redshift

Building Routstr.

Public Key

npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh

NIP-05 Address

Profile Code

nprofile1qqsy44h695tw92d4wmyx8dx0wsz2wr2dcvsvp3z86y9ddl6cnyl2ejqpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dshj0vqm

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh

Show more details

Last Notes

2026-09-01 09:50:42 UTC

If you've noticed, routstrd now requires a lot fewer sats reserved for even big models like Kimi-K3.
Kimi-K3 requires just 1.2k sats to make a request and usually 50 sats per request.
#nevent1q…2ts9

2026-08-31 13:03:28 UTC

- reply

I understand your concern. I'm also learning this as we go.
Just to note:
routstrd: routstr daemon that you run locally. it's a client that communicates with Routstr nodes
Routstr nodes (they run routstr-core): providers who offer AI models for sats.
routstrd uses our sdk which in turn uses tinfoil's SDK (https://github.com/Routstr/routstr-sdk/blob/main/client/TinfoilSecure.ts) to encrypt the message body on the client side to Tinfoil's servers where decryption happens. They have TEE-terminated TLS so you can be sure that the decryption happens inside their TEE. (source: https://confidentialinference.net/providers/tinfoil).
So Routstr nodes, @npub1fet…jzfa in this case, cannot read any of your messages, you can verify this by running a node yourself. If there is a way to read your messages it has to be on the Tinfoil side, that is if their system is not well set up. Let's look at Tinfoil:
Here we are attesting that their TEEs run their released binary on Github. Yes, we are essentially trusting that new Github binaries don't contain any backdoors. These are the guarantees we have:
1. The enclave is real AMD SEV-SNP hardware. SEV-SNP report verified against AMD VCEK.
2. The code inside is the auditable open-source release. Sigstore bundle for tinfoilsh/confidential-model-router.
3. The measured code is what's actually running. Enclave measurement == signed release measurement
4. The request body can only be read by that enclave. EHBP encrypts to attested HPKE public key
5. A third party can't pretend to be Tinfoil. Certificate SANs bind HPKE key + attestation hash to *.tinfoil.sh
Auditing everything is very hard for us. Happy to hear any feedback here and improve our setup and/or support better TEE inference providers if any.
Tagging the TEE experts I know here: @npub1alj…g9jp @npub1gzu…a5ds @npub1226…grkj

2026-08-31 11:27:15 UTC

- reply

Please say more.
How can we ensure that we're getting the attestation of the actual code? What can we demand from Tinfoil?
This is what Routstr does right now (from my LLM):
Trust boundaries worth knowing
A few nuances from reading the code:
1. We follow the latest signed release. routstr-sdk does not pin a specific releaseTag or expected codeFingerprint by default. It verifies whatever the verifier says is the latest signed release of tinfoilsh/confidential-model-router. That's a "trust the release process" model, not a hardcoded allow-list.
2. The bundle is assembled by Tinfoil's ATC. The SDK fetches a pre-assembled bundle from https://atc.tinfoil.sh/attestation rather than fetching GitHub/Sigstore/AMD itself. This is safe because every component is independently signed/verified — the ATC is a proxy, not a trust root — but it's still a component in the path.
3. The verified repo is the router, not the model server directly. By default, configRepo is tinfoilsh/confidential-model-router. Downstream model enclaves are covered by Tinfoil's enclave chaining — the router verifies the inference enclave. Our client only directly verifies the first hop.
4. Headers are plaintext to the proxy. EHBP encrypts the JSON body. The auth token, X-Routstr-Model, and other headers are visible to the Routstr proxy. The model id inside the body is stripped to the bare upstream id and encrypted.
5. Dependency supply chain. We trust [email protected], [email protected], and their transitive deps. The package lock pins them, but the npm registry supply chain is a boundary.
So: yes, routstr-sdk does real Tinfoil attestation — client-side, before token spend, with encryption bound to the attested key. The main trust decisions are "follow latest signed router release" and "trust AMD + GitHub Actions/Sigstore + the Tinfoil npm package."

2026-08-26 20:35:34 UTC

- reply

Yes. A release is coming soon. I’d say update to the latest main for now. The latest main should fix the issue with Minibits. Plus, we’re gradually moving away from Minibits. Finding mints that charge fees and have better rate limits

2026-08-26 19:59:52 UTC

- reply

Yeah. We’re in search for faster and more reliable mints. Cuba is good but slow. Would you say Macadamia has been more reliable?

2026-08-23 20:17:28 UTC

- reply

> I am finding it not possible to replicate Claude's output
What about Claude’s output that doesn’t work well with other models? (Proprietary or open source)?
All 1M context models via APIs run the same way as using Claude inside Claude Code

2026-08-21 13:02:23 UTC

- reply

Run “routstrd models —refresh” to force refresh.
And “routstrd clients add —opencode” to install in there.
(It’ll automatically pull and update opencode every 21mins)

2026-08-21 13:00:47 UTC

- reply

please use routstrd
It should solve your issue.
https://routstr.com/routstrd
Re: api.routstr.com, we have sunset that node. Will refund any stuck/lost funds.

2026-08-21 12:55:24 UTC

- reply

Sorry. I forgot to publish it. Try now.

2026-08-20 08:17:08 UTC

- reply

Awesome. Will try it out! Ty

2026-08-19 20:12:47 UTC

- reply

What do you use to run a Debian VM?

2026-08-19 17:22:47 UTC

Took us a while to fix all these issues. Sorry about the inconvenience in the last week or so. Please update and try the new version of Routstrd let me know if you have issues:
routstrd update
Or if you’re using an older version:
bun i -g routstrd
Officially releasing soon!
#nevent1q…77ah

2026-08-16 08:54:40 UTC

- reply

it’s some work on the routstr-core side not on the routstrd side. Just started working on it yesterday. 😁

2026-08-13 15:33:06 UTC

- reply

Bricked in the sense that you can still withdraw your balance from the wallet, but you cannot use it with routstr anymore.
Damn, I should let my clanker handle our communications. So many typos and mistakes.

2026-08-13 15:29:21 UTC

A part of this is also because of the recent Minibits changes. If your default mint has been Minibits, you're essentially bricked until we handle this migration. This and in combination with a commit that went into the last routstr-core release will cause you disruptions despite having a second mint in your Routstrd.
#nevent1q…wxu8
#nevent1q…pxmg

2026-08-13 15:24:04 UTC

- reply

Typos :(
I'm really sorry that the infrastructure**

2026-08-13 15:02:50 UTC

To all routstrd users, I’m really sorry about the infrastructure has been sub par this last week or two. We are not actively being attacked by anyone.
It is simply because this stuff is hard (combining Nostr, Bitcoin+Lightning+Cashu and LLMs) and we made a mistake of pushing a change to Routstr Core that should’ve been in the next release.
We are responsible for the disruptions in your workflow. I’m really sorry about it. We’re working on making it a lot more stable with the upcoming release of routstrd.

2026-08-13 10:33:55 UTC

- reply

Oh okay. Yeah. I might have hit your node when you were restarting.

2026-08-12 13:48:26 UTC

- reply

It was an issue with Minibits btw, it was down, it should be up now.
Also, i get 502 on your node sometimes:
```
[req:234dfs45:RoutstrClient] [RoutstrClient] _handleErrorResponse: status=502, baseUrl=https://privateprovider.xyz/, mode=apikeys, token preview=sk-<KEY>e1c0ab, requestId=undefined, errorType=unknown, errorCode=unknown, errorMessage=<html>
<head><title>502 Bad Gateway</title></head>
<body>
<center><h1>502 Bad Gateway</h1></center>
<hr><center>nginx</center>
</body>
</html>
```
PS: tinfoil-kimi-k3 was updated there right after.

2026-08-11 18:05:38 UTC

- reply

Done

2026-08-11 18:01:53 UTC

- reply

Ooo. I didn’t know tinfoil had kimi-k3. Will do

2026-07-31 06:24:27 UTC

- reply

It’s probably the event that triggered this whole thing: https://x.com/zherbert/status/2082993276324319713

2026-07-29 23:11:56 UTC

- reply

Can you pls ask your agent to do a PR to https://github.com/Routstr/routstrd with the codex integration like we have claude code/pi integrations? 🙃

2026-07-28 22:02:04 UTC

- reply

when I tried buzz I already had routstrd configured inside my claude code and it worked without any extra setup. I’ve never used codex.

2026-07-25 09:06:47 UTC

- reply

WORD5 #659 6/6
⬛⬛⬛🟧🟪
⬛⬛🟧⬛⬛
⬛🟪🟪⬛⬛
⬛🟪🟪⬛🟪
🟧🟧⬛⬛⬛
🟪🟪🟪🟪🟪
https://otherstuff.ai/word5/

2026-07-24 08:19:03 UTC

- reply

Can you switch your model?

2026-07-24 08:14:27 UTC

- reply

Hey @npub1vl6…pu6v , how much does a request cost you? Do you know how to check?

2026-07-23 12:00:09 UTC

- reply

which tool are you using to convert credits to API?

2026-07-23 11:48:23 UTC

- reply

Would be cool to have an easy way to sell these AI credits for sats when the users aren't using them. 👀

2026-07-23 11:46:25 UTC

- reply

tbh, we cannot shut it down even if we wanted to lmao.
We can stop working on it ofc.

2026-07-23 11:45:57 UTC

- reply

Use kimi-k3 and glm-5.2
Or use gpt-5.6-sol.
the goal is to always have the option to switch. say no to being locked in on subscriptions.

2026-07-23 11:44:51 UTC

- reply

Use open source models.
Use sovereign tech.
Use Routstr. If you tell me why Routstr is a bad idea, we'll shut it down today :).

2026-07-23 11:18:39 UTC

We wanna become popular *before* a model gets banned you know :)
Waiting for a black swan to happen just to profit is kinda a weird feeling
#nevent1q…ux68

2026-07-23 11:16:36 UTC

- reply

Sir this is wendy's
GM

2026-07-23 11:13:42 UTC

Kimi k3's cache hit rates are always at 50%. Which upstream are you using @npub1m2m…r8p9 ?
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/dc4c6f806c36315fa49e0ca260d8cd64b90593963cee970114871cb749d29e88.png

2026-07-22 16:03:48 UTC

- reply

Is this compatible with concord or is it it’s own system spec?

2026-07-22 13:10:35 UTC

- reply

And if i download the mac app from their main website I'll need an email to sign up to a relay, no way to add my own relay. Maybe they'll fix all these things. But yeah, theyre building for normies.

2026-07-21 16:01:31 UTC

- reply

Bark! A good CLI based alternative to Cashu. You trade off privacy for self custody.
https://static.klipy.com/ii/39f2394ae36df6e199be9eb7c9fa1012/d6/87/77SZPXYP.gif

2026-07-21 15:58:38 UTC

Signal is a great messaging app but it is NOT Sovereign Tech.
What is sovereign tech? IMO, sovereign tech is where you have 100% control over every aspect of the solution you’re using. The entire stack!
But in Signal, I cannot even log into multiple accounts in the app. It’s been a decade since Signal has been around. They have chosen this for you.
That’s why I really want a group chat protocol on Nostr to go live. Concord seems to be the most reliable take on this problem. You have sovereignty over the whole stack. Let’s make it happen.
#nevent1q…quhf

2026-07-20 14:54:12 UTC

- reply

Once you topup to PPQ you cannot switch identities without losing all the funds. Whereas Routstr nodes allow refunding your balance with them. So easier to switch identities.
On the other hand, if you’re using TEE models there are no downsides only upsides. But if you’re using normal models, the Routstr node COULD also be logging your requests (along with potentially PPQ and OpenRouter). But I trust that the big Routstr nodes (in my WoT) are not logging.

2026-07-20 14:10:08 UTC

- reply

If you really want to use a direct API key, I recommend Routstr.otrta.me node.
We’re sunsetting api.routstr.com

2026-07-20 14:09:35 UTC

- reply

Hey, I highly recommend using https://routstr.com/routstrd instead of a direct API key. You can self host to use it with a simple URL and API key https://github.com/routstr/routstrd-auth
For this API key, can you pls DM me with details?

2026-07-20 14:07:16 UTC

- reply

Self plug but @npub130m…l29s was built for this. We have PPQ through Routstr as well. And we allow for easier switching between nyms (your API keys are refunded every 21 mins or you can manually refund). https://routstr.com/routstrd
This client pays for AI on a per request basis. You can choose a routstr node that proxies to PPQ. We also have TEE end to end encrypted models as well, also through PPQ.
And a nice TUI for usage tracking! https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/e4ae7835e4fa11aa8f1e8792775b81800fbbcc0d4a2b64492804d056f96d5636.png

2026-07-17 14:30:48 UTC

- reply

Requiring emails to sell VMs is a big turn off for automated VM deployments. If you’re being spammed a lot, you can require users to pay with Cashu and/or pay to create invoices.

2026-07-16 07:03:56 UTC

Concord is the most reliable nostr based Discord/Slack alternative app I’ve used so far. It works. Multiple clients and they interoperate well.
So far I’ve tried Armada and Accordion. They work great. 🚀
#nevent1q…vyxs

2026-07-16 07:00:13 UTC

I use @npub1h0u…rwx8 and it is the most reliable Nostr DM app to my knowledge.
#nevent1q…55mt

2026-07-15 06:05:27 UTC

- reply

Yep. Gotta fix this. Lots of work to do!

2026-07-15 05:32:21 UTC

- reply

and since we're both priced exactly the same and bitcoin prices fluctuate every min and our nodes fetch bitcoin prices are different times within a minute, sometimes your node is the cheapest and sometimes mine is 😂
that explains why you saw Private (E2EE) GLM 5.2 in your opencode intermittently haha

2026-07-15 05:30:26 UTC

- reply

Yeha, it shows the name from the first provider in the list.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/8806c1e1f2b069b72b4d867deb0395ed24eef06afbc8e312e975947488789183.png

2026-07-15 05:00:20 UTC

- reply

even after all this, there's something wrong when i use it with opencode. it works fine with pi. will investigate!

2026-07-15 04:57:07 UTC

- reply

this is your entry: https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/89fbb9960d794e1875061660acc2f6af0423ff6683c7fd312c25bc731bca2464.png

2026-07-15 04:56:18 UTC

- reply

okay, i tried it on my opencode and i see the issue:
I set a custom name for tinfoil-glm-5-2 on my node and that's why it shows up as Private E2EE GLM 5.2 whereas your default name is "glm-5-2". that's the tinfoil model.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/be78fefbdbe53ace710534cbe4cfe78a53dbd70157db1a5cae0418897e0d3a78.png
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/a0402bfccaa5634171b5a590dd7746501d56d54c3fff672dd456458d3816708a.png

2026-07-15 04:40:14 UTC

- reply

can you run "routstrd stop" and check again?

2026-07-15 04:29:00 UTC

- reply

hmm. weird. do you see different lists switching between each other?
can you run "ps aux | grep routstrd"? and see how many routstrd instances are running?
also run "routstrd --version" and tell me whcih one you're using?

2026-07-15 04:27:19 UTC

- reply

routstrd updates client lists periodically, you can also force update with "routstrd clients add --opencode"
yeah, we have a lot of little quirks like this. Different harnesses behave differently too. Also, be sure to have the latest version of routstrd "routstrd update"

2026-07-15 02:40:54 UTC

- reply

I use VMs. And give full access. I don’t use API keys anymore, anywhere. And deleting the whole VM matters little.

2026-07-15 02:33:45 UTC

- reply

Woah. 1 year!! Time flies 🥲

2026-07-15 02:23:45 UTC

- reply

—refresh haha

2026-07-15 02:23:09 UTC

- reply

I see privateprovider in my routstrd. Routstrd refreshes models every now and then. You can force refresh like this: routstrd providers list —fresh
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/217007d1ade659d3cbdc4d972a82029edce5a5d07d78ea902c496b14659d81d9.jpg

2026-07-14 10:11:13 UTC

- reply

One small config change and it should show up :)
And we're working with PPQ to enable TEE models through PPQ as well!
#nevent1q…ctjl

2026-07-14 10:08:24 UTC

- reply

hey, one small thing in config that needs to be done.
can you pls override models in tinfoil upstream and add "tinfoil-" before all the model ids?
For example: `tinfoil-glm-5-2` instead of just `glm-5-2` in the client-alias-id section.
You can do this in the admin UI in the providers section.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/9a92299b0e1fa93db9e1255187a8ce92c2e030ef5aad50fcdfcf475663421404.png

2026-07-13 09:22:52 UTC

- reply

cc: @npub1fet…jzfa @npub1ak6…y8fx

2026-07-13 05:03:01 UTC

Run a routstr node and help your fellow bitcoiners who live off grid use machine intelligence without breaking a sweat! 👻
#nevent1q…stz7

2026-07-12 07:08:37 UTC

- reply

Try out Sol, on par with Fable!
#nevent1q…pkp8

2026-07-12 06:44:51 UTC

- reply

But GPT models seem to like to write A LOT. If you ask for PR review, they'll write a book.

2026-07-12 06:43:52 UTC

- reply

If you're curious, look at the I/CR/CW/O column, it's showing Input, Cache Read, Cache Write and Output token data.

2026-07-12 06:42:54 UTC

OpenAI's prompt caching is the best I've seen. On par with Deepseek V4 prompt caching!
Sol, Terra and Luna are all great models.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/cffb91c92f5d0e2c171c48c32d36765cdc2ca3558eb44b0eba69207aea3f011e.png
#nevent1q…s8pd

2026-07-10 08:21:50 UTC

- reply

Hey, here's the tagged branch that you should checkout and run, it includes all the changes in 0.4.4 already. e2ee will also be merged to main soon, we just had a small thing that we needed to add, it's added now.
ask your agent to checkout this tag and restart your node:
docker compose down && docker compose up --build -d
https://github.com/Routstr/routstr-core/releases/tag/v0.4.5-e2ee

2026-07-10 06:12:45 UTC

- reply

You can DM me on your node npub if you don't want to dox yourself.

2026-07-10 06:12:25 UTC

If you're running this Routstr node, please reach out to me, I'm currently testing your node. @npub1jk3…9w0w
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/acaa8285a8c7264ae8604693e848a21b47a8892b936e352332e5e9487abbc33c.png

2026-07-09 15:36:20 UTC

- reply

Hey, tinfoil code is in a different branch, let me get that for you soon :).

2026-07-09 14:34:34 UTC

- reply

Yeah, that’s what Routstr nodes see when you use TEE models!

2026-07-09 13:13:39 UTC

We worked hard to bring this to production. Would love to have you all test it and give us feedback :)
PS: @npub1alj…g9jp @npub1226…grkj, thanks for recommending Tinfoil! I mention both of you in the article. https://confidentialinference.net/ is great!
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/5fcd8cbcb3941bae2758675e3a143b26c42b9e8de0b743fdcc7a12eec5a77c58.jpg
#nevent1q…u294

2026-07-08 16:23:55 UTC

Dario: I need a trillion dollar IPO, let's fuck with prompt caching on Fable 5 and bleed these plebs dry!
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/6b42f5601bc4774f0bed866309951b5eba2ce6db6fc4c9a315773e520f479b9f.png
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/260e235c45373defd33e6833df74618e6145d55f8d31a95fd55dcc04a91c5c84.png

2026-07-08 02:36:32 UTC

- reply

Once you do this, you’ll have a local proxy daemon that connects to Nostr relays to discover all the available Routstr node and fetches their cost info for models. It finds the cheapest and also audited Routstr node to route your requests for a model through that node. :)

2026-07-08 02:34:09 UTC

- reply

The link above helps you set it up for an agent. These are the commands:
bun install -g routstrd
routstrd onboard
routstrd receive 2100

2026-07-07 02:51:04 UTC

- reply

Which harness are you using? I recommend trying out Pi (pi.dev) or omp (omp.sh)

2026-07-07 02:49:17 UTC

- reply

Here, no accounts, no lock in, pay per request using bitcoin:
https://routstr.com/routstrd

2026-07-06 15:24:34 UTC

- reply

Did you try to compare it to GLM 5.2?

2026-07-06 15:22:25 UTC

- reply

Kimi 2.7 is okayish compared to GLM 5.2. GLM seems to be better. Don’t see the need to switch to Kimi.
For tasks that don’t require GLM’s expertise, I use DeepSeek-V4-Pro which is much much cheaper and would be comparable to Kimi K2.7 (benchmarks are lower but cannot beat on costs).
My current stack is:
Fable 5 - very very high end task planning.
GPT 5.5 - for high end execution.
GLM 5.2 - medium to high end execution (70% of my usage rn)
Deekseek V4 Pro - low to medium (20% of my usage)
And finally.
DeepSeek V4 Flash - low, grunt agentic work.

2026-07-05 07:09:06 UTC

- reply

Thank you for the diagnosis. Will look into this further.
Actually, Routstrd's default mode is streaming mode. Most of Routstr's requests go through routstrd, and fugu model worked for 8 requests, in streaming mode, before it broke.
Will try to reproduce this and see where the issue is.

2026-07-04 17:05:57 UTC

- reply

I understand your angle here. I mostly use GLM 5.2 these days. But I still think it's important to understand the gap between the closed source frontier and the open source frontier. People seem to like Fable a lot and i don't think it's bad to use better tools if available. It is strictly better to use Deepseek V4 Pro for medium tier tasks. It costs 100x cheaper.
Also, if there's a better tool than Routstr but it doesn't fully follow my principles, I'd use that. But Routstr is the best tool for my needs right now. :)

2026-07-04 16:48:29 UTC

- reply

Okay! that makes sense. We should handle cache read data in responses for custom upstreams so that your users don't get billed full prompt prices vs 90% discount cache read prices. Will verify this asap!

2026-07-04 16:45:15 UTC

- reply

Just use routstr:
#nevent1q…f3gu

2026-07-04 16:44:09 UTC

- reply

Use routstrd and pay per request using Bitcoin. never let Dario affect your workflow.
bun i -g routstrd
routstrd onboard (choose opencode)
routstrd receive 4200 (scan lightning invoice in yoru terminal).

2026-07-04 16:40:30 UTC

- reply

fugu-ultra is available from a lot of upstream (through Openrouter i think) but fugu is available only from you. idk the difference.
will get back to you on updating the node. I'll wait for @npub1ygj…lajc to get back to me first.

2026-07-04 16:38:03 UTC

- reply

I just wanted to try it out and it was definitely NOT worth it lol. Both the Fable + GPT 5.5 combo and Deepseek V4 Pro ended up doing the exact same thing.
Maybe I should try again with a more difficult task. Will report back.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/81aedcf36a4ac28435f3affab6c62e648708047403c00d80f7fff732085e56f2.png

2026-07-04 16:27:53 UTC

- reply

Insanely expensive and ended abruptly. Got:
` Unexpected token 'd', "data: {"id"... is not valid JSON`
Weird.
Okay, I think your node isn't taking care of cache pricing as it's a custom upstream. I see that the response contains the cache read data in it. The current main fixes this for some direct upstreams like deepseek, maybe it also works for custom upstreams. \
@npub1ygj…lajc
In the usage chunk i got:
```
\"usage\": {\"prompt_tokens\": 45763, \"completion_tokens\": 193, \"total_tokens\": 45956, \"prompt_tokens_details\": {\"cached_tokens\": 43520, \"orchestration_input_tokens\": 0, \"orchestration_input_cached_tokens\": 0}, \"completion_tokens_details\": {\"reasoning_tokens\": 77, \"orchestration_output_tokens\": 0}
```
But i was charged full prompt token price here.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/707df5f614ec413dbe35c87ea9efa84b714e4432daeea19c45a7156dfe490991.png

2026-07-04 16:04:36 UTC

- reply

Not yet. Let me try rn!

2026-07-04 15:44:45 UTC

- reply

I think this probably has something to do with the way Anthropic does cache reads and writes.
#nevent1q…xf3j

2026-07-04 15:42:57 UTC

- reply

Okay, I think I figured it out. Unlike most other modern LLMs, that write to cache incrementally, Anthropic models only seem write to cache initially and never again?

2026-07-04 15:38:42 UTC

- reply

Caching seems to be working fine for @npub1ygj…lajc!! I might be cursed :(
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/1acae99ad499e868073a5f9edf8490af32ddf713c28780bb301256bb4a0ee5f0.png

2026-07-04 15:21:04 UTC

- reply

Deepseek, same prompt as Fable, cost me 57 sats.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/4e581927cc8ebd1db458f99b92dbef7b7af914c7b3ec9e6b6b6693780781187c.png

2026-07-04 15:20:28 UTC

Insane that Fable doesn't write to cache after the initial system prompts. Notice that it wrote 4.9k tokens to cache when the session began and never again. I feel rugged! This planning session cost me 8.9k sats. 😭
Deepseek V4 Pro caches everything and costs much less, proportionally. (see image below)
ps: this is routstrd top that shows cache r/w data
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/adef595ebf6d8773e4afa1985e500e93d3c47de031293523e7c393f09ae5d3a6.png

2026-07-04 13:45:25 UTC

Why aren't we inviting Bitcoiners to Nostr?
I invite @npub1gjy…ektz and @npub1ate…mf3z to come back to Nostr.
I'm currently trying bark-cli
https://second.tech/docs/getting-started/bark-cli
Let's scale Bitcoin!

2026-07-04 13:24:22 UTC

But but i paid for the blue check mark!! These scammers!! 😭
And adding phone feature is broken lol. App by a $1.5T fiat company.
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/f6ee61f67144df88171b163255682d96144961bfb75610b373947bc8c55dfdfc.png
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/cda3d1548e24855e1d6ddaba50ec79623edcd323a6e369cc8936958ecf8a30e5.png

2026-07-04 13:23:24 UTC

But but i paid for the blue check mark!! These scammers!! 😭
And adding phone feature is broken lol. App by a $1.5T fiat company. (pls also try and shit on our bitcoin apps ahah)
https://npub1ftt05tgku25m2akgvw6v7aqy5ux5mseqcrzy05g26ml43xf74nyqsredsh.blossom.band/f6ee61f67144df88171b163255682d96144961bfb75610b373947bc8c55dfdfc.png

2026-07-04 13:17:08 UTC

- reply

no, we have x-cashu mode where it's not linked by payments at all. The response contains the refund cashu token.
plus, since you're issuing tokens for a fixed price that the user paid for beforehand, users sending a fixed 42 sat cashu token per request eliminates the linking payments issue.

2026-06-30 09:43:49 UTC

- reply

Please give @npub130m…l29s a try as well. Especially https://routstr.com/routstrd a tool that discovers inference providers via Nostr and fetches you the best price for any model. We have also just added E2EE models like Maple and PPQ.ai.

2026-06-30 09:16:42 UTC

- reply

Anyone can run open-source models on their own hardware and sell their inference on @npub130m…l29s for sats right now.
Someone is already doing this here: https://routstr.gnostr.cloud/
Running on their DGX spark.
All you have to do is run https://routstr.com/routstrd and choose that model. It's a pretty decent model.