How to Cut OpenClaw API Costs: Model Routing, Caching, and Batching

By Beau Johnsonยทยท8 min read

Quick answer

One builder cut his four-agent OpenClaw bill from $420 to $168 over 20 days, about 60% less. He sent simple tasks to cheaper models like Claude Haiku 4.5 instead of Opus 4.6, turned on prompt caching, batched support emails, and capped output length.

In this video I go through how one builder cut the cost of his OpenClaw setup by about 60%, from $420 to $168 over 20 days, without losing any quality. If you run AI agents and your API dashboard keeps surprising you, this is for you. The four fixes he used are the same ones I use on my own setup.

I recorded this in March 2026. Model names, prices, and OpenClaw settings change fast, so check the current docs and pricing pages before you change anything.

What I cover in this video

  • 0:00 Why most people running agents pay far more than they need to
  • 0:47 His setup: four agents and a $420 bill after 20 days
  • 2:19 Sorting every API call by how complex the task was
  • 5:23 Fix #1: model routing, the biggest win
  • 8:26 Fix #2: prompt caching, a five-minute change
  • 9:11 Fixes #3 and #4: batch processing and output caps
  • 10:42 Before and after costs for each agent
  • 13:48 My four-tier model framework and an action plan

The audit: where the money was actually going

The builder (his original Reddit post is here) ran four agents for a side project. One handled website analytics, one wrote blog content, one reviewed code in his GitHub repos, and one handled customer support. It's a very standard setup. After 20 days the bill was $420, roughly $21 a day, for a project with no paying customers.

He didn't shut everything down. For the next 20 days he logged every API call: input tokens, output tokens, which agent made the call, and what task it was doing. Then he sorted the calls by complexity:

  • 70% simple: FAQ answers, basic formatting, one-line summaries, "what's the status of this deployment?"
  • 16% medium: longer email drafts, moderate code reviews, multi-paragraph summaries
  • 9% complex: deep code analysis across multiple files, long-form blog content
  • 6% heavy reasoning: architecture decisions, complex multi-step debugging, new problems

The problem was that all four agents ran on Opus 4.6 for every call, even yes-or-no questions. At $5 per million input tokens and $25 per million output tokens, that's like hiring a brain surgeon to put a band-aid on your knee.

Fix #1: Model routing

He set up tiered routing. Simple tasks go to budget models, medium tasks go to mid-tier models, and only the complex 6โ€“9% goes to Opus. For customer support he chose Claude Haiku 4.5, which costs $1 per million input tokens and $5 per million output tokens, five times cheaper than Opus on both. Switching models alone took the support agent from $175 to $85, with the same system prompt and the same customer experience.

He also tested Kimi K2.5. It handled long context well but was wordy, so he had to add output caps to keep it in check.

The big lesson: budget models don't behave like Opus. When he first switched to Haiku, the tone and format were off. Haiku doesn't pick up tone or style from vague instructions. Once he rewrote his system prompt to spell out the reply structure, tone, length, and format, the replies were solid. Treat a cheaper model like a new employee who can do the job but needs clear instructions. This is also why I give my agents a voice profile trained on my own content, like past emails or Zoom recordings.

In my own setup, Opus runs my execution and coordination agent. Lower-cost models like Sonnet 4.6 and Gemini 2.5 Flash handle cron jobs and basic web scraping.

Fixes #2โ€“4: Caching, batching, and output caps

Prompt caching

This was the easiest fix. He turned on prompt caching for his support agent, which took about five minutes, and input costs dropped about 40%. Every API call sends the full system prompt. A 2,000-token prompt across 500 calls a day adds up to a million tokens of system prompt alone. With caching, you pay only a fraction of the cost for the repeated part.

Batch processing

His support agent used to make an API call for each email as it came in. He switched to 30-minute batch windows, which cut that agent's calls from about 3,000 to 1,400 over 20 days. For email support, a 30-minute delay is fine.

Output caps

The support agent was writing 500+ token answers to questions that needed two sentences, like a five-paragraph security essay in reply to "How do I reset my password?" He set a 300-token limit, and customers noticed no drop in quality.

The results and the four-tier framework

Total cost went from $420 to $168 over 20 days, a saving of $252. That's almost $400 a month. By agent:

  • Support: $175 โ†’ $58 (Haiku, better prompting, and caching)
  • Code review: $112 โ†’ $48 (Sonnet for routine PRs)
  • Content: $85 โ†’ $40 (shorter prompts, Sonnet for first drafts)
  • Analytics: $48 โ†’ $22 (mostly batching)

Model routing and prompt caching made up 85% of his total savings. Batching and output caps were smaller wins.

Here's the framework I use for my 13 AI agents:

  1. Tier 1, simple: FAQs, formatting, lookups, yes/no questions. Use the cheapest model that passes your quality bar, like Haiku, Gemini Flash, or MiniMax. You can compare options on OpenRouter.
  2. Tier 2, standard: longer drafts, moderate analysis, multi-turn conversations. Use mid-range models like Sonnet, GPT-4, or Gemini Pro.
  3. Tier 3, complex: deep analysis, long-form content, multi-file code review. Use Opus.
  4. Tier 4, heavy reasoning: architecture, complex debugging, new problems that need extended thinking. Use the top frontier reasoning model, which when I recorded this was Opus 4.6.

Tiers 1 and 2 make up most agent work, around 70โ€“85% of calls. Paying premium prices for that work wastes money.

My action plan:

  1. Open your API dashboard (console.anthropic.com or your OpenRouter usage page) and break costs down by agent.
  2. Check whether each agent uses the right model for its job.
  3. Turn on prompt caching.
  4. Look at your most expensive agent and ask whether you could send its tasks to a cheaper model.

Key takeaways

  • Audit first. The builder's biggest lesson was that once he could see where the money went, the fixes were obvious.
  • Most agent tasks are simple. Running Opus on everything is like using a Ferrari to get groceries.
  • Cheaper models need more explicit prompts. Spell out tone, format, and length.
  • Prompt caching takes about five minutes and cut his support agent's input costs about 40%.
  • Batching and output caps add smaller but real savings.

FAQ

How do I reduce my OpenClaw API costs?

Start by auditing your API usage by agent and task. Then send simple tasks to cheaper models, turn on prompt caching, batch tasks that don't need instant replies, and cap output length. In the video, these four fixes cut one builder's costs by about 60%.

Is Claude Haiku 4.5 good enough for a customer support agent?

In this case it was. The builder's customers noticed no difference after switching from Opus. He did have to rewrite his system prompt to spell out the tone, format, and length he wanted, because Haiku doesn't pick these up from vague instructions.

How much does prompt caching save?

Turning on prompt caching for the builder's support agent cut input costs about 40%, and setup took about five minutes. It works because the system prompt sent with every call gets cached, so you pay only a fraction of the cost for it.

Which model should I use for each OpenClaw agent task?

Use a four-tier approach. Send simple tasks to budget models like Haiku, Gemini Flash, or MiniMax, and standard work to mid-range models like Sonnet or Gemini Pro. Save Opus for complex analysis and heavy reasoning.

Chapters

  1. 0:00Are you overpaying for OpenClaw?
  2. 1:30The $420 wake-up call
  3. 3:00Auditing 13,500 API calls
  4. 5:00Fix #1: Model routing (the biggest win)
  5. 7:30Fix #2: Prompt caching (5-minute setup)
  6. 9:00Fix #3: Batch processing
  7. 10:00Fix #4: Output caps
  8. 11:00Full before vs after breakdown
  9. 12:30The 4-tier model framework
  10. 14:00Your action plan (do this today)

Transcript

Read the full transcript

0:00I'm going to be going over something that every single person running Open Claw and AI agents needs to know because if you're running agents right now, there is a very good chance you're spending way more money than you need to. I'm not talking about pennies more. I'm talking about double to triple what you should be paying. And the worst part is is you probably don't even know it. Have you been opening up your API usage dashboard and getting surprised? Well, in this video, we're going to fix that and go over to the reasons why that could be. So, there's this guy that was spending $420 uh after 20 days of use on his Open Claw setup just running four AI agents, nothing crazy. And this was like nothing that required a computer science degree. It's a very simple pipeline system, but he was able to bring it down 60% basically producing the same quality

0:47with the same agents and producing the same results just for 60% less. So, I'm going to set the stage here. This guy was running Open Claw for his side project. There wasn't anything fancy about it. It's basically just four agents, you know, one handles the website data analytics, one creates blog content, one does code review for his GitHub repos, and the other one handles his customer support. It's a very standard setup. It's probably very similar to what a lot of people are running right now or planning to run. So, you know, one day he opens up his API dashboard and the number shocks him. $420 for just 20 days worth of usage. That's $21 a day and for just a side project that wasn't making sense. You know, he had this moment that I think a lot of us have where you look at the bill and you're like, "This isn't reflecting for the output that I've been

1:33getting. This feels expensive for these agents and what they're actually doing." Most of their tasks are pretty simple. And why is this costing so much? He does something smart though. Instead of just complaining about it and shutting everything down and giving up, he decides to figure out where the money's actually going. He logs for the next 20 days where these API calls are going, every single API call, every token, you know, input tokens, output tokens, which agent with which call, and what task was it doing. Basically just a full audit of the whole pipeline. And what he did find is what how he was able to change his complete pipeline overall. You know, 30,500 API calls across four agents. That's insane. Even for my system, that's crazy. And the raw numbers were not what matters here. What matters is that what

2:19he found when he categorized these calls by complexity, this is the part that changed everything. 70% of those calls were simple like FAQ answers and basic text formatting, basically one-line summaries, things like, "Summarize this PR that changes a readme typo." It's crazy, you know, or "What's the status of this deployment?" It's very simple stuff. And it's the kind of work that literally any model can handle. Gemini 2.5, MiniMax, Kimmy K2. You don't need the most powerful AI model in the world to answer what your business is actually doing. 16% were medium complexity, you know, like longer email drafts, moderate code reviews, where you need to understand a little more context. Multi-paragraph summaries worth a decent

3:04model, but still not the top tier. 9% were actually pretty complex where it required deep code analysis across multiple files, long-form blog content, things that require holding a lot of context and producing thoughtful output. And only 6% needed real heavy reasoning, you know, those like architecture decisions that you would need like an Opus 4.6, complex multi-step debugging, actual novel problem-solving where the model actually needs to think hard. The main problem about all this is all four of his agents were running on Opus 4.6 for every single API call, the most expensive model out there. Even the ones that were basically just reformatting a sentence or answering a yes or no question. He set up his agents at first

3:51in picking Opus as the frontrunner. He wanted the best results and I understand that. If you have the money to spend, go ahead. But for a side project, that adds up quick, especially if you don't have any paying customers. So, he configured it across the whole board and then he basically forgot about it for weeks. And then the bill came in and he saw that wow, he was shocked. He needed to make changes. And I've made this mistake before. When I first started running my agents on Open Router, I was spending $10, $15 a day and I just thought that that was not going to add up for what I was getting in return. So, it is the best model, Opus, and I do have it for my execution and coordination agent, but everything else like the cron jobs is handled with lower end models like Sonnet 4.6 and Gemini 2.5 Flash for basic web scraping. And the results are

4:37amazing. But when you look at the per token pricing, Opus costs $5 per million input tokens and $25 per million output tokens. For simple tasks, that's like hiring a brain surgeon to put on a band-aid on your knee. It works, but you're massively overpaying. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full Open Claw course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching,

5:23so you could book a call with me. We could share screens and I can help you get Claude code or Open Claw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. So, here's what he did to fix it and I want you to pay attention because these are the exact same strategies I use for my setup. The number one fix is you need model routing. This is the biggest win. Instead of running Opus on everything, he created a tiered routing system where simple tasks go to budget models, medium

6:10tasks go to mid-tier models, and only the genuinely complex stuff, that 6 to 9% usually, gets sent to Opus, the deeper reasoning model. He tested a bunch of models to find the right ones for each tier. For customer support, he landed on Claude Haiku 4.5, which is great. It's fast, it follows formatting instructions well, and it costs $1 per million input tokens and $5 for a million output tokens. So, if you compare that to Opus at 5 and 25, that's five times cheaper on input and five times cheaper on output. For the customer support agent alone, this dropped his cost from $175 to $85 just by switching the model. Same

6:55agent, same system prompt, and same customer experience. He also tested Kimmy K2.5. They have 2.7 out now, which is even cheaper at 60 cents per million input and $3 per output. It handles the long context well, which is nice for multi-turn conversations, but he found it tends to be more verbose. It writes three paragraphs and then one sentence would do. It just gets a little too fluff, too much fluff. So, he had to set output caps to keep it in check. Good to know if you're thinking about trying it. Now, here's something really important that he mentioned that I want to make sure you hear this. Not all budget models behave the same. Haiku was reliable for support, but you have to be very explicit in your prompts. It won't infer tone or style from vague

7:41instructions the way Opus does. So, when he first swapped to Haiku, the responses were kind of off, wrong tone, wrong format, but he once wrote his system prompt to spell it exactly how he wanted replies structured, the tone he wanted, the length he wanted, and the format he wanted, and it was solid from there. This is why I always go back to making sure your agents have a voice profile system or a humanized file based and trained off of your content. So, if you have Zoom recordings, thousands of emails you've sent in the past, feed that through your agent that's writing a lot of the copy material and you want it sounding in your voice, and have it follow those directions. You know, this person's customer didn't notice any difference when he switched from Opus to Haiku. You just make sure you need to

8:26spend the time training it on how you want it to speak and talk. You want it to be in your voice, right? And that's the biggest lesson right there is, you know, when you switch to a cheaper model, don't just swap it in and expect the identical results. You do need to do some testing. Adjust your prompts, be more specific, be more explicit about what you want. The model can absolutely handle the work, but if you have to meet it halfway, think of it like training a new employee. You know they're capable, but you do need to give them clear instructions until they learn the ropes. The second fix is prompt caching and this was pretty much honestly the easiest one of all. And he turned on prompt caching for his support agent and input cost dropped about 40%. It took him 5 minutes and that's it. Pretty crazy. Open Claw, if you're running Open

9:11Claw right now and you haven't turned on prompt caching, you need to stop this video and go and do it. Here's why prompt caching matters. Every time your agent makes an API call, it sends the entire system prompt along with the conversation. If your system prompt is 2,000 tokens and you make 500 calls a day, that's a million tokens in just system prompts. With caching, the model remembers the system prompt from previous calls and you only pay a fraction of the cost. It's literally free money sitting on the table. Fix number three is batch processing. His support agent was handling customer emails one at a time. Every email that came in, boom, immediate API call. He switched to 30-minute batch windows. So, instead of processing 50 emails

9:56individually throughout the day, it batches them up and process them in groups every 30 minutes. His call count for that agent went from roughly 3,000 to 1,400 over the 20-day period. Cut almost in half, fewer calls means fewer tokens means less money spent. And for customer support, a 30-minute delay is totally acceptable. Nobody expects an instant reply to an email anyways. And if they do, that's just weird. That's totally AI. You want to give us some time. The fourth thing that you could do when you're running Open Claw and you're trying to optimize your tokens is output caps. This is one is very subtle, but it does add up fast, especially if you're building and building and building. Basically, his support agent was generating 500-plus token responses to

10:42questions that needed two sentences. Someone asked, "How do I reset my password?" And the agent writes a five-paragraph essay account about security best practices, two-factor OAuth, why strong passwords matter, and here's a link to our blog post about cybersecurity. Nobody needs that. He set a 300-token limit, customers didn't notice any difference in quality, and the bill got much lighter. So, to look at the results in the full breakdown, before optimization, 20 days in was $420. After optimization, 20 days later, $168. About a 60% savings. By agent, support went from $175 to $58. Haiku plus prompting and caching did the heavy lifting. Code review went from $112 to

11:29$48 using Sonnet for routine PRs instead of Opus. Content went from $85 to $40. Shorter prompts and Sonnet for first drafts. Analytics went from $48 to $22. Batching was the main lever there. Total savings, $252 over 20 days. And I know it doesn't sound like a lot of money, but a lot of people that are getting started in this AI space don't want to be spending a ton right off the bat. That's why I'm making this video to help you save some money. You know, that's almost $400 a month in savings for the same amount of work. Same quality, same customer experience, and nobody noticed. And nobody complained. If you're struggling to keep up with content, I'm about to save you about 40 days worth of work. [music] I built something called Content Machine. It's 10 AI agents that run on the Open

12:17Claw orchestration, and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning, I wake up [music] and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. It works for any niche, fitness, finance, real estate, marketing, whatever you're building, and it is 100% completely customizable to your use case. So, you get the Mission Control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the

13:03community. So, you plug in your own thing, and it molds it to you. It learns how you talk, and it writes so it doesn't sound like AI slop. $97 one-time, it's not a subscription. I'll put the link down below, and you'll thank me later. Now, I'm going to talk to you why this matters so much and how I think about it for my own setup. I run 13 agents on my Mac Mini, and I think it costs I run 13 agents on my Mac Mini, and I think about costs constantly. Because when you're running all those agents, even small optimizations multiply pretty quickly. So, if I save $10 a month per agent through a better model routing, that's $130 a month. Over a year, that's $1,560. Just from picking the right model for each task. And that adds up quick. It's not nothing. This is something we talk

13:48about all the time in the community because when people first start building with Open Claw, the first thing they do is pick the most powerful model and slap it on everything. It works, but you're going to be spending a ton of money. The results are great, but then you look at your bill in at the in the month, and you get slightly depressed. Where did all the money go? And the answer is always the same. You're using a Ferrari to go get groceries. So, let me give you the framework I use. You know, think about your agent tasks in like four tiers. Tier one is like simple FAQ answers, formatting, basic lookups, and yes or no questions. Use the cheapest model that gives you acceptable results. You could use Haiku, Gemini Flash, Mini Max. Check out those models on Open Router. It's like one of the best marketplaces out there. So, basically, test whatever passes your quality bar.

14:33The second tier is standard longer drafts, you know, moderate analysis, multi-turn conversations that need same context awareness. Sonnet or GPT-4, Gemini Pro, solid mid-range models. Use those. The third thing, more of complex reasoning and deep analysis and long-form content, you know, even multi-file code review. Anything that requires basically holding a lot of context and producing nuanced output. That's where you want to use a model like Opus. Tier four would be like reasoning and heavy lifting and architecture decisions and complex debugging with multiple failure models, novel problem-solving where the model needs extended thinking. Opus with Ultra Think or whatever the latest frontier reasoning model is, and I still think that's Opus 4.6. The inside is this.

15:19Most of your agents' work, and I mean the vast majority of it, is tier one and tier two. 70 to 85% of all the API calls you're making. So, if you're paying a premium prices for that work, you're going to be throwing money away. Here's what I want to do after you watch this video. Step one is go look at your API dashboard. If you're using Anthropic, go to the console.anthropic.com. If you're using Open Router, check your usage page. Look at your costs. Break them down by agent if you can possibly. Set it up in Mission Control, too. It's much easier to track costs that way. Second step is ask yourself, "Is each agent using the right model for what it's actually doing?" The third step is turn on prompt caching if you haven't already. Come on now. That's a five-minute win. Step four is look at your most expensive agent and ask,

16:05"Could I route that task to a simpler, cheaper model?" Nine times out of 10, the answer is yes. The guy who did this said the biggest lesson was that he had no idea where the money was going. The moment he could see the numbers, the fixes were obvious. Model routing and prompt caching alone saved 85% of his total savings. Batching and output caps helped, but they were smaller wins. And if you're building with Open Claw and you wanted to learn how to do setups like this and set it up properly, the right way from the start so you're not burning through a ton of money, that's exactly what we do inside of Shipping School. 114 members in over 11 days. Thank you to all that have joined. We do have courses on Cloud Code and Open Claw for setup, but we're really focused on community first, which is why we do six live boot camp calls every single week.

16:51We don't just show you how to build agents, we show you how to actually run them efficiently and how to optimize them. We want you to turn them into actual business tools that make more money than the cost. That's ideal. I'll put the link in the description down below, and if you haven't subscribed already, please do because we release about four to six videos every single day. Leave comments down below. Were you surprised about your first bill? That's why I made this video. Apply what we talked about here. Give a link to this video to your agent and say, "Do what he just said. You're going to save money." We'll see you in the next one. Bye-bye.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading