StarSling's YC P25 CI Pitch Rests on Vendor Speedups Up to 36.4x, and Its Post and Site Disagree
Startups / news
StarSling's YC P25 CI Pitch Rests on Vendor Speedups Up to 36.4x, and Its Post and Site Disagree
The runner price is 33% under GitHub's 4-core rate. The agent-written pull requests are the part with no independent check.

StarSling, a Y Combinator P25 company, sells GitHub Actions runners at $0.008 a minute for a 4 vCPU Linux machine, which its pricing page calls 33% cheaper than GitHub. That is arithmetic GitHub's own price list supports ($0.012 for its 4-core runner), and it is the least interesting number in the pitch. The expensive claim is that AI agents will open pull requests that make your pipeline faster, backed by benchmarks the company ran on its own customers and on open-source repositories.
The company announced itself in a Launch HN post that reached 89 points, and its site carries the price list and customer logos. No funding figure appears in either, so this is a product story, not a round story.
What StarSling charges
The runners are a one-line change to runs-on. New accounts get 2,000 free minutes with unlimited concurrency and no card. After that the 4 vCPU Linux rate is $0.008 a minute, billed by the second, with queue time never billed and a 99.99% SLA. Review Runners, an AI code review job that runs inside the customer's CI, cost $0.004 a minute. GPU runners start at $0.02522 a minute for an RTX 4090.
GitHub's runner pricing page lists Linux 2-core at $0.006 and Linux 4-core at $0.012 a minute. StarSling's discount is therefore 33.3% against the 4-core price and a 33% premium against the 2-core price, so the comparison only holds if the customer was already paying for larger runners.
The speedups are vendor-supplied
Every performance figure below comes from StarSling's launch post or site. None was run by an independent party.
- mirotalksfu36.4 x
- gogs12.8 x
- transfer.sh11.7 x
- Mastra5.87 x
- shellcheck5.5 x
- chatwoot3.9 x
Source: StarSling Launch HN post, accessed 2026-11-21 (vendor-supplied)
The Mastra figure is the customer one: its slowest workflow fell from 29 minutes 56 seconds to 5 minutes 6 seconds, and the company says Mastra merges more than 1,800 pull requests a month. The post puts the open-source average at about 10x wall-clock and 70% less runner time. The five open-source repositories the post names average about 14x by simple arithmetic, above the roughly 10x it quotes, so the quoted average must cover repositories not listed. Without the 36.4x outlier, the other four average about 8.5x.
Where the post and the site disagree
The two documents do not match on three points a buyer would check.
| Point | Launch HN post | Site |
|---|---|---|
| Disk cache speedup | 4.9x faster than actions/cache | 4.5x faster than actions/cache |
| When agent PRs start | About 10,000 minutes | Paid plans only, off by default |
| Better Auth result | One test job 5.4x faster | E2E tests from 2m 22s to 1m 04s, about 2x |
The post also offers a "CI Challenge": $1,000 if StarSling cannot save at least 10% on the next CI bill.
What commenters pressed on
Commenters asked whether an agent chasing speed could weaken tests to win. A founder identified in the thread as Yonas, who previously founded StackShare, replied that wins require matching test IDs and coverage and that trimming tests is not something the agents do yet. Another asked whether the agent diffs test bodies, since a looser timeout or a mocked dependency would look like a speedup. A third asked whether inference cost would eat the savings on small repositories; the reply was that savings usually outweigh the cost on large repositories, with no size threshold given.
The other founder in the post, Daniel, helped build Netflix's internal developer portal. Neither founder's title appears in the post. For comparison with other agent tools that publish a single-task result, see our report on OpenComputer's 3.3x Devin claim, and for another CI-focused agent, OpenSRE.
What would confirm it
The free tier ends at 2,000 minutes and the agents begin opening pull requests at roughly 10,000, so a team can only judge the optimisation claim after spending about five times the free allowance. The first independent check will be a customer publishing before-and-after bills, or a repository maintainer posting what an agent's pull request actually changed.

A Zen 5 desktop chip. StarSling's post says its runners use Zen 5 machines; it does not name this part.
Sources
More in Startups
- 01Traya's FY26 Marketing Spend Doubled to ₹275 Crore While Revenue Rose 73%, Ahead of Verlinvest TalksThe hair-care company's loss widened to ₹29 crore on ₹585.1 crore of revenue. Two reports of the same filings disagree on EBITDA margin and stock purchases.
- 02StayTouch's Parent Seeks ₹1,500 Crore From Apple Over NameDrop, Against a $2.75 Million Spend CertificateGG Technologies has filed in the Bombay High Court on top of two US suits. The damages figure is a self-described conservative estimate with no shown calculation.
- 03Everful's $48.5 Million Series A Comes With a $100 Million Revenue Forecast and No Disclosed ValuationVertex and January Capital co-led the round for the San Francisco and Singapore sourcing company, but margin, prior funding and valuation are all missing.
- 04Paytm Pays Rs 162 Crore for 19.1% of CardUp, Which Implies Rs 848 Crore, and Its Filing Dates DisagreeThe Singapore payments firm's turnover rose to Rs 299.2 crore in 2025, so the minority stake prices it at under three times that figure.