StarSling's YC Launch Claims 10x Faster CI at $0.008 a Minute, but Its Own Pages Give Three Speedups
Startups / news
StarSling's YC Launch Claims 10x Faster CI at $0.008 a Minute, but Its Own Pages Give Three Speedups
The Spring 2025 company sells GitHub Actions runners that open optimisation pull requests. Every performance figure is the company's own.
StarSling, a Y Combinator Spring 2025 company, launched agent-driven GitHub Actions runners on Hacker News with a claimed average speedup of about 10x across five open-source benchmarks, priced at $0.008 a minute against GitHub's $0.012. The Launch HN thread had 85 points when this article was reported on Nov. 18. Every performance figure below is vendor-supplied, and the benchmark repository the company links does not publish the numbers in its README.
What the product does
StarSling replaces ubuntu-latest in a workflow file with its own runners, per its site. Each job runs in an ephemeral sandbox on AMD EPYC hardware, and an agent reads workflows, logs and machine telemetry, then opens pull requests for caching, dependency installs, build parallelisation and test sharding. A deterministic check is meant to confirm a gain is statistically real before a PR is opened, and the PR carries benchmark evidence.
The founders are Yonas Beshawred, founder and CEO, and Daniel Worku, founder and CTO, according to the Y Combinator company page, which lists a team of 3. In the thread, Yonas says he previously founded StackShare.
Pricing and the free tier
The site lists 2,000 free minutes with no credit card. Agents start opening PRs at around 10,000 minutes, according to the launch post. A "CI Challenge" promises $1,000 if StarSling does not save at least 10% on your next CI bill, subject to terms on a blog post this article did not read.
| Product | Price | Note |
|---|---|---|
| Linux runner, 4 vCPU | $0.008 a minute | GitHub Actions listed at $0.012 |
| Review Runners | $0.004 a minute | Model tokens billed by your provider |
| Disks | $0.50 per GB-month | |
| RTX 4090 GPU runner | $0.02522 a minute | H100 listed as coming soon |
The $0.008 price is 33% under the $0.012 GitHub figure the page cites, which matches the "up to 33% cheaper" claim.
Speedup claims, and where they disagree
The launch post reports five open-source benchmarks from 3.9x on chatwoot to 36.4x on mirotalksfu, averaging about 10x in wall-clock time and 70% less runner time. It also says Mastra's slowest workflow fell from 29 minutes 56 seconds to 5 minutes 6 seconds, which is 5.87x.
- chatwoot (lowest of five)3.9 x faster
- StarSling Disks vs actions/cache4.9 x faster
- better-auth test job5.4 x faster
- Mastra slowest workflow5.87 x faster
- Five-benchmark average10 x faster
- mirotalksfu (highest of five)36.4 x faster
Source: Launch HN post by StarSling, accessed 2026-11-18; all figures vendor-supplied
The company's own pages are not consistent. The site says up to 10x faster, the YC page says both up to 10x and up to 6x in different places, and the site reports over 3 million CI jobs run while the YC launch text says over 1 million during private beta. The two job counts may describe different periods; neither states one.
The linked benchmark repository is MIT licensed and describes fresh, pinned sandboxes with disclosed failures. Its README names no providers, gives no speedup figures and carries no run dates, and says only that results sit in a separate leaderboard file. The page does not say who ran the tests, beyond that the repo belongs to the StarSling account.
What commenters pushed on
One commenter asked what stops Anthropic from reproducing the product from agent logs, and the thread does not answer it directly. Another said buyers who care are likely already doing these optimisations by hand with prompts, and the founders agreed that audience of easy gains is the first target.
The sharpest question was whether test trimming could skip tests that catch bugs. The founders said changes must match test IDs and coverage, with a 5% runner-time cost cap by default, and that judging test trimming is not automated yet.
The pull-request flood this serves is covered in Harvard's coding-agent paper on review bottlenecks, and Claude Code's team published 39 verification checks for agent output. StarSling's claim sits on the same problem from the build side: Mastra, one of its listed customers, merges over 1,800 PRs a month, according to the launch post.
The 2,000 free minutes are the cheapest way to test the 10x figure on your own repository. The company has not published a date for its benchmark leaderboard.
Sources
More in Startups
- 01Unsiloed AI Raises $7.5 Million Seed and Claims 88.0 on olmOCR-Bench Against Reducto's 66.0The Y Combinator Fall 2025 company reported the score itself, and its 99.78% accuracy figure on 542 financial documents has not been independently verified.
- 02Onlook, a YC W25 Startup, Launches a Free Desktop Design Editor With 27,400 Stars and No Pricing PageThe Apache 2.0 app edits a team's own React code on a canvas through the user's Claude Code or Codex account, and neither source says whether the company has raised money.
- 03Genesis AI Raises $500 Million at $3.5 Billion With 30 to 40 Humanoid Robots BuiltThe two reports of the November 17 round disagree on the founding year and on Théophile Gervet's title, and neither discloses revenue or a pre- or post-money basis.
- 04Razorpay's Reported $5 Billion IPO Value Is 12 Times FY25 RevenueThe payments company is expected to publish its updated draft prospectus next week, which will be the first time anyone outside Sebi sees recent numbers.