{"id":265666,"date":"2026-08-04T08:30:14","date_gmt":"2026-08-03T23:30:14","guid":{"rendered":"https:\/\/designcopy.net\/en\/?p=265666"},"modified":"2026-10-21T19:53:12","modified_gmt":"2026-10-21T10:53:12","slug":"chain-of-thought-prompting-gpt4o-claude-gemini-2026","status":"publish","type":"post","link":"https:\/\/designcopy.net\/en\/chain-of-thought-prompting-gpt4o-claude-gemini-2026\/","title":{"rendered":"Chain-of-Thought Prompting in 2026: What Works and When to Skip It"},"content":{"rendered":"<div style=\"background:#EFF6FF;border-left:4px solid #3B82F6;padding:20px 24px;margin:24px 0;border-radius:6px;\">\n<strong style=\"color:#1E40AF;font-size:1.05em;\">Quick Answer: Chain-of-Thought Prompting<\/strong><\/p>\n<ul style=\"margin:10px 0 0;padding-left:20px;color:#1E3A5F;line-height:1.8;\">\n<li>Chain-of-thought (CoT) prompting asks a model to work through intermediate steps before answering. Wei et al. (2022) showed it lifting large models on multi-step maths and reasoning; Kojima et al. (2022) showed a single line, &ldquo;Let&rsquo;s think step by step&rdquo;, could do much of the same.<\/li>\n<li><strong>It matters less on today&rsquo;s reasoning models.<\/strong> OpenAI&rsquo;s reasoning models, Claude with extended thinking and Gemini&rsquo;s thinking mode already reason before they answer. OpenAI says &ldquo;think step by step&rdquo; prompts may not help them and can hinder them.<\/li>\n<li>CoT still earns its place on standard (non-reasoning) models, and in the form of <em>structure<\/em>: telling the model which checks to make and what to output.<\/li>\n<li>The main variants are zero-shot CoT, few-shot CoT, self-consistency and tree of thoughts.<\/li>\n<\/ul>\n<\/div>\n<p style=\"background:#f8fafc;border:1px solid #e2e8f0;border-radius:6px;padding:12px 16px;font-size:0.95em;\"><strong>How we wrote this:<\/strong> the research and vendor guidance come from the original papers and the providers&rsquo; current documentation, checked on October 21, 2026. The four content templates are starting points to adapt; we haven&rsquo;t benchmarked them across models.<\/p>\n<figure style=\"margin:24px 0;text-align:center;\"><img decoding=\"async\" src=\"https:\/\/designcopy.net\/wp-content\/uploads\/2026\/07\/chain-of-thought-prompting-gpt4o-claude-gemini-2026-internal-1-hero.jpg\" alt=\"What chain-of-thought prompting is and why step-by-step reasoning helps\" style=\"max-width:100%;height:auto;border-radius:8px;\" loading=\"lazy\" title=\"\"><\/figure>\n<h2>What Is Chain-of-Thought Prompting?<\/h2>\n<p>Chain-of-thought prompting gets a language model to write out its intermediate reasoning before committing to a final answer, like a student showing their working.<\/p>\n<p>The idea was formalised by Google researchers in <a href=\"https:\/\/arxiv.org\/abs\/2201.11903\" target=\"_blank\" rel=\"noopener nofollow external noreferrer\" data-wpel-link=\"external\">&ldquo;Chain-of-Thought Prompting Elicits Reasoning in Large Language Models&rdquo;<\/a> (Wei et al., 2022). With eight worked examples in the prompt, a 540-billion-parameter model reached state-of-the-art accuracy on GSM8K, a benchmark of grade-school maths word problems. The same paper found the benefit appeared only in very large models of that era; small models didn&rsquo;t gain.<\/p>\n<p>A few months later, <a href=\"https:\/\/arxiv.org\/abs\/2205.11916\" target=\"_blank\" rel=\"noopener nofollow external noreferrer\" data-wpel-link=\"external\">&ldquo;Large Language Models are Zero-Shot Reasoners&rdquo;<\/a> (Kojima et al., 2022) showed you don&rsquo;t always need examples. Adding &ldquo;Let&rsquo;s think step by step&rdquo; took one model from 17.7% to 78.7% on the MultiArith benchmark and from 10.4% to 40.7% on GSM8K.<\/p>\n<p><strong>Why it works:<\/strong> a model generates each token based on everything before it. When it writes the reasoning first, the final answer is conditioned on those steps rather than produced in one leap, which cuts down errors on problems with several dependent parts.<\/p>\n<h2>The Main Variants<\/h2>\n<table style=\"width:100%;border-collapse:collapse;margin:20px 0;\">\n<thead>\n<tr style=\"background:#1e3a5f;color:#fff;\">\n<th style=\"padding:10px 14px;text-align:left;\">Technique<\/th>\n<th style=\"padding:10px 14px;text-align:left;\">How it works<\/th>\n<th style=\"padding:10px 14px;text-align:left;\">When to use it<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr style=\"background:#f8fafc;\">\n<td style=\"padding:10px 14px;\">Zero-shot CoT<\/td>\n<td style=\"padding:10px 14px;\">Add a trigger such as &ldquo;Let&rsquo;s think step by step&rdquo; (Kojima et al., 2022)<\/td>\n<td style=\"padding:10px 14px;\">Quick reasoning boost on a standard model<\/td>\n<\/tr>\n<tr>\n<td style=\"padding:10px 14px;\">Few-shot CoT<\/td>\n<td style=\"padding:10px 14px;\">Include worked examples that show the reasoning and the answer (Wei et al., 2022)<\/td>\n<td style=\"padding:10px 14px;\">Specialised tasks and fixed output formats<\/td>\n<\/tr>\n<tr style=\"background:#f8fafc;\">\n<td style=\"padding:10px 14px;\">Self-consistency<\/td>\n<td style=\"padding:10px 14px;\">Sample several reasoning paths and take the most common answer (Wang et al., 2022)<\/td>\n<td style=\"padding:10px 14px;\">When one wrong chain is costly and you can afford several calls<\/td>\n<\/tr>\n<tr>\n<td style=\"padding:10px 14px;\">Tree of thoughts<\/td>\n<td style=\"padding:10px 14px;\">Explore branching reasoning paths, evaluate them, and backtrack (Yao et al., 2023, Princeton and Google DeepMind)<\/td>\n<td style=\"padding:10px 14px;\">Planning and search problems with no single path<\/td>\n<\/tr>\n<tr style=\"background:#f8fafc;\">\n<td style=\"padding:10px 14px;\">Built-in reasoning<\/td>\n<td style=\"padding:10px 14px;\">The model reasons in a separate phase before answering (OpenAI reasoning models, Claude extended thinking, Gemini thinking)<\/td>\n<td style=\"padding:10px 14px;\">Hard multi-step tasks; replaces manual CoT on these models<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<figure style=\"margin:24px 0;text-align:center;\"><img decoding=\"async\" src=\"https:\/\/designcopy.net\/wp-content\/uploads\/2026\/07\/chain-of-thought-prompting-gpt4o-claude-gemini-2026-internal-2-hero.jpg\" alt=\"Chain-of-thought prompting variants from zero-shot to tree of thoughts\" style=\"max-width:100%;height:auto;border-radius:8px;\" loading=\"lazy\" title=\"\"><\/figure>\n<h2>Reasoning Models Changed the Advice<\/h2>\n<p>The biggest change since 2022: the leading providers now ship models that do chain-of-thought internally. That changes how you should prompt them.<\/p>\n<ul>\n<li><strong>OpenAI.<\/strong> OpenAI&rsquo;s <a href=\"https:\/\/developers.openai.com\/api\/docs\/guides\/reasoning-best-practices\" target=\"_blank\" rel=\"noopener nofollow external noreferrer\" data-wpel-link=\"external\">reasoning best practices<\/a> tell you to avoid chain-of-thought prompts with its reasoning models. They already reason internally, and asking them to think step by step &ldquo;may not enhance performance (and can sometimes hinder it)&rdquo;.<\/li>\n<li><strong>Anthropic.<\/strong> With extended thinking, Claude&rsquo;s <a href=\"https:\/\/platform.claude.com\/docs\/en\/build-with-claude\/prompt-engineering\/claude-prompting-best-practices\" target=\"_blank\" rel=\"noopener nofollow external noreferrer\" data-wpel-link=\"external\">prompting guide<\/a> recommends general instructions over prescriptive steps. A prompt like &ldquo;think thoroughly&rdquo; often beats a hand-written step-by-step plan.<\/li>\n<li><strong>Google.<\/strong> Gemini&rsquo;s thinking mode reasons before responding. You control how much it thinks with a budget, not with prompt phrasing.<\/li>\n<\/ul>\n<p>So: on a reasoning model, state the goal, the constraints and the output you want, and let it plan. Save explicit &ldquo;step 1, step 2&rdquo; instructions for standard models, or for cases where the steps are business rules the model can&rsquo;t know (&ldquo;check the price against our sheet before recommending&rdquo;).<\/p>\n<div style=\"background:#fff7ed;border-left:4px solid #ea580c;padding:14px 18px;margin:20px 0;border-radius:4px;\">\n<strong style=\"color:#c2410c;\">Warning<\/strong><\/p>\n<p style=\"margin:6px 0 0 0;color:#7c2d12;\">Thinking costs money. Reasoning tokens are billed (as output tokens on the major APIs), even when you never see them. Set a thinking or effort budget, and measure on your real workload before running a batch of hundreds.<\/p>\n<\/div>\n<h2>Four Templates for Content and SEO Work<\/h2>\n<p>These are written for standard models, where explicit structure helps. On a reasoning model, keep the goal, inputs and output format, and drop the numbered steps.<\/p>\n<h3>Template 1: Content brief<\/h3>\n<pre style=\"background:#1e293b;color:#e2e8f0;padding:20px;border-radius:6px;overflow-x:auto;font-size:14px;line-height:1.6;\">Keyword: \"[target keyword]\"\nWork through these before writing anything:\n1. The likely search intent (informational, commercial, navigational)\n2. The questions a searcher has next\n3. What the current top results cover, and what they leave out\n4. Three angles that fill that gap\nThen write a 300-word brief based on your analysis.<\/pre>\n<h3>Template 2: Title tag comparison<\/h3>\n<pre style=\"background:#1e293b;color:#e2e8f0;padding:20px;border-radius:6px;overflow-x:auto;font-size:14px;line-height:1.6;\">Compare two title tags for the same page:\nTitle A: \"[first title]\"\nTitle B: \"[second title]\"\nFor each, assess: keyword match, specificity, clarity of the promise,\nand how it would stand out among the current results.\nThen recommend one and explain why in two sentences.<\/pre>\n<h3>Template 3: Internal link map<\/h3>\n<pre style=\"background:#1e293b;color:#e2e8f0;padding:20px;border-radius:6px;overflow-x:auto;font-size:14px;line-height:1.6;\">Here are pages from my site: [paste URLs and titles]\n1. Group them into topic clusters\n2. Pick the strongest hub page in each cluster\n3. Decide which pages should link to each hub\n4. Flag pages with no natural link target\nOutput a JSON array of [source_url, target_url, anchor_text].<\/pre>\n<h3>Template 4: Search Console error triage<\/h3>\n<pre style=\"background:#1e293b;color:#e2e8f0;padding:20px;border-radius:6px;overflow-x:auto;font-size:14px;line-height:1.6;\">Here is my Search Console page indexing report: [paste]\nFor each issue type: the usual causes, the likely traffic impact,\nand a fix priority from 1 to 3.\nThen list the three fixes to do this week.<\/pre>\n<div style=\"background:#f0fdf4;border-left:4px solid #16a34a;padding:14px 18px;margin:20px 0;border-radius:4px;\">\n<strong style=\"color:#15803d;\">Pro Tip<\/strong><\/p>\n<p style=\"margin:6px 0 0 0;color:#14532d;\">End any reasoning prompt with a check: &ldquo;List any assumption you made and any fact you could not verify.&rdquo; It doesn&rsquo;t make the model right, but it shows you where to look before you act on the output.<\/p>\n<\/div>\n<p>For more structures like these, see our <a href=\"https:\/\/designcopy.net\/en\/advanced-prompting-techniques-guide\/\" data-wpel-link=\"internal\" rel=\"follow noopener noreferrer\" class=\"wpel-icon-right\">advanced prompting techniques guide<i class=\"wpel-icon dashicons-before dashicons-admin-page\" aria-hidden=\"true\"><\/i><\/a> and the <a href=\"https:\/\/designcopy.net\/en\/prompt-engineering-seo-guide\/\" data-wpel-link=\"internal\" rel=\"follow noopener noreferrer\" class=\"wpel-icon-right\">prompt engineering for SEO guide<i class=\"wpel-icon dashicons-before dashicons-admin-page\" aria-hidden=\"true\"><\/i><\/a>.<\/p>\n<h2>When Chain-of-Thought Goes Wrong<\/h2>\n<ul>\n<li><strong>Confident wrong reasoning.<\/strong> A fluent chain can lead to a wrong answer, and the explanation makes the error harder to spot on review. Check the facts, not the prose.<\/li>\n<li><strong>Reasoning without benefit.<\/strong> On simple lookups or rewrites, CoT adds length and cost without improving the result.<\/li>\n<li><strong>Inconsistent examples.<\/strong> In few-shot CoT, examples that reason in different styles produce inconsistent output. Keep them uniform.<\/li>\n<li><strong>Double reasoning.<\/strong> Asking a reasoning model to also think step by step in its answer wastes tokens and, per OpenAI, can hurt results.<\/li>\n<\/ul>\n<figure style=\"margin:24px 0;text-align:center;\"><img decoding=\"async\" src=\"https:\/\/designcopy.net\/wp-content\/uploads\/2026\/07\/chain-of-thought-prompting-gpt4o-claude-gemini-2026-internal-3-hero.jpg\" alt=\"How reasoning models handle chain-of-thought differently from standard models\" style=\"max-width:100%;height:auto;border-radius:8px;\" loading=\"lazy\" title=\"\"><\/figure>\n<h2>The Benchmarks You&rsquo;ll See Cited<\/h2>\n<table style=\"width:100%;border-collapse:collapse;margin:20px 0;\">\n<thead>\n<tr style=\"background:#1e3a5f;color:#fff;\">\n<th style=\"padding:10px 14px;text-align:left;\">Benchmark<\/th>\n<th style=\"padding:10px 14px;text-align:left;\">What it tests<\/th>\n<th style=\"padding:10px 14px;text-align:left;\">Origin<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr style=\"background:#f8fafc;\">\n<td style=\"padding:10px 14px;\">GSM8K<\/td>\n<td style=\"padding:10px 14px;\">About 8,500 grade-school maths word problems needing several steps<\/td>\n<td style=\"padding:10px 14px;\">OpenAI (Cobbe et al., 2021)<\/td>\n<\/tr>\n<tr>\n<td style=\"padding:10px 14px;\">MATH<\/td>\n<td style=\"padding:10px 14px;\">Competition-level maths problems<\/td>\n<td style=\"padding:10px 14px;\">Hendrycks et al., 2021<\/td>\n<\/tr>\n<tr style=\"background:#f8fafc;\">\n<td style=\"padding:10px 14px;\">MMLU<\/td>\n<td style=\"padding:10px 14px;\">Multiple-choice questions across 57 subjects<\/td>\n<td style=\"padding:10px 14px;\">Hendrycks et al., 2020<\/td>\n<\/tr>\n<tr>\n<td style=\"padding:10px 14px;\">HumanEval<\/td>\n<td style=\"padding:10px 14px;\">Writing Python functions that pass unit tests<\/td>\n<td style=\"padding:10px 14px;\">OpenAI (Chen et al., 2021)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Benchmarks tell you how models compare on standard problems. They don&rsquo;t tell you how a prompt will do on your content tasks, so test on a handful of your own real examples.<\/p>\n<div style=\"background:#eff6ff;border-left:4px solid #2563eb;padding:16px 20px;margin:24px 0;border-radius:4px;\">\n<strong style=\"color:#1d4ed8;font-size:1.05em;\">Key Takeaway<\/strong><\/p>\n<ul style=\"margin:8px 0 0 0;padding-left:20px;color:#1e3a8a;line-height:1.7;\">\n<li>Chain-of-thought is a research-backed technique (Wei et al. 2022; Kojima et al. 2022).<\/li>\n<li>On standard models, explicit steps and worked examples still help with multi-step tasks.<\/li>\n<li>On reasoning models, state the goal and output, and skip &ldquo;think step by step&rdquo;. The providers say so.<\/li>\n<li>Reasoning costs tokens; budget it and test on your own workload.<\/li>\n<\/ul>\n<\/div>\n<p>New to prompting? Start with our <a href=\"https:\/\/designcopy.net\/en\/prompt-engineering-guide\/\" data-wpel-link=\"internal\" rel=\"follow noopener noreferrer\" class=\"wpel-icon-right\">prompt engineering guide<i class=\"wpel-icon dashicons-before dashicons-admin-page\" aria-hidden=\"true\"><\/i><\/a>.<\/p>\n<h2>Frequently Asked Questions<\/h2>\n<h3>Does chain-of-thought prompting work on all AI models?<\/h3>\n<p>It helps most on large standard models. In the original 2022 research, small models didn&rsquo;t benefit. Reasoning models already reason internally, and OpenAI advises against chain-of-thought prompts for them.<\/p>\n<h3>What is the difference between zero-shot and few-shot CoT?<\/h3>\n<p>Zero-shot CoT adds a trigger such as &ldquo;Let&rsquo;s think step by step&rdquo; with no examples. Few-shot CoT includes worked examples showing the reasoning and the answer. Few-shot takes more effort and helps most when the task or format is specialised.<\/p>\n<h3>Should I tell a reasoning model to think step by step?<\/h3>\n<p>Generally no. OpenAI says it may not help its reasoning models and can hinder them, and Anthropic recommends general instructions such as &ldquo;think thoroughly&rdquo; over prescriptive steps when Claude&rsquo;s extended thinking is on. Give the goal, the constraints and the output format instead.<\/p>\n<h3>Can chain-of-thought reduce hallucinations?<\/h3>\n<p>Sometimes, on reasoning errors. It does not stop a model inventing facts, citations or statistics, and a confident chain can make an invented fact more convincing. Verify specifics against primary sources.<\/p>\n<h3>What is tree-of-thoughts prompting?<\/h3>\n<p>Tree of thoughts (Yao et al., 2023) has the model explore several reasoning branches, evaluate them and backtrack from dead ends. It suits planning and search problems; for everyday content tasks, a single structured prompt is usually enough.<\/p>\n<p style=\"font-size:0.85em;color:#64748b;margin-top:40px;border-top:1px solid #e2e8f0;padding-top:12px;\">Last updated: 2026-10-21. Sources: Wei et al. 2022 (arXiv:2201.11903), Kojima et al. 2022 (arXiv:2205.11916), Wang et al. 2022, Yao et al. 2023, OpenAI reasoning best practices, Anthropic prompting best practices.<\/p>\n<p><!-- designcopy-schema-start --><br \/>\n<script type=\"application\/ld+json\">\n{\n  \"@context\": \"https:\/\/schema.org\",\n  \"@type\": \"FAQPage\",\n  \"mainEntity\": [\n    {\n      \"@type\": \"Question\",\n      \"name\": \"Does chain-of-thought prompting work on all AI models?\",\n      \"acceptedAnswer\": {\n        \"@type\": \"Answer\",\n        \"text\": \"It helps most on large standard models. In the original 2022 research, small models didn\u2019t benefit. Reasoning models already reason internally, and OpenAI advises against chain-of-thought prompts for them.\"\n      }\n    },\n    {\n      \"@type\": \"Question\",\n      \"name\": \"What is the difference between zero-shot and few-shot CoT?\",\n      \"acceptedAnswer\": {\n        \"@type\": \"Answer\",\n        \"text\": \"Zero-shot CoT adds a trigger such as \u201cLet\u2019s think step by step\u201d with no examples. Few-shot CoT includes worked examples showing the reasoning and the answer. Few-shot takes more effort and helps most when the task or format is specialised.\"\n      }\n    },\n    {\n      \"@type\": \"Question\",\n      \"name\": \"Should I tell a reasoning model to think step by step?\",\n      \"acceptedAnswer\": {\n        \"@type\": \"Answer\",\n        \"text\": \"Generally no. OpenAI says it may not help its reasoning models and can hinder them, and Anthropic recommends general instructions such as \u201cthink thoroughly\u201d over prescriptive steps when Claude\u2019s extended thinking is on. Give the goal, the constraints and the output format instead.\"\n      }\n    },\n    {\n      \"@type\": \"Question\",\n      \"name\": \"Can chain-of-thought reduce hallucinations?\",\n      \"acceptedAnswer\": {\n        \"@type\": \"Answer\",\n        \"text\": \"Sometimes, on reasoning errors. It does not stop a model inventing facts, citations or statistics, and a confident chain can make an invented fact more convincing. Verify specifics against primary sources.\"\n      }\n    },\n    {\n      \"@type\": \"Question\",\n      \"name\": \"What is tree-of-thoughts prompting?\",\n      \"acceptedAnswer\": {\n        \"@type\": \"Answer\",\n        \"text\": \"Tree of thoughts (Yao et al., 2023) has the model explore several reasoning branches, evaluate them and backtrack from dead ends. It suits planning and search problems; for everyday content tasks, a single structured prompt is usually enough.\"\n      }\n    }\n  ]\n}\n<\/script><br \/>\n<!-- designcopy-schema-end --><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Chain-of-thought prompting asks an AI model to generate intermediate reasoning steps before committing to a final answer.<\/p>\n","protected":false},"author":1,"featured_media":265671,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_et_pb_use_builder":"","_et_pb_old_content":"","_et_gb_content_width":"","rank_math_title":"","rank_math_description":"Chain-of-thought prompting explained from the original research, the main variants, why reasoning models change the advice, and four templates for content and SEO work.","rank_math_focus_keyword":"chain of thought prompting","footnotes":""},"categories":[1483],"tags":[],"class_list":["post-265666","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-advanced-prompting-techniques","et-has-post-format-content","et_post_format-et-post-format-standard"],"_links":{"self":[{"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/posts\/265666","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/comments?post=265666"}],"version-history":[{"count":3,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/posts\/265666\/revisions"}],"predecessor-version":[{"id":266665,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/posts\/265666\/revisions\/266665"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/media\/265671"}],"wp:attachment":[{"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/media?parent=265666"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/categories?post=265666"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/designcopy.net\/en\/wp-json\/wp\/v2\/tags?post=265666"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}