Ask
23
@kesslerj ·

OpenRouter after Stripe: 5.5% fee, provider drift, RTK, litelm

Stripe is buying OpenRouter: announced on 19 August, reported at over $7 billion, closing expected within weeks. The announcement promises the same product, name and roadmap, with routing decided by what is best for the user. The visible price today is a 5.5% platform fee on pay-as-you-go credits, 8% on the Business plan, and bring-your-own-key is free up to $25,000 of list-price inference a month, 5% after that.

Two write-ups from the same fortnight measured the costs that fee does not show.

One operator, running about 18 million messages with roughly a third on open models through OpenRouter, tested the providers behind a single model id. On OpenRouter's own per-provider board, DeepSeek V4 Flash 0731 scored 90% on GPQA and 81% on TAU-Bench Airline at DeepSeek itself, and 75% and 58% at another host serving the same weights. Some hosts accepted reasoning.effort and ignored it, some returned HTTP 200 with empty content, and filtering for declared fp8 did not predict quality.

The other measured RTK, the terminal output compressor with 79k stars, across 1,740 Terminal-Bench 2.1 attempts and more than $1,500 of tokens. The bill fell 5% with Claude Code on Fable 5.0 and rose 5% with OpenCode on DeepSeek V4 Pro. Almost all of the Fable saving came from one task, and rtk gain reported 349 million tokens saved on runs that cost more.

And for anyone who only wanted the call path, litelm cuts LiteLLM down to routing and message translation: about 2,900 lines and two dependencies.

13 answers Share
Report

Answering anonymously, a moderator will review it first.

  • @anouk_v · 2d ago · 3 replies

    The per-provider board is the most useful page on the site and almost nobody opens it. Pick the benchmark closest to your workload, prefer the two or three providers within a couple of points of first party, and leave fallbacks on.

    26
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @grayfox_31 · 2d ago · 2 replies

      Leave fallbacks on is doing a lot of work there. The same write-up pinned three providers with fallbacks off, and within two weeks one was rate limiting everything, one had dropped the model and the third started returning 429s too.

      15
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
      • @anouk_v · 2d ago

        Which is exactly why they stay on. A pinned order is a preference, not a promise that anyone will still be serving it next month.

        10
        Share
        Reply

        Answering anonymously, a moderator will review it first.

        Report
  • @marlowe_t · 2d ago · 3 replies

    Nothing changes is what every acquired company says in week one. I will be watching whether routing drifts toward whoever is the biggest Stripe customer, and whether that per-provider board stays this blunt.

    20
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @olumide · 3d ago · 2 replies

      Stripe earns on the transaction happening, not on which lab wins it. Of the plausible buyers, a payments company has the least reason to tilt routing. A model lab or a cloud buying it would have been the real worry.

      14
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
      • @marlowe_t · 2d ago

        Fair point on incentives. I am still exporting my activity logs every month.

        9
        Share
        Reply

        Answering anonymously, a moderator will review it first.

        Report
  • @plumline · 3d ago · 3 replies

    The RTK result makes sense once you look at where the tokens were. Terminal output was about 7% of Fable's input tokens, and later turns read it back from cache at a tenth of the price. Shrinking a small, mostly cached slice barely moves the bill, and changing what the agent sees can add turns.

    17
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @yuki_t · 3d ago · 2 replies

      So is it useless?

      11
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
      • @plumline · 2d ago

        As a cost tool in that run, yes: DeepSeek tasks came out 17% more expensive on average and Fable was a wash. If context length is your constraint rather than money, maybe not.

        12
        Share
        Reply

        Answering anonymously, a moderator will review it first.

        Report
  • @yuki_t · 3d ago · 2 replies

    Moved a small service from LiteLLM to litelm last week. The import change was one sed and the dependency tree went from pages to two lines. You lose the router and cost tracking, which I was already doing in my own table.

    14
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @kesslerj · 2d ago
      @kesslerj OP ·

      Your own table is the underrated part. Every gateway dashboard reports the price it believes you paid, and nobody reconciles it against the invoice.

      1
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
  • @grayfox_31 · yesterday

    HTTP 200 with content null is my favourite genre of bug.

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
  • @olumide · 13h ago

    Stripe for LLMs, now literally.

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report