# Claude Haiku 5.5 Earned a Spot in My Real-World Workflows > Source: > Published: 2026-10-12 14:31:02+00:00 Last week I was staring at my AI usage, wondering if $4 was too much to pay to complete my project. Cost-avoidance is a feature and flaw of my personality, and it’s made Anthropic a recurring thorn in my side. They’ve been slinging models that accidentally hack websites and possibly threaten the species, while their workhorse Haiku 4.5 model sat in the parking lot with no updates and a price tag that made it look like the gas-guzzling used car it really was. Then Haiku 5.5 landed a few days ago, and I had to open up a new parking space at my “useful models club.” This is for people who run models on their own hardware and care about the bill at the end of the month, not just the benchmark score. If you’re a developer or maker with a decent GPU, a low tolerance for waiting, and a genuine distaste for paying for things, this is your article. If you want the smartest model regardless of cost, or if you don’t care whether your AI runs on your machine or someone else’s, keep scrolling. As a reminder from [my past articles](https://medium.com/@codysandahl/forget-the-benchmarks-a-real-world-local-vs-cloud-llm-cage-match-4a29cc5f1d40), I am a cheapskate who doesn’t like to pay for things. **Cost matters**. I don’t have any huge secrets to protect from the AI APIs, so I run local models to keep my costs down and not for data privacy reasons. I am also impatient along with being a cheapskate, so I don’t like to wait for the results I don’t want to pay for. **Speed matters**. I have a Strix Halo AI Max+ 395 with 96GB dedicated to VRAM ([more info on my local setup](https://medium.com/@codysandahl/i-finally-have-my-dream-local-ai-stack-and-it-runs-on-amd-c9f4935788f9)), so I can run some sizable models but they don’t run super quickly. The current king of the AI hill on my machine is [Ornith 1.5 35B-A3B](https://ornith.ai/ornith_1_5.html). Ornith uses a novel self-improvement loop on top of Qwen’s models, and the 35B-A3B version uses a mixture-of-experts architecture to speed up inference. If those words don’t mean anything to you, just know that it’s smart and fast if you have enough VRAM to fit a 21 GB model. My Strix Halo machine is my daily driver, not just an AI appliance, and it only cost a couple hundred dollars more than the other machines I was looking at. So before you roll your eyes at me saying local LLM inference is “free”, know that I’m not wielding a multi-GPU flamethrower that costs as much as a used car. I’m using something that’s a slightly premium regular desktop and it happens to be good at AI. It made up for the price delta long ago, so I’m calling local LLM inference “free”. You can gripe if you want and I won’t hold it against you. At least not too much. When Ornith can’t handle something, I have to reach for a bigger and slower local model (annoying my impatient side), or I need a faster and more capable Cloud model (annoying my cheapskate side). **What I need is a very cheap Cloud model that is noticeably better than Ornith 1.5 to get me over the top when the local model taps out.** I find the Artificial Analysis Intelligence Index to be the fastest way to compare models without too much hype. They make a composite score across a bunch of benchmarks, so a few points can mean you’ll notice the difference in capabilities. Here are the stats showing Claude Haiku 5.5, Claude Haiku 4.5, ChatGPT Luna 6, GLM-5.2, GLM-5.3, and Qwen 3.6 35B-A3B. I actually use Ornith 1.5, but that’s not in the index. It comes in slightly better than Qwen 3.6, but this is a useful baseline for my local model capability. *Source:* [*https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index?models=glm-5-3%2Cclaude-haiku-5-5-medium%2Cglm-5-2%2Cgpt-6-luna-medium%2Cqwen3-6-35b-a3b%2Cclaude-4-5-haiku-reasoning*](https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index?models=glm-5-3%2Cclaude-haiku-5-5-medium%2Cglm-5-2%2Cgpt-6-luna-medium%2Cqwen3-6-35b-a3b%2Cclaude-4-5-haiku-reasoning) **Quick summary:** **Intelligence Index Score** **Intelligence Index vs Cost per Task** **Token Usage** **Cost** **An important note on Haiku 5.5 pricing**: There are two WILDLY different pricing regimes for Claude Haiku 5.5. Under 100K of context tokens, Haiku 5.5 is priced identically to ChatGPT Luna 6. They’re both $0.10 for 1M input tokens and $0.50 for 1M output tokens, and that’s an absolute steal. Above 100K of context tokens, Haiku 5.5’s price jumps 5x to $0.50 for input and $2.50 for output. That makes Haiku 5.5’s value proposition go from “amazing” to “good”. ChatGPT Luna 6 also has a price escalator, but it’s at 272K of context and it only doubles the price instead of Haiku’s 5x. I break my model comparison into two broad categories: agentic tasks and coding. For **agentic tasks**, I have a stable harness of Open WebUI with a custom-developed agent. It has access to Tavily for web search, Defuddle and Playwright for scraping and web navigation, Open Terminal for local files and bash scripts, and Open WebUI’s built-in tools for notes, image generation, tasks, time, etc. As such, the agentic comparison is much more apples-to-apples because everything other than the model remains the same. I test agentic capabilities by asking the model to conduct research about how AI models have evolved over the past 12 months, and then it writes a report with some instructions. I monitor how thorough the research is, how efficient it is (getting the right information with 8 searches is better than the same info over 20 searches), how quickly it completes the task, how insightful the report is, and how succinct the report is. My **coding comparisons** are admittedly less directly comparable, because I am often testing with different coding harnesses. This can make a HUGE difference. GLM-5.2 inside of [Little Coder](https://github.com/itayinbarr/little-coder) kept running out of thinking budget and got stuck forever. GLM-5.2 inside of OpenCode is a very capable combination. So I will note which coding harness I used for the testing. Every harness has [Ponytail](https://github.com/DietrichGebert/ponytail), [Caveman](https://github.com/JuliusBrussee/caveman), [Defuddle MCP](https://github.com/codysandahl/defuddle-mcp), [Playwright MCP](https://github.com/microsoft/playwright-mcp), and the harness’ built-in tools. And I test models at Medium effort or thinking level wherever the model allows me to specify this. Why not Max or Extreme or Super Awesome effort level? Because I am impatient, dear reader. If you don’t mind waiting for your model to write a novella about its internal anxiety trying to satisfy your conflicting requirements, be my guest. I save the higher effort levels for when I’m running things in the background or overnight. My primary coding benchmark is to use the intent-driven development files for my local AI image generator and create a finished implementation. [Intent-driven development](https://medium.com/towards-artificial-intelligence/kill-the-ceremony-intent-driven-development-and-the-minimum-viable-spec-f8c65c5b5b8d) gives high level goals and constraints and then lets AI fill in the details. Every model starts from the same point with an intent, a wireframe, and a few notes. You can see the [baseline and several implementations here](https://github.com/codysandahl/ai-coding-harness-comparison). This coding benchmark contains a few intentional poison pills that the model has to overcome. The baseline intent file has some purposeful inaccuracies and contradictions, and that allows me to quickly measure a model’s debugging capability. The best result is for the model to notice it before implementing, ask me about it, and recommend fixing the intent before implementation. Very few models have achieved top marks on this. Beyond those intentional issues with the intent, I want the coding model to complete the task quickly, effectively, and with minimal interaction with me. **For this article, I tested Claude Haiku 5.5 inside Claude Code’s harness. I didn’t notice any regressions when I tested in OpenCode, but all coding observations in this article assume Claude Code.** Haiku 5.5 garnered top marks in four of my agentic categories, and got second place on the other two. It didn’t exhibit any noteworthy weaknesses. By contrast, Luna 6 tends to over-research. GLM-5.2 tends to over-research and over-think. Ornith 1.5 is just a step down in quality across the board, though it is still very good. **Verdict: Haiku 5.5 is the clear agentic winner.** Haiku 5.5 handled the intent spec and my poison pills beautifully, and while it didn’t win every category it had no clear weaknesses in the main test. In more complicated tasks beyond my benchmark, GLM-5.2 was noticeably better at creating an in-depth (albeit very verbose) and accurate plan. ChatGPT Luna 6 was a small step behind Haiku 5.5 on most of my metrics for the same price. But past 100K of context, Haiku costs 5x what Luna does and that’s going to happen frequently in coding scenarios. GLM required a bit more human feedback to reach a good solution, and Ornith was on the verge of annoying me by how much it needed my guidance. Luna and Haiku barely needed me after the initial prompt. For the intent spec, Luna and GLM fixed all the poison pills but never updated the intent spec, while Haiku and Ornith remembered to fix the underlying spec file. As a result of this mixed bag, I’m going to highlight when you would want to choose each model. Look at the “I Dub Thee” row to see when you should use each model. An important note: I ran my initial benchmark with GLM-5.2, but I have since switched to GLM-5.3 and it’s a major upgrade. Haiku 5.5 can kinda keep up with GLM-5.2, but it’s clearly not in GLM-5.3’s category for coding and planning. **Verdict: Per dollar, Haiku 5.5 is the coding winner IF you stay under the 100K context barrier. Past that, it costs 5x what Luna does and each model earns a niche.** I bought a nice computer for local LLM inference. I spent months figuring it all out and setting up the software — all in the pursuit of a fully-local AI workflow. I’ve invested blood, sweat, tears, and money into this local LLM concept. Granted, the blood was only a paper cut when I was unboxing the computer, but still. ChatGPT Luna 6 and Claude Haiku 5.5 have presented me with a dilemma. They are both clearly better than the best local model I can run at decent speeds, and they’re so cheap that my cheapskate side is waffling. Is it better to accomplish my task over an hour working with Ornith 1.5 for free, or over 15 minutes with Luna or Haiku for less than a dollar? My internal voices of achievement, cheapskateness, and impatience are at war with each other over this. So while it hurts my pride and decreases my ability to smugly mock my friends spending money on Cloud APIs, I have found myself leveraging Luna 6 and Haiku 5.5 more frequently at the expense of my local LLMs. But I haven’t completely abandoned the local LLM cause! Any task that can be run in the background and churn through a bunch of input tokens goes immediately to Ornith 1.5. For example, I needed to review a private Git repo to ensure I didn’t accidentally check in anything sensitive before I made it public. I tossed that to Ornith 1.5 inside of OpenCode, told it to spin up sub-agents for each root folder, and then I did other things while it checked. 28 minutes later, I had a zero-cost clean bill of health across a TON of input tokens. That would have cost several dollars even with Luna 6 or Haiku 5.5, and it didn’t matter that Ornith took five times as long to do the work because it wasn’t an interactive session. I do the same thing with some large research tasks. If I need to comb through hundreds of web pages and I don’t mind waiting for the report, Ornith 1.5 knocks them out one-by-one, takes notes, and delivers an excellent summary for free. If I need a quick answer across 10–20 web pages interactively, I find myself reaching for the speed of the Cloud models. I also still hand anything that touches API keys and secrets to Ornith 1.5, because then I’m guaranteed to not be leaking private info. This isn’t a high frequency use case for me, but when it arises I have my trust local LLM at the ready. Claude Haiku 5.5 has raised the bar for low cost Cloud LLMs — especially for any task that stays under the 100K context point. I’m still testing it, but I expect Haiku 5.5 to take over from Luna 6 as my main agentic Cloud model in Open WebUI, because most of those conversations stay under 100K of context. For coding, Luna 6 is likely to be the first car I pull out of the parking lot because of the much lower cost on coding tasks. Haiku 5.5 is a clear step up, but I also have GLM-5.3 sitting there with MUCH higher capability if I need a good plan that Luna 6 or Ornith 1.5 can implement. The highly capable and very cheap Cloud models are chipping away at my commitment to using local LLMs for everything, but I’m still glad I *can* do real work locally and I’m glad I can throw big token, slow-running tasks at my local LLM for free. Just promise me you won’t tell my friends that I’m reaching for Cloud models more than my local LLMs these days. They might give me some comeuppance if they find out. [Claude Haiku 5.5 Earned a Spot in My Real-World Workflows](https://pub.towardsai.net/claude-haiku-5-5-earned-a-spot-in-my-real-world-workflows-b7461598640f) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.