New AI Models Force Reasoning On - 7 of 18 Won’t Let You Disable

👁 116 reads

When working with models through the API, you could until now disable reasoning with a single parameter. The model would stop thinking out loud and answer immediately, which saved time and money on simple tasks. We tested eighteen models today and found something unexpected.

Seven of them reject such attempts with an error. All seven are latest-generation models. All eleven that still allow it are older models.

Which Models Block the Choice

The message is the same every time and leaves no room for discussion:

Reasoning is mandatory for this endpoint and cannot be disabled.

The division is surprisingly clean. It’s not about the vendor or price, because both sides include models from OpenAI, Anthropic, Google and Z.ai. It’s about generation. GLM 5.2 allows it, GLM 5.3 doesn’t. Claude Sonnet 4.6 allows it, Claude Fable 5.1 doesn’t. Same company, same parameter, different answer.

Why This Costs Money, Not Just Annoyance

Reasoning tokens are billed like regular output tokens, meaning at the more expensive side of the price list. You pay for text you never see.

We measured how much there is. In a test of six short tasks (run in Polish, a heavily inflected language), reasoning’s share of total output was:

• GPT-6 Astra: 52 to 95 percent
• GLM 5.3 Flash: 68 to 96 percent
• Gemini 3.8 Flash when writing an article: 985 reasoning tokens
• Kimi K3 on the same article: 4,667 reasoning tokens

On a task where we asked for two sentences, one model produced 659 output tokens, of which 625 were thinking. We got two sentences and a bill for twenty times more text.

We described the most extreme case separately: when asked for an article, GLM 5.3 Flash used its entire eight-thousand-token limit on reasoning and returned zero words, while still billing for it.

What This Means for Listed Prices

If reasoning accounts for an average of eighty-five percent of output tokens, and you can’t disable it, then the price shown on the model’s card describes something different from what you actually pay for. For every token you see, you pay roughly as much as for six you don’t see.

This is the second multiplier after the one we described earlier: the same model is hosted by a dozen companies with their own rates, and the spread reaches fourteen times. Both work in the same direction and both are invisible until you start reading the cost from the bill instead of estimating it from the price list.

Is This Bad

Not for every use case. On difficult tasks, reasoning genuinely improves results and paying for it is justified. The problem is that on trivial tasks you pay exactly the same, and you have no choice.

The vendor has its logic here: a model with reasoning disabled performs worse in tests and damages the product impression. We understand that. It doesn’t change the fact, however, that when transcribing text from one format to another, nobody needs the model to first consider three approaches.

What to Do About It in Practice

For simple and high-volume tasks, choose previous-generation models. For transcription, summarization and data organization, DeepSeek V4 Flash with reasoning disabled does the same thing for a fraction of the cost. We measured that it wrote a nine-hundred-word article for $0.00039, without a single reasoning token.

Handle the 400 error in your code. If your program sends this parameter hardcoded, it will stop working entirely after switching models. That’s exactly what happened to us: our news generator sent reasoning disablement with every request, and after switching to a newer model it would have returned an error instead of text. We added detection for this error and one retry without the parameter.

Log the cost from the response, not the price list. Just ask the interface for billing, and with each response you’ll get the actual amount and number of reasoning tokens. Without this, you won’t notice that half the bill goes to text that’s not on the screen.

With models that have mandatory reasoning, set a higher token limit. Otherwise the model will burn through it on thinking and return an empty response with success status. This is a failure you won’t see in logs, because formally nothing broke.

How We Tested This

Eighteen models, September 10, 2026, one minimal query to each with the reasoning-disabling parameter and a limit of ten tokens to keep the test cheap. Only the response code and error message content matter, so the result is unambiguous and easy to replicate.

The list of models that block disablement will grow, because it follows new versions. We’ll update this article when more arrive.

We collected more measurements of the same models, but from the quality and billing side, in articles about which model writes best in non-English languages, how much an AI-written article costs, and whether the promised context window actually works.

// read moreGPT-6.1 Terra Reasoning Levels: Low, Medium, High or Max?BenchmarksDeepSeek: Free Chat, API Costs and Limits ExplainedGuidesAI Context Window Test: Two Models Failed at 25,000 TokensBenchmarks

Kinetyra