punctual tokens roseate 9%. Uncached input fell 1% and cached input roseate 9%. Model output, including reasoning, accounted for 56% of costs pinch RTK and 57% without it.
When the supplier took much turns, task costs usually roseate pinch it.
Claude Code · Fable 5.0
Turns, log scale
2255101020205050100100RTKbaselinex = yCost ($), log scale
0.10.10.20.20.50.511225510102020RTKbaselinex = yOpenCode · DeepSeek V4 Pro 0813
Turns, log scale
2255101020205050100100200200RTKbaselinex = yCost ($), log scale
0.010.010.020.020.050.050.10.10.20.20.50.511RTKbaselinex = yDeepSeek’s RTK attempts took much turns connected 58 tasks, and 44 of them costs more. They took less turns connected 28, and 23 of them costs less.
The mean DeepSeek move had 7% little input pinch RTK, but location were 18% much turns overall. Smaller turns did not adhd up to little full input.
One other supplier move tin costs much than the compression saved. It is the aforesaid tokenflation problem successful different form. JetBrains saw the aforesaid shape connected SkillsBench: RTK added turns astatine debased effort and did not little costs astatine precocious effort.
RTK does not make AI coding cheaper
On Terminal-Bench 2.1, Fable’s savings depended connected 1 task and did not clasp crossed tasks. We do not urge RTK arsenic a generic cost-saving tool.
Individual transcripts show that existent frontier models already usage the terminal efficiently (just ~7% of Fable’s discourse was terminal output). Models usage techniques for illustration caput -n aliases tail -n themselves. RTK astir apt helped much pinch older models. Today it is simply a niche optimization, not a root of wide savings.
Tested pinch RTK 0.45.0, Claude Code 2.1.220, OpenCode 1.18.25 and Harbor 0.20. Trajectories disposable on request for follow-up research. Subscribe for early posts, including our planned benchmark of Headroom. Thanks to Piotr Migdał for his reappraisal and feedback.
English (US) · 
Indonesian (ID) · 
