TechNews DevTop — Reasoning models are now a default setting across major AI platforms, but the label hides a commercial trade-off. Before answering, these systems generate internal reasoning tokens: an invisible draft that helps with code debugging, planning, math and long-document analysis. That draft is billed, occupies context and adds latency. The practical question is not whether reasoning is superior, but whether a task justifies paying for it.
- What changes: reasoning models produce hidden intermediate tokens before visible output, and providers bill them as output even when users cannot see them.
- Where costs bite: OpenAI says reasoning can range from hundreds to tens of thousands of tokens per request and recommends reserving at least 25,000 tokens for reasoning plus response in early tests.
- When to use it: multi-step debugging, competition math, planning and ambiguous research benefit; simple retrieval, classification, voice and routine drafting usually do not.
What the thinking layer actually does
OpenAI describes reasoning tokens as a planning stage that helps models use tools effectively, examine alternatives, recover from ambiguity and solve multi-step tasks. Anthropic says Claude reformulates a request, tests approaches, checks intermediate results and abandons weak paths. Google says Gemini models adjust effort dynamically according to complexity.
Visible thought summaries are not raw cognition. Anthropic describes thought blocks as summaries, never the unprocessed reasoning chain. OpenAI does not expose reasoning tokens through its API. Google returns only the final answer by default; thought summaries must be enabled. Anthropic also says Claude can think between one tool call and the next, which matters for agentic workflows.
The bill is measured in output tokens and time
All three providers charge reasoning as output tokens, even when the user never sees them. OpenAI says reasoning tokens consume context-window space and are billed as output. Anthropic says customers pay for the full thinking process, not the summary shown in the interface. Google calculates response price as the sum of output tokens and thinking tokens.
Scale is the issue. OpenAI says a model may generate from a few hundred to tens of thousands of reasoning tokens per request. In an example in its own documentation, 1,024 of 1,186 output tokens—about 86%—were reasoning. The company recommends reserving at least 25,000 tokens for reasoning and response when experimenting.
There is also a failure mode more expensive than ordinary overthinking: running out of budget midway. If the token limit is reached, OpenAI warns that users can pay for input and reasoning without receiving any visible answer. Google describes a similar outcome: output may be truncated or empty while generated thinking tokens are still billed. Its guidance is to lower the thinking level rather than cut the maximum output. Latency is the second invoice. OpenAI calls its low level a modest latency increase and says xhigh should be used only when evaluations show a clear benefit that justifies the extra latency and cost.
Matching effort to the task
The providers’ guidance converges on a simple split. Direct answers need little or no reasoning. Drafting, data analysis and customer support usually need low effort. Comparing concepts, research and spreadsheet work fit medium effort, the default on many models. Complex code debugging, competition mathematics and multi-step planning justify high or maximum effort.
OpenAI frames reasoning models as planners and faster models as workhorses. The planners excel at ambiguous tasks and needle-in-a-haystack searches across contracts and financial statements; the workhorses handle well-defined jobs where speed and cost matter. OpenAI also warns that telling a reasoning model to think step by step can fail to improve results or even make them worse. Direct instructions with a clear objective work better.
As of October 2026, the model landscape is blurring the category. OpenAI recommends gpt-6-astra for most reasoning workloads, gpt-5.6-terra for lower cost and gpt-5.6-luna for the lowest cost and latency. Effort levels run from none to max depending on the model; gpt-6-astra does not support none. Anthropic says reasoning is already enabled on Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, among others. On Opus 5.5 adaptive thinking cannot be disabled and defaults to medium effort; the model decides per request whether and how much to think. Google lists gemini-4-argon and gemini-3.8-flash with thinking on by default at medium, plus low, medium and high levels. Some lighter models, such as gemini-2.5-flash-lite, ship with it off.
The direction of travel is clear: reasoning is becoming an effort dial rather than a separate product tier. That shifts responsibility to buyers. Defaults are commercial choices, and Anthropic’s move to medium effort on Opus 5.5 where earlier Opus models used high effort shows how quietly the setting can change. Teams should start low or at the provider default, measure success on multi-step tasks and raise effort only when a specific evaluation shows a better answer. A model that thinks longer does not gain new knowledge; it reasons more thoroughly over what it already has. If the improvement cannot be demonstrated, the extra tokens are payment for waiting.