# MIT Technology Review reports Google’s Gemini reasoning dial targets overthinking [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-reports-google-s-gemini-reasoning-dial-94106996) MIT Technology Review’s report, published April 17, 2025, says Google DeepMind added a dial to Gemini Flash 2.5 so developers can limit how much computing power it uses to reason through a problem. The report presents this as a way to address a costly mismatch: reasoning models can spend more time than simple prompts need, increasing developers’ costs and AI’s environmental footprint. DeepMind’s Jack Rae says it is hard to know which tasks need more reasoning; Hugging Face engineer Nathan Habib says companies use these models even where they are not needed. The report says outputs cost about six times more to generate when reasoning is on, while citing coding and expert-level research reports as tasks where developers might accept the expense. It also notes that reasoning can go awry: Habib describes a model repeating itself during an organic chemistry problem, and DeepMind’s Kate Olszewska says Google models can get stuck in loops. The dial is for developers building apps, not the consumer version of Gemini. More testing and developer feedback are needed to learn when medium or low settings are sufficient. The report closes with DeepMind CTO Koray Kavukcuoglu’s view that reasoning will underpin future AI models that act on people’s behalf and solve problems for them. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Apr 17, 2025 (2025-04-17T19:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-14T13:33:26.405Z. First observed: 2026-10-29T19:00:18.246Z. Retrieved: 2026-10-29T19:01:52.014Z. - [A Google Gemini model now has a “dial” to adjust how much it reasons](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Limits Text as retrieved on October 29, 2026; originally published April 17, 2025. The report does not give a measured estimate of the environmental footprint attributable to the reasoning dial. ## Open question What testing and developer feedback will show when medium or low reasoning settings are sufficient for a task? ## Source-stated Google DeepMind added a dial to Gemini Flash 2.5 that lets developers control how much the model reasons. > Gemini Flash 2.5, the model released today that includes a slider for developers to dial back how much it thinks. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated The report says reasoning can make simple prompts more expensive to run and worsen AI’s environmental footprint. > When a model spends longer than necessary on a problem, it makes the model expensive to run for developers and worsens AI’s environmental footprint. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated The report says outputs are about six times more expensive to generate when reasoning is turned on. > Outputs from the model are about six times more expensive to generate when reasoning is turned on. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated The report says more testing and feedback from developers are needed to find when medium or low settings are good enough. > But more testing and feedback from developers will be needed to find out when medium or low settings are good enough. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated Koray Kavukcuoglu says DeepMind views reasoning as a foundation for future AI models that act on people’s behalf and solve problems for them. > In DeepMind’s view, this reasoning will be the foundation of future AI models that act on your behalf and solve problems for you. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Reasoning model A language model trained to work through a problem in written steps before it answers. - [Definition reference](https://en.wikipedia.org/wiki/Reasoning_model) ## Google DeepMind Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold. - [Explore this context](/entities/google-deepmind) - [Reference](https://deepmind.google/blog/) - [Reference](https://deepmind.google/science/alphafold/) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Hugging Face Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems. - [Explore this context](/entities/hugging-face) - [Reference](https://huggingface.co/blog) ## Editorial provenance Version 1. Publication ID: 83e39d94-6a04-f616-6667-218fa219b749. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-14T13:28:47.425Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: bc6a0a34121cb9428789bbc266de0bebe5f216e9eac3a3a46a54fd17459a4bfd. Source text hash: ee0ea94823ff1b9bc6812fdf5e729c731aa986cc94776edd16083b041fa63f9f. ## Revision and source-check history ### Version 1 — 2026-11-14T13:33:26.405Z MIT Technology Review’s report, published April 17, 2025, says Google DeepMind added a dial to Gemini Flash 2.5 so developers can limit how much computing power it uses to reason through a problem. The report presents this as a way to address a costly mismatch: reasoning models can spend more time than simple prompts need, increasing developers’ costs and AI’s environmental footprint. DeepMind’s Jack Rae says it is hard to know which tasks need more reasoning; Hugging Face engineer Nathan Habib says companies use these models even where they are not needed. The report says outputs cost about six times more to generate when reasoning is on, while citing coding and expert-level research reports as tasks where developers might accept the expense. It also notes that reasoning can go awry: Habib describes a model repeating itself during an organic chemistry problem, and DeepMind’s Kate Olszewska says Google models can get stuck in loops. The dial is for developers building apps, not the consumer version of Gemini. More testing and developer feedback are needed to learn when medium or low settings are sufficient. The report closes with DeepMind CTO Koray Kavukcuoglu’s view that reasoning will underpin future AI models that act on people’s behalf and solve problems for them. ## Original source MIT Technology Review · Article Original source date: Apr 17, 2025 (2025-04-17T19:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-14T13:33:26.405Z. First observed: 2026-10-29T19:00:18.246Z. Retrieved: 2026-10-29T19:01:52.014Z. - [A Google Gemini model now has a “dial” to adjust how much it reasons](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Limits Text as retrieved on October 29, 2026; originally published April 17, 2025. The report does not give a measured estimate of the environmental footprint attributable to the reasoning dial. ## Open question What testing and developer feedback will show when medium or low reasoning settings are sufficient for a task? ## Source-stated Google DeepMind added a dial to Gemini Flash 2.5 that lets developers control how much the model reasons. > Gemini Flash 2.5, the model released today that includes a slider for developers to dial back how much it thinks. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated The report says reasoning can make simple prompts more expensive to run and worsen AI’s environmental footprint. > When a model spends longer than necessary on a problem, it makes the model expensive to run for developers and worsens AI’s environmental footprint. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated The report says outputs are about six times more expensive to generate when reasoning is turned on. > Outputs from the model are about six times more expensive to generate when reasoning is turned on. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated The report says more testing and feedback from developers are needed to find when medium or low settings are good enough. > But more testing and feedback from developers will be needed to find out when medium or low settings are good enough. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source-stated Koray Kavukcuoglu says DeepMind views reasoning as a foundation for future AI models that act on people’s behalf and solve problems for them. > In DeepMind’s view, this reasoning will be the foundation of future AI models that act on your behalf and solve problems for you. - [Supporting source](https://www.technologyreview.com/2025/04/17/1115375/a-google-gemini-model-now-has-a-dial-to-adjust-how-much-it-reasons/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Reasoning model A language model trained to work through a problem in written steps before it answers. - [Definition reference](https://en.wikipedia.org/wiki/Reasoning_model) ## Google DeepMind Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold. - [Explore this context](/entities/google-deepmind) - [Reference](https://deepmind.google/blog/) - [Reference](https://deepmind.google/science/alphafold/) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Hugging Face Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems. - [Explore this context](/entities/hugging-face) - [Reference](https://huggingface.co/blog) ## Editorial provenance Version 1. Publication ID: 83e39d94-6a04-f616-6667-218fa219b749. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-14T13:28:47.425Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: bc6a0a34121cb9428789bbc266de0bebe5f216e9eac3a3a46a54fd17459a4bfd. Source text hash: ee0ea94823ff1b9bc6812fdf5e729c731aa986cc94776edd16083b041fa63f9f. ### passed — 2026-11-14T13:28:47.425Z — anthropic/claude-opus-5.5 Publication: 83e39d94-6a04-f616-6667-218fa219b749 attribution — pass: The reading presents this as MIT Technology Review's report and attributes quoted views to the right people: Rae, Habib, Olszewska, Doshi's overthinking context and Kavukcuoglu. Linking the 'In DeepMind's view' sentence to Kavukcuoglu fits the paragraph, which centres on his remarks and is followed by his quote. The source is a single article, not a listing. > In DeepMind’s view, this reasoning will be the foundation of future AI models that act on your behalf and solve problems for you. summary — pass: The reading leads with the actor (Google DeepMind) and the change (a developer-facing dial on Gemini Flash 2.5 that limits reasoning compute). It explains why this matters through the cost, the environmental footprint and the overthinking problem. 'Reasoning model' is defined beside the reading. It keeps the source's uncertainty about which tasks need reasoning, along with Habib's critique. The question follows directly from the source's stated open issue. > Developers can set a budget for how much computing power the model should spend on a certain problem date — pass: April 17, 2025 matches the UTC day of originalPublished. The source was modified the same day, per its note that it was 'updated to clarify the problem of overthinking'. The reading's overthinking framing appears in the article body as captured, and nothing the reading uses comes from the later sidebar items, such as the AlphaGo ten-year story. The limitation discloses the later retrieval. No hindsight framing appears, and no forecast is presented as an outcome. > This story was updated to clarify the problem of "overthinking." support — pass: Each claim is directly stated in the source: the developer dial, the cost and environmental footprint from overthinking, the roughly sixfold output cost, the need for more testing, and the DeepMind view on future agents. Habib's critique and the looping examples are faithfully conveyed. No benefits are invented beyond what the source states, and the costs are framed as the report's assertions. > For simple prompts, the model does think more than it needs to. media — pass: No media was selected, and no verified media candidates were supplied, so nothing is misidentified or irrelevant. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/94106996c1cfd63765d82d2bb00fecbab46d49cf225d7ac26f04e6f50ab8b297/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-reports-google-s-gemini-reasoning-dial-94106996.json)