# Opus 5.5 is not more expensive than Qwen4 Max

> Source: <https://hominis.eu/blog/effort-levels/>
> Published: 2026-11-18 01:03:16+00:00

**Recommendation:** Compare models across multiple reasoning levels rather than one shared label, since a pricier model at Medium can cost less per task than a cheaper model at Max.
More in the playbook "Controlling LLM cost": <https://wpnews.pro/playbooks/llm-cost>

Today, Qwen4 Max was publicly announced and subsequently got listed on Artificial Analysis. A straight forward read of the graphs shows Alibabas newest 2.4T model scoring **42.9** at **$3.27** per task, quite a lot cheaper when compared with, to pick a random example, Claude Opus 5.5 at **$10.04**.

Would be nice to have a third the cost, but that is not the case. Those cost numbers are on **Max** reasoning. Note here, this is independent of the “Max” which Alibaba uses as part of the name for Qwen4 Max, which I am certain does not cause added confusion on top. Regardless, most models, those from Alibaba and Anthropic included, don’t have one reasoning level, but multiple, generally four to six:

“Is model A more expensive than B?” has no clear yes or no answer as it depends on the reasoning level chosen. As an example, Opus 5.5 at High falls within the ballpark of Qwen4 Max on average per-task-pricing yes scores nine points higher. Meanwhile, at Medium Opus 5.5 is both cheaper and still scores higher. Problem, if we only every compare the Max reasoning level, that gets easily overlooked.

## Effort levels are just labels

Effort levels are just marketing, a series of labels that only within a single model can be relatively consistent. Other models, even of the same lab, pre-train or “generation” (more on that below) can yield widely different results for the same reasoning level, including in relative improvement. Just because Model X 5.1 did slightly better on High vs Medium, doesn’t mean automatically that Model X 5.2 couldn’t do better on Medium vs High. We have seen this in the past, Opus 4.7 tended to use a higher average budget on the same reasoning level vs Opus 4.6 yielding to better average success but a slight regression in cost/performance on higher effort levels. A more recent example is the Gemini 4 series, those models play tightrope with their effort levels, improving and regressing in generally hard to predict ways.

High reasoning on one model can be cheaper, faster and/or more token efficient than Low or Medium on another. Even within labs and generations of models, even more so as version numbering and naming is, at this stage, mainly market positioning. If OpenAI tomorrow decided based on data to call what is speculated to be an upcoming release of Astra “GPT-7 Astra Pro MaxS”, they can and will do that, but it doesn’t mean that said model is a wholly new pre-train or that a hypothetical GPT-6.2 Sol couldn’t be based on a newer, more advanced training run that just so happened to also yield better efficient results. This also partly answers the eternal question of “Why would anyone use Fable 5.1 if Opus 5.5 is better and cheaper…”, though let’s stay on topic.

Effort levels are a problem in this regard, as they are often employed/tortured in an attempt to “normalise” comparisons between models. Have each model at Max, High or another level of the same label between labs and you have a fair comparison, that is the idea. Unfortunately, you tend to get the opposite. Grossly overgeneralising, a more advanced model can be more token efficient, completing the same task with a similar or higher success rate in similar or shorter time vs a less advanced model, all while full end-to-end cost is also lower.

If you normalise for such an arbitrary label, any nuance gets lost. A more capable and often per million tokens more expensive LLM on the same reasoning level may use more tokens than required to complete the task, thus spending what isn’t actually needed, thus seeming more expensive than it really is.

## The AA Default and its issue

[Artificial Analysis](https://artificialanalysis.ai/) is a great first stop to get a feel for model performance across many evals and viewpoints (token efficiency, throughput and latency, cost). Unfortunately their website by default only shows the highest effort level results they have available for any model. This leads to confusion, as on Max reasoning e.g. Qwen4 Max costs $3.27,- on average per task at a score of 43, whereas Opus 5.5 on Max costs $10.04,- on average per task at a score of 59. Again, I ask why you’d name your model after one of many effort levels it can be used at, but that’s not the main point right now.

Considering the score, the problem becomes quickly apparent. If we normalise by an actual user facing metric such as cost, Opus 5.5 on Medium reasoning costs only $2.45,- on average per task at a score of 48, getting both a better score at a lower average cost than Qwen4 Max. Similar is the case when looking at token efficiency, time to task completion, etc.

## Use Cost instead

In reporting and discussions of these models, we must get away from comparing effort levels between models, as this simply does not work. Instead, normalise by something measurable and comparable like cost or duration.

This graph provided get models whose average per-task-cost lands within a set range, sorted by score. By default each model appears once, at the best scoring effort setting in range. Note that even if you have an infinite budget, Max on many models tends to have diminishing returns and in specific evals, might even lead to regressions. Qwen4 Max is marked as a reference point.

Compare based on Cost, regardless of the reasoning label

## There is a better way (and why we won't get that one)

None of this would need to be a problem if we just learned from Nissan. Yes, the car brand. Their engine codes are frankly ingenious. As an example, let's take the venerable "VR38DETT". Might sound like gibberish, but:

- VR = Series
- 38 = Displacement (3.8l)
- D = Double Overhead Cams
- E = Electronic Fuel Injection
- TT = Twin Turbo

You see how easy this is to follow? If I now give you another engine, e.g. the "RB26DETT", you will instantly understand that it's a 2.6l engine, also with double cams, EFI and twin turbos.

For models, at least in the open-weight arena, we have something similar and I do like it. Easy to parse naming, nothing opaque, provides the needed context to get some high level understanding. I'd really like if all, including closed weight models followed such a naming (PreTrainGen+Param+PostTrainDate) but know that is never going to happen, because to anyone besides myself, it reads like a visual car crash and marketing doesn't exactly like that for some reason.

So how about this as a compromise: Labs should be more willing to communicate what their reasoning levels do and how they may differ between their models and between models by other labs. This should ideally go beyond the limited effort vs score graphs we have been getting since give or take the Opus 4.5 release post.

Labs should ideally state exactly how much each reasoning level raises the minimum and maximum tokens a model may budget for reasoning, whether this is consistent with other models of the same family and/or generation, etc.

## There is an even better way (and why we will get that one even if you might not like it)

Reasoning levels are poor UX. Heck, model selectors are poor UX. There, I said it, feel free to lob tomatoes in my general direction at your earliest convenience. But it's true, these opaque, inconsistent, constantly changing, sometimes even harness-dependent labels need to go. I know a person who, despite advice against it, runs regular back-and-forth chat queries with Opus 5.5 on Max Reasoning. Meanwhile, others may take a more measured approach in choosing their reasoning level but during long-running, multi-hour coding tasks, their model may encounter a single step that might benefit from a higher reasoning level, whereas the rest does not. You can try to gauge it, be clever and have the model use a higher-reasoning subagent if you can predict such a step before starting, but in any situation, why should you? Model choice within an offering shouldn't require a manual decision, same for reasoning. Note, this distinctly only applies for models within an organisation, switching on the fly between different closed-weight labs would just mean the loss of reasoning and other helpful but generally obfuscated information.

Now, I know for a fact a large contingent of users will not want this to happen. The loss of control will feel like just another lever for labs to introduce regressions, stealth nerfs and the like. And lest we forget, OpenAI tried and (due to rollout issues more than the concept) failed with a model router concept not unlike what I have described here during the early GPT-5 days, so I do hope that, like features such as auto-compaction, the context window, etc., models and efforts will remain something users can choose if they so desire. But if a major lab implements such an offering, I will test it. And if it does as well as or better than me manually selecting what models to use in which subagents on what specific reasoning level when, I will happily use it.

## Sources

All numbers come from the [Artificial Analysis Data API](https://artificialanalysis.ai/data-api), fetched on .

Note, in my experience, while the Intelligence Index can be a decent approximation of model performance, it doesn’t fully reflect my own experience or measurements. In my private evals, certain models deviate from their results, under- or overperforming the Intelligence Index results, with certain labs clearly teaching to the test, whether intentionally or on accident. That however does not affect the take-aways on this page concerning reasoning levels and relative performance.
