Provide an API for the Assistant

ncwade

Would love API access, more than happy to pay for all usage via the API. Gives me A) a single subscription, and B) a source I trust to sit in front of the companies I don’t.

madhatter

ncwade

B) a source I trust to sit in front of the companies I don’t.

I used to think like this earlier but the impression I get is that it's just a case of "we have these LLM providers but if you care about privacy, just go and look up each one's privacy policy".

I am serious about privacy and had a couple of questions about privacy and the assistant but from the response I received, I failed to come to any interpretation which says Kagi gives us more privacy in this area.

I do concede that we get the convenience of having multiple providers in one place (and each of them being able to use Kagi search).

slate

I feel that people are completely unrealistic about the costs of providing an API to "access their credits".

In a previous post Vlad stated that search "Costs us about 1.25 cents per search". If that search count lines up with what is shown in the billing portal then I was burning over $30 in search costs a month for a $10 subscription. And that didn't include any LLM costs.

I bumped my subscription up to $25 only to find I burned $1 in the first half of the first day as !code now maps to Claude Sonnet 4 in the $25 plan instead of Kimi K2 as it does in the $10.

This seems like a huge mistake. The Ultimate plan costs are simply unsustainable.

The costs of Anthropic models like Claude Sonnet 4 are absurd. OpenRouter shows Sonnet 4 costing $3-$6/M input tokens and $15-$22.50 output. Meanwhile Kim K2 0905 on OpenRouter ranges around $0.50-0.75/M on input and $2-$2.50/M. No wonder Ultimate users are net negative for Kagi the minute they touch the assistant.

Also keep in mind Anthropic and OpenAI are charging these insane API costs and subscription plans costing hundreds of USD a month that still get rate limited. And these two companies are still burning cash and wildly unprofitable.

Look up Edward Zitron's article "Why Everybody is Losing Money On AI". These prices are subsidized and will get higher over time as they run out of other people's money to burn. And while the cost of some unit of inference on a GPU might be going down there are two trends that will make it more expensive. The new models are generating so many reasoning tokens that you will pay for and the cost of electricity in the U.S. is going up and will continue to go up. Kagi is a reseller of LLM APIs here. The API providers will have to raise costs dramatically on Kagi to reach profitability themselves.

I don't see any future were unlimited AI assistant works for Kagi without it focusing on local hosting of smaller open weight models to keep costs as low and controlled as possible. API access will have to be pay per call.

I use Kagi assistant because the search integration is amazing and I prefer LLM models to use search results to answer my question with citations for everything instead of any "world knowledge" baked into the LLM.

I don't see how this functionality is able to be offered long term with a measly $25 plan.

I'd love to see a pay as you go search API where Kagi makes a small profit with each call. And I fully expect any future subscription plans with the Assistant UI to only offer X base credits and you'll have to pay extra for more.

id

slate

slate I don't see how this functionality is able to be offered long term

By charging a premium on API calls, like openrouter does? Nobody is expecting to get unlimited access for $25.

maedoc

just a word of support that I would also pay higher rate to be able to use Kagi tomorrow, next year, 5 years from now. I think a several ideas here are worth exploring: small credit + pay per usage + integrated API usage.

I also second the "offload to local" idea: I can run a 7 to 24b model locally and provide the openai compat endpoint but I'd like to do so with the rest of Kagi features intact (else I'd have moved to Openwebui or similar; tho even then I don't want to, Kimi k2 is great).

edit I would see the local offload working similar to how GitHub/GitLab allow registering self-hosted CI runners. I use this for resources not available on their clouds (GPUs, large datasets, many cores, etc).

I also realize my comment contributes to derailing the conversation from the original topic. In the meantime I also saw that one can move credit from account to API which seems to largely address the original topic in short term. I'll go try that out 😃

Matted4505

Just my two cents on this topic: I'm a Pro user who would very much like to upgrade to Ultimate if API access were provided. Currently, the cheaper models are perfectly fine for typical question-based searches (my monthly use is around $1). However, my struggle is that I can't use an API to connect to my Home Assistant instance, improve text grammar, or integrate AI into my code editor when needed. For these use cases, I'd have to subscribe directly to one of the LLM companies or use a service like OpenRouter. If I do that, I don't see the value of an Ultimate subscription beyond supporting Kagi (which matters to me, but $15 is a lot for just that).

I don't believe these use cases would increase my monthly costs to an exorbitant level, so I'm really in favor of reducing the Ultimate LLM use to $10. Either way, users who hit their limit should consider switching to a different type of service or paying extra for increased limits, at least that's my opinion.

stefreak

I personally need to use multiple subscriptions next to my Kagi ultimate subscription in order to use agent-based workflows. I'd rather spend more on Kagi to support your other efforts and avoid the "subscription creep" with subscriptions for several providers.

It would even be OK for me if the inference API would not be part of the ultimate plan, or if there's a reduced "fair use" limit for the API use case. But then there would need to be an "add-on" subscription (e.g. 20$ additional inference use + including 5000 searches via Kagi API/MCP) or the ability to top up on-demand fees.

WhatMatters

I do hope it's not a case of "it was fun while it lasted" because $25 is just the right amount for me, and I use search extensively and often hit my LLM limit in about 20 days, so need to renew earlier. But $75... that's just... boutique territory really. I understand why they're still keeping $25 around, because mindshare matters more than just being able to generate money for a certain time and then slide into oblivion, but I do indeed hope it's possible to be profitable and stay more or less within $25 limits.

Fagioletti

Why not leave the pricing as is and set limits in one or the other form (my mobile provider gives me X GB/month. When not using it up, unused go into the next month as a buffer. max allowance is 2 months worth of data). Something easy and understandable.

pbronez

My preference is:

Ultimate Plan stays around $25/month and includes access to everything... with usage limits that make it sustainable for Kagi
Ultimate Plan includes access to all the APIs. Using them counts against that sustainable limit.
Usage beyond the sustainable limit costs extra
extra costs are billed in buckets that keep processing costs sustainable. Maybe $5 or $10 increments
roll over unused overage balances. If I use $2 of APIs beyond the sustainable limit, and you bill me $5 to keep processing costs in check, I should have $3 in credit for overages the next month.

I want the same capability on the team plan, but across multiple users. This would give me a central way to tell my workforce "you want the AI? here is the AI." I'd probably still maintain subscriptions directly with the major providers, but I'd buy many fewer seats.

IvanCherepanov

+1 for OpenAI-compatible API for Kagi-enabled assistants. We have all our agents at work built with OAI API and it's a blessing in terms of integrating them everywhere.

Costs should absolutely be sustainable for Kagi, from that point of view any kind of model with fixed predicted usage will be prone to abuse, usage-based pricing seems to be the only reasonable way - where Kagi just transparently takes some margin (or acts similar to OpenRouter and negotiates discounts from providers for bulk use).

mateja

Ultimate plan user here and I appreciate the discussion so far.

Here's an idea: two buckets for AI credits: (1) monthly and (2) extra.

The monthly bucket is included in your plan, and the number of credits is based on your plan level.

The monthly bucket renews each month but credits from the monthly bucket DO NOT roll over. Whatever you don't use each month, you lose.

Credits for the extra bucket can be purhased at any time, and you DO NOT lose these credits each month. They stay until you use them. (Or you lose them eventually if your account becomes inactive for a year, let's say)

All API calls would draw from the monthly bucket first, and then the extra bucket.

All credits would be priced sustainably to ensure Kagi's longevity. Potentially a higher margin on monthly plan credits than extra credits, but this is up to the Kagi finance and product teams.

Essentially this is just the existing plan with the addition of the extra bucket instead of having to renew early when you run out of credits.

Coincidentally, this is the same plan my metro card uses!

Thoughts?

TheSingular

mateja I agree with that. If that happened, I could totally picture myself just plugging the API into an MCP server and using it in my AI workflows. Especially for search.

pbronez

mateja That sounds good to me. I like the clear delineation of the "extra" bucket.

FWIW, I've seen several of my AI-focused vendors adopt a Seat + Usage model like this. It's complex, but done well I think it can align incentives correctly.

The upfront per-seat cost needs to cover the vendor's fixed costs to support my account and make a profit without being so expensive that I start scrutinizing who gets a license. The main way vendors can reduce this upfront cost is reducing the included usage.
The Usage cost needs to have good margins for the vendor so they make more money when the product is good and we use it more
I want options to reduce costs if I wind up with consistently high usage. These can include bulk discounts, pre-pay discounts, etc. They should reduce but not eliminate the vendor's margin.

That's all very enterprise-y. From a personal perspective, I love my Ultimate Account, and I would love it more if it cost less. I don't expect "Unlimited" usage. I'd prefer a lower subscription cost with the extra bucket.

TomatootamoT

Is there any idea on if there is going to be a rate limit? Or will there be different limits based of plan type?

pineapplemachine

I'm also interested in this, especially since the Kagi-specific Quick and Research assistants came out and they do not seem to have any existing API that I could use. I'm not that particular about pricing or billing. It would be nice if API usage would first go through the token allowance already included with my Ultimate plan, but that's not an absolute requirement for me. I have appreciated the usage-based billing for FastGPT, though, meaning that while my API usage is low, so is the pricing.

« Previous Page Next Page »