--- title: 'Before "thinking models" — the dark ages (2022–2024)' description: From ChatGPT's launch in November 2022 to September 2024, chat models answered in one pass, with no step to check their work. How thinking models arrived, what they fixed and what they didn't, and what thinking costs in my own apps. date: 2026-11-04 tags: AI, LLMs, reasoning models, hallucinations, AI costs --- Before thinking models, from ChatGPT's launch in November 2022 to September 2024, a chat model answered in one pass. It wrote its reply one word at a time, with no step to check its work. Thinking models, starting with OpenAI's o1, write out their reasoning first and answer after. I wasn't sure of the dates myself. This week, I asked Nova, one of my agents, on Telegram. I remembered that the first ChatGPT and Claude didn't think, and that today they all do. I guessed the change came about a year ago. It came two years ago. Nova sent back a timeline and called the years before it "the dark ages". I liked the title. So I checked the timeline for this note, and added what thinking changed in my own apps: what they cost, how fast they answer, and sometimes whether they answer at all. ## Two years of answers in one pass "The dark ages" is a joke, but the mechanism behind it is real. Until September 2024, a chat model had no place to think before it spoke. A language model writes one word at a time. Strictly, it writes tokens, pieces of words, but words are close enough. For each one, it picks a likely next word given everything before it, then moves on. It can't go back and change a word it already wrote. So the answer was the thinking. If the model took a wrong step early in a sum, every word after it built on that step. The reply still read smoothly. A wrong answer looked as sure as a right one. Researchers found a workaround in 2022. A team from the University of Tokyo and Google added one sentence before the answer: "Let's think step by step." On a set of simple arithmetic word problems, an OpenAI model went from 18% right to 79%. That's from fewer than one problem in five to almost four in five. So the model could reason, as long as it wrote the steps down. But someone had to ask for it, and nobody had trained the model to do it well.

ChatGPT, November 2022

One pass

Writes the answer word by word

Each word is a guess from the words before

No step back to check

then

o1, September 2024

Think, then answer

Writes its reasoning first

Can catch a mistake before you see it

More words, more time, more cost

2022 trick"Let's think step by step": from 18% to 79% right on simple word problems.

The word chips are a schematic. The two scores are from Kojima and others, 2022, with an OpenAI model of the time.

## September 2024: the model learns to think first On September 12, 2024, OpenAI released o1-preview. It was the first widely used model trained to write its reasoning before every answer. The training method is reinforcement learning. The model tries a problem, writes its reasoning, and gets a reward when the final answer is right. Over many tries, it keeps the habits that pay. OpenAI's post names some of them: it "learns to recognize and correct its mistakes", breaks hard steps into simpler ones, and tries another way when one fails. OpenAI kept the raw reasoning hidden. In ChatGPT, you saw a short summary of it, and a line saying how long the model had thought. The gain was largest where an answer can be checked. AIME is a 15-question exam that qualifies US high school students for the math olympiad. On the 2024 exams, GPT-4o, the best ChatGPT model before o1, solved about 2 questions out of 15 on average. o1 solved about 11. OpenAI also found that o1 got better the longer it thought. That's a new trade: you buy a better answer with more time and more tokens. It's the trade I now meet in my own apps.

GPT-4oAbout 2 of 15

o1, one tryAbout 11 of 15

64 triesWith a vote on the most common answer, o1 reached 12.5 of 15.

OpenAI's averages on the 2024 exams, 1.8 and 11.1 of 15, rounded to whole squares.

## Six months later, everyone had one The full o1 came out on December 5, 2024. Two weeks later, Google released an experimental thinking version of Gemini 2.0 Flash. Then everyone followed. - **DeepSeek-R1, January 2025.** A Chinese lab released a reasoning model close to o1 on math and code. Its weights were open, under the MIT license, so anyone could download and run it. A week later, Nvidia's share price fell 17% in one day. - **Claude 3.7 Sonnet, February 2025.** Anthropic called it a "hybrid reasoning model": the same model answers fast, or thinks first when you ask, and you can read the thinking. - **o3 and o4-mini, April 2025.** OpenAI's next reasoning models could use tools while they thought: search the web, run code, look at images. - **GPT-5, August 2025.** In ChatGPT, a router decides, message by message, whether to answer fast or think. Today, thinking is a setting, not a separate product. Even the smallest Claude, Haiku 5.5, thinks unless you tell it not to. Nova's dates held up when I checked them. One number didn't. Nova said DeepSeek trained R1 for about $6 million. That figure belongs to another model: about $5.6 million was the final training run of DeepSeek's base model, V3. R1 is built on top of V3, and its reasoning training cost $294,000 more, according to DeepSeek's paper in Nature. Nova runs on Claude Sonnet 5, a thinking model, and it still mixed the two up. So what does thinking really fix?

Nov 2022
ChatGPT, then ClaudeOne pass, no thinking step. Claude arrives in March 2023.
Sep 12, 2024
o1-previewTrained to think before it answers.
Dec 2024
o1, Gemini 2.0 Flash ThinkingOpenAI and Google, two weeks apart.
Jan 2025
DeepSeek-R1Open weights, MIT license. Nvidia falls 17% in a day.
Feb 2025
Claude 3.7 SonnetOne model: answer fast, or think first.
Apr 2025
o3 and o4-miniThink while using tools: search, code, images.
Aug 2025
GPT-5A router decides when ChatGPT thinks.
Today
Thinking by defaultEven Claude Haiku 5.5 thinks unless told not to.

Dates from each company's announcement. The grey row is the "dark ages"; the last row is from my own apps.

## Thinking checks the steps, not the facts Thinking helps most when each step can be checked: a sum, a proof, code that tests can run. It doesn't give the model a fact it never learned. OpenAI's own reports show both sides. For the o1 report, in December 2024, a monitor read the reasoning behind 102,443 answers. In 43 of them, o1 made up information, and its reasoning showed it knew. Most were references: a book or a web page it couldn't check without search, so it wrote a likely-looking one. That's rare, about 4 answers in 10,000. But thinking didn't stop it. It only made it visible to whoever read the reasoning. The o3 report, in April 2025, went further. On PersonQA, OpenAI's test of questions about public facts on people, o3 made things up in 33% of its answers. That's twice o1's 16%, and o4-mini reached 48%. OpenAI's explanation is that o3 makes more claims overall, right and wrong. It also wrote that "more research is needed". In September 2025, OpenAI published a paper on why models make things up at all. Their answer: most tests score a guess above "I don't know", so training teaches models to guess. Thinking doesn't change that score. I hit the same limit at home. In August, my fitness agent, Rex, had to log a food product from my photos, nutrition table included. I tried five models on the same step. Gemini 3 Flash came back with a fish dish from earlier in the chat, not the product in the photo. Gemini 3.1 Pro said it couldn't read the values and asked me for them. The problem was my photos: none showed the table clearly. No model, thinking or not, reads numbers that aren't in the picture. I needed a better photo, not a better model.

Thinking helps

Steps you can check

A math exam: from about 2 to 11 of 15

Code, when tests run it

A plan with tools, step by step

but

Thinking can't help

Facts it doesn't have

A reference it can't open

A fact about a person it never learned

A number that isn't in the photo

PersonQA: answers with made-up facts

o1, December 202416%

o3, April 202533%

o4-mini, April 202548%

o1 report43 of 102,443 answers made up knowingly, mostly references.

Bars to scale on a 0 to 100% axis, from OpenAI's o3 and o4-mini system card. Red marks a rate above o1's.

## Thinking has a bill Thinking is extra text the model writes before the answer. You pay for it, you wait for it, and it counts against the answer's size limit. I've met all three in my own apps. - **The cost.** In September, my meeting notes app, Scribelex, sent a 16-minute meeting to Gemini 3.5 Flash. About 80% of what the model wrote was hidden thinking, billed like the answer. The meeting cost €0.39. My estimate had said €0.08, because it didn't count the thinking. - **The empty answer.** In October, the live chat in the same app showed a blank reply. Claude Haiku 5.5 thinks by default, and its thinking shares the answer's limit. With a limit of 700 tokens, one call spent all 700 thinking. It returned no text and no error. With thinking off, answers now come back in about 2 seconds instead of 6. - **The local model.** My commit tool, brb-git, can run on the Mac mini's own models. There it uses IBM's Granite, a model that doesn't think. Thinking models came back empty: they spent the budget on reasoning, the library dropped the reasoning, and nothing was left. So I choose per task. For a commit message or a quick answer during a meeting, thinking stays off. For agents, it's the opposite. When I set up my agent framework, OpenClaw, it skipped any local model that wasn't marked as able to reason. Agents plan, call tools, read the results and fix their own mistakes. That's the step-by-step work thinking is good at.

AppWhat thinking didWhat I changed

Meeting notes80% of the output was hidden thinking: €0.39 against an €0.08 estimateThe cost now counts it. Turning it down: undecided

Live chat700 of 700 tokens spent thinking: an empty answer, no errorThinking off: 2 seconds instead of 6. An empty answer is an error

Commit messagesLocal thinking models: empty textIBM's Granite 4.1, a model that doesn't think

AgentsPlan, call tools, fix their own mistakesOnly models marked as able to reason

Per taskShort answers: thinking off. Agents that plan: thinking on.

From my own apps, this year. Red marks what broke.

## Try it - OpenAI, [Learning to reason with LLMs](https://openai.com/index/learning-to-reason-with-llms/), September 2024: how o1 was trained, and the AIME results. - The [o1 system card](https://arxiv.org/abs/2412.16720), December 2024, with the reasoning monitor, and the [o3 and o4-mini system card](https://openai.com/index/o3-o4-mini-system-card/), April 2025, with PersonQA. - Kojima and others, [Large Language Models are Zero-Shot Reasoners](https://arxiv.org/abs/2205.11916), 2022: the "Let's think step by step" paper. - DeepSeek, [DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning](https://www.nature.com/articles/s41586-025-09422-z), Nature, September 2025. - OpenAI, [Why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/), September 2025. - Two of my notes on what AI costs: [counting what tokens buy](/notes/tokenomics-dont-count-tokens.html) and [spending more tokens](/notes/tokenmaxxing-fat-skills-real-tests.html). Pick one prompt you send every day. Run it once with thinking off and once with it on, and compare the answer, the wait and the cost. If you find a task where thinking isn't worth it, tell me which one.