{"id":1138,"date":"2025-09-20T22:01:51","date_gmt":"2025-09-20T20:01:51","guid":{"rendered":"https:\/\/www.bittner.fr\/blogs\/pierre\/?p=1138"},"modified":"2025-09-23T10:28:25","modified_gmt":"2025-09-23T08:28:25","slug":"agentic-ai-with-openai-gpt-5-does-the-empire-strike-back","status":"publish","type":"post","link":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/","title":{"rendered":"Deep Research with OpenAI GPT-5: Does the Empire strike back?"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><em>Pushing OpenAI\u2019s new flagship model to beat GPT-4.1 and a GPT-4.1 + o3-mini hybrid across Agentic AI (CoT\/ReAct and Plan\u2013Execute) workflows<\/em> for <a href=\"https:\/\/github.com\/BittnerPierre\/AI-Agent-Casebook\/tree\/feature\/optimize-agentic-workflows-gpt5\/experiments\/agentic-research\">Deep Research<\/a>.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In my <a href=\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/08\/29\/integrating-gpt-5-promises-challenges-and-early-lessons\/\">previous post<\/a>, the <strong>rebels<\/strong> (<a href=\"https:\/\/mistral.ai\/news\/mistral-small-3-1\">Mistral AI small<\/a>) held strong. I did a <strong>drop-in<\/strong> swap of <a href=\"https:\/\/openai.com\/index\/introducing-gpt-5\/\">GPT-5<\/a> with <strong>zero adaptation<\/strong> (model defaults, no prompt changes), and the result was parity at best.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This time, I came back with a more disciplined benchmark on <strong><a href=\"https:\/\/www.youtube.com\/watch?v=kSZyluqgRWo\">Deep Research<\/a><\/strong> (YT) using two orchestration patterns: <strong>Agentic Research<\/strong> (Chain-of-Thought\/ReAct-style supervisor) and <strong>Deep Agent<\/strong> (Plan\u2013Execute). The goal wasn\u2019t verbosity; it was <strong>structured reasoning under constraints<\/strong>.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>TL;DR \u2014 Short version<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GPT-5 wins.<\/strong> Even with <strong>minimal reasoning<\/strong> and <strong>low verbosity<\/strong>, it solves the tricky case, follows instructions more strictly, executes tool calls <strong>and<\/strong> produces structured outputs correctly, and needs <strong>no sidekick<\/strong> model. The hybrid <strong>GPT-4.1-mini + o3-mini<\/strong> remains strong, but it\u2019s <strong>slower<\/strong>, <strong>costlier in tokens<\/strong>, and <strong>more complex<\/strong> to operate. <strong>GPT-4.1-mini<\/strong> alone consistently struggles on the complex planning test. If you\u2019re starting new work on deep research pipelines, <strong>pick GPT-5<\/strong>.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Why Mistral first, and why a second wave?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">In the first wave I wanted a <strong>\u201cfair drop-in\u201d scenario<\/strong>:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Keep <strong>GPT-5 default parameters<\/strong>  (which are <em>medium<\/em> reasoning &amp; verbosity; which actually hurt performance in small-model-tuned workflows).<\/li>\n\n\n\n<li><strong>No prompt surgery<\/strong> or tool routing tweaks.<\/li>\n\n\n\n<li>Just swap models into existing workflows and watch.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Mistral AI<\/strong> came out as a <strong>legit contender<\/strong> under those rules. GPT-5 didn\u2019t collapse, but it also didn\u2019t obviously dominate. That told me: <em>drop-in parity \u2260 true capability.<\/em><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">So I launched a <strong>second wave<\/strong> to answer: <em>If I meet GPT-5 halfway\u2014just enough to unlock its sweet spot\u2014does it pull ahead decisively?<\/em><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">What changed in the second wave (and why)<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">I kept changes <strong>surgical<\/strong> and transparent:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Use GPT-5-mini with Agents SDK default values<\/strong><\/li>\n<\/ol>\n\n\n\n<pre class=\"wp-block-code\"><code>model_settings = ModelSettings(reasoning=Reasoning(effort=\"minimal\"), verbosity=\"low\")\n# or\nmodel_settings = get_default_model_settings(extract_model_name(config.models.search_model))\n# optional if you standardize on GPT-5\n# OPENAI_DEFAULT_MODEL=gpt-5\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">I stayed at <strong>minimal reasoning \/ low verbosity<\/strong> intentionally\u2014if GPT-5 is as good as claimed, it shouldn\u2019t need \u201chigh-effort\u201d to win.<\/p>\n\n\n\n<ol start=\"2\" class=\"wp-block-list\">\n<li><strong>Upgrade evaluation: add LLM-as-judge (content + spec)<\/strong><\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">My original <strong>path evaluation<\/strong> (did the agent call the right tools, in the right order?) wasn\u2019t sufficient. I added an <strong>LLM-as-judge<\/strong> to verify <strong>content<\/strong> and <strong>contract<\/strong>:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>It\u2019s a <strong>research report<\/strong>, not a syllabus\/podcast\/script.<\/li>\n\n\n\n<li>Required <strong>sections<\/strong> are present (Raw Notes \u2192 Detailed Agenda \u2192 Report).<\/li>\n\n\n\n<li><strong>Adherence to length &amp; style specs<\/strong> (e.g., 250\u2013300 words for the report).<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">This also solved a practical gotcha: <strong>GPT-5 tends to over-structure<\/strong> outputs (adds parentheses in headings, translates constants like \u201cRaw Notes\u201d into French, etc.). My old <code>grep<\/code> checks broke on that strict matching. <strong>LLM-as-judge<\/strong> made evaluation robust without penalizing readability.<\/p>\n\n\n\n<ol start=\"3\" class=\"wp-block-list\">\n<li><strong>Fix the input contract<\/strong><\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">A subtle bug: I had wrapped the user request in <code>&lt;syllabus\/&gt;<\/code>, so GPT-5 kept returning <strong>course outlines<\/strong> instead of <strong>research reports<\/strong>. GPT-5 is <strong>prompt-strict<\/strong>; if your input is contradictory, it will <strong>refuse<\/strong> or follow it <strong>literally<\/strong>. That\u2019s a <em>great developer asset<\/em> (fails fast), but the payload had to be cleaned up.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Where GPT-5 immediately pulled ahead: the writer agent<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The writer has a <strong>dual goal<\/strong> in one turn:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Execute function calls \/ handoffs<\/strong>; and<\/li>\n\n\n\n<li><strong>Emit a structured report<\/strong> (three sections: Raw Notes \u2192 Agenda \u2192 Report).<\/li>\n<\/ol>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>GPT-4.1-mini<\/strong> consistently <strong>refused<\/strong> to do both simultaneously (wanted JSON <strong>or<\/strong> function calling, not both).<\/li>\n\n\n\n<li><strong>Mistral<\/strong> <em>could<\/em> do it after iterations.<\/li>\n\n\n\n<li><strong>GPT-5<\/strong> did it <strong>on the first try<\/strong>.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">This mattered later: if a model can\u2019t meet <strong>two end goals<\/strong>, you burn time building glue code.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Two workflows, one question<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">I ran the <strong>same research ask<\/strong> through two orchestration patterns:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Agentic Research<\/strong> = <strong>CoT\/ReAct-style supervision<\/strong> (one supervisor that thinks <em>and<\/em> acts).<\/li>\n\n\n\n<li><strong>Deep Research<\/strong> = <strong>Plan\u2013Execute<\/strong> (a planner proposes; an executor runs the plan).<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"382\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?resize=1024%2C382&#038;ssl=1\" alt=\"\" class=\"wp-image-1145\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?resize=1024%2C382&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?resize=300%2C112&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?resize=768%2C287&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?resize=1536%2C573&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?w=1656&amp;ssl=1 1656w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">Deep Research avec Chain-of-Thought &amp; ReACT<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager-scaled.jpeg?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"516\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager.jpeg?resize=1024%2C516&#038;ssl=1\" alt=\"\" class=\"wp-image-1146\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager-scaled.jpeg?resize=1024%2C516&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager-scaled.jpeg?resize=300%2C151&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager-scaled.jpeg?resize=768%2C387&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager-scaled.jpeg?resize=1536%2C774&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Deep-Research-Manager-scaled.jpeg?resize=2048%2C1033&amp;ssl=1 2048w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">Agentic AI Research avec plan-execute<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">Step 1 \u2014 Baseline (unit) test<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">A deliberately easy prompt: <em>\u201cWrite a short, neutral, technical report on <strong>MIPS + ANN<\/strong> with two algorithms (HNSW, FAISS) and compare with <strong>ReWOO<\/strong>; 250\u2013300 words; 4\u20135 sections; no off-topic multi-agent content.\u201d<\/em> The goal is <strong>format &amp; content sanity<\/strong> across models and both workflows.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Expectation:<\/strong> everyone should pass. If there\u2019s no clear winner here, move to a stressor.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Step 2 \u2014 Tricky planning test (stressor)<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">A <strong>planning-first puzzle<\/strong> to force stepwise reasoning during <strong>plan generation<\/strong>:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Hard budget:<\/strong> cap searches to <strong>5\u20137<\/strong>. If evidence is missing, <strong>stop<\/strong> (don\u2019t keep digging forever).<\/li>\n\n\n\n<li><strong>Split evidence across files:<\/strong> the model must <strong>connect facts<\/strong> scattered across different docs in the knowledge base.<\/li>\n\n\n\n<li><strong>Inject inconsistencies:<\/strong> include <strong>contradictory hints<\/strong> that must be reconciled.<\/li>\n\n\n\n<li><strong>No pre-specified sources:<\/strong> the model must <strong>choose<\/strong> the right files on its own.<\/li>\n\n\n\n<li><strong>Strict output contract:<\/strong> produce the <strong>three sections<\/strong> (Raw Notes \u2192 Detailed Agenda \u2192 Report). No syllabus, no podcast, no script.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Why this is hard:<\/strong> non-reasoning models tend to either (a) <strong>over-search<\/strong> (latency\/cost blow-ups) or (b) <strong>guess<\/strong> and miss cross-file links. A <strong>reasoning-capable<\/strong> model should plan, check, course-correct, and still respect the budget.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Results<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Baseline (unit) test \u2014 everyone passes<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">All models produced valid reports with the right scope and sectioning across both workflows. No material functional gaps here.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Run<\/th><th>Workflow<\/th><th>Model<\/th><th>Content (0\u20135)<\/th><th>Spec (0\u20135)<\/th><th>Notes<\/th><\/tr><\/thead><tbody><tr><td>1<\/td><td>Agentic Research<\/td><td>GPT-5-mini<\/td><td>5<\/td><td>5<\/td><td>Clean, concise; minor stylistic flourishes ok.<\/td><\/tr><tr><td>2<\/td><td>Deep Research<\/td><td>GPT-5-mini<\/td><td>5<\/td><td>5<\/td><td>Same quality; slightly tighter by default.<\/td><\/tr><tr><td>3<\/td><td>Deep Research<\/td><td>GPT-4.1-mini<\/td><td>5<\/td><td>4\u20135<\/td><td>Passes; occasional minor spec drift.<\/td><\/tr><tr><td>4<\/td><td>Agentic Research<\/td><td>GPT-4.1-mini<\/td><td>5<\/td><td>4<\/td><td>Passes; supervisor sometimes over-elaborates.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">Latency &amp; tokens for the baseline were <strong>close<\/strong> for this short task, so I focus on the stressor where gaps actually appear.<\/p>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\">Tricky planning test \u2014 this is where GPT-5 separates<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Here the plan quality, budget discipline, and structured output matter. I also added the <strong>hybrid arm<\/strong> (GPT-4.1-mini + o3-mini) <strong>only<\/strong> for this test to benchmark a \u201creasoning-assist\u201d strategy.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Run<\/th><th>Workflow<\/th><th>Model<\/th><th>Content (0\u20135)<\/th><th>Spec (0\u20135)<\/th><th>Tokens (\u2248)<\/th><th>Latency (\u2248)<\/th><th>Notes<\/th><\/tr><\/thead><tbody><tr><td>1<\/td><td>Agentic Research<\/td><td><strong>GPT-5-mini<\/strong><\/td><td><strong>5<\/strong><\/td><td><strong>4<\/strong><\/td><td>~250k<\/td><td>~40 s<\/td><td>Strong; tiny verbosity drift.<\/td><\/tr><tr><td>2<\/td><td>Deep Research<\/td><td><strong>GPT-5-mini<\/strong><\/td><td><strong>5<\/strong><\/td><td><strong>5<\/strong><\/td><td>~220k<\/td><td>~35 s<\/td><td><strong>Winner<\/strong>: clean plan; respects 5\u20137 search cap; nails structure.<\/td><\/tr><tr><td>3<\/td><td>Deep Research<\/td><td>GPT-4.1-mini<\/td><td>3<\/td><td>1<\/td><td>~500k<\/td><td>~80 s<\/td><td>Off-spec (drifts), misses structure; weaker plan discipline.<\/td><\/tr><tr><td>4<\/td><td>Agentic Research<\/td><td>GPT-4.1-mini<\/td><td>4<\/td><td>3<\/td><td><strong>~1.88M<\/strong><\/td><td><strong>~265 s<\/strong><\/td><td><strong>Loops searches<\/strong> beyond budget; expensive; slower.<\/td><\/tr><tr><td>5<\/td><td>Agentic Research<\/td><td>GPT-4.1-mini + <strong>o3-mini<\/strong><\/td><td>5<\/td><td>4<\/td><td>~500k<\/td><td>~90 s<\/td><td>Solves the puzzle; <strong>slower + costlier<\/strong>; operationally heavier.<\/td><\/tr><tr><td>6<\/td><td>Deep Research<\/td><td>GPT-4.1-mini + <strong>o3-mini<\/strong><\/td><td>5<\/td><td>4<\/td><td>~430k<\/td><td>~75 s<\/td><td>Strong; still <strong>slower<\/strong> than GPT-5-mini (Run 2).<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"259\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?resize=1024%2C259&#038;ssl=1\" alt=\"\" class=\"wp-image-1151\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?resize=1024%2C259&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?resize=300%2C76&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?resize=768%2C194&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?resize=1536%2C388&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Benchmark.jpg?resize=2048%2C517&amp;ssl=1 2048w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">Performance comparison on Langsmith between GPT-5-mini, GPT-4.1-mini and GPT-4.1-mini + o3-mini<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Takeaways<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Baseline:<\/strong> all pass; no big winner \u2192 hence the stressor.<\/li>\n\n\n\n<li><strong>Tricky:<\/strong> <strong>GPT-5-mini (Run 2)<\/strong> wins; the <strong>hybrid<\/strong> is competitive but <strong>slower &amp; pricier<\/strong>; <strong>GPT-4.1-mini alone<\/strong> struggles with plan discipline and spec adherence.<\/li>\n\n\n\n<li><strong>Budget respect:<\/strong> GPT-5 reliably stops at <strong>5\u20137 searches<\/strong>; GPT-4.1 often keeps digging.<\/li>\n\n\n\n<li><strong>Dual goal:<\/strong> GPT-5 handled <strong>function-calling + structured output<\/strong>; GPT-4.1 refused the dual objective. Here for parity, dual goals was remove (no save_report function calling in Writer agent). This to avoid penalty on latency and token consumption).<\/li>\n<\/ul>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>Note on Mistral<\/strong>: I excluded Mistral from the <strong>global workflow ranking<\/strong> here because our <code>file_search<\/code> agent relies on OpenAI\u2019s Response API for retrieval, biasing comparisons. It remains a strong contender, but not comparable under this setup.<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Why GPT-5 pulled ahead<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Instruction fidelity<\/strong> \u2014 It\u2019s <em>prompt-strict<\/em>. Clean inputs \u2192 clean behavior. Contradictory inputs (e.g., a stray <code>&lt;syllabus\/&gt;<\/code>) make it <strong>refuse<\/strong> or <strong>follow literally<\/strong>, which is a <strong>developer asset<\/strong>: failures surface early.<\/li>\n\n\n\n<li><strong>Programmability<\/strong> \u2014 It reliably does <strong>function calls \/ handoffs<\/strong> <em>and<\/em> returns <strong>structured outputs<\/strong> in the same turn, which is crucial for supervisors and writer agents.<\/li>\n\n\n\n<li><strong>Search discipline<\/strong> \u2014 It <strong>respects budgets<\/strong> (\u00ab\u00a0max 7 searches\u00a0\u00bb), protecting <strong>latency<\/strong> and <strong>cost<\/strong>; GPT-4.1 tends to chase.<\/li>\n\n\n\n<li><strong>Writing quality &amp; structure<\/strong> \u2014 Even with <strong>low reasoning<\/strong>, outputs are technical and well-structured. GPT-5 also tends to <strong>over-structure<\/strong> (parentheses, translated headers), which broke my old <code>grep<\/code> checks; switching to <strong>LLM-as-judge<\/strong> fixed evaluation without penalizing clarity.<\/li>\n\n\n\n<li><strong>No sidekick required<\/strong> \u2014 The GPT-4.1-mini + o3-mini hybrid is good, but <strong>slower, costlier, and more operationally complex<\/strong>. GPT-5 delivers the capability in <strong>one<\/strong> model.<\/li>\n<\/ol>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"903\" height=\"1024\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?resize=903%2C1024&#038;ssl=1\" alt=\"\" class=\"wp-image-1147\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?resize=903%2C1024&amp;ssl=1 903w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?resize=264%2C300&amp;ssl=1 264w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?resize=768%2C871&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?resize=1354%2C1536&amp;ssl=1 1354w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/file_search_gpt-4-1_endless_loop.png?w=1444&amp;ssl=1 1444w\" sizes=\"auto, (max-width: 903px) 100vw, 903px\" \/><\/a><figcaption class=\"wp-element-caption\">GPT-4.1 going in an endless loop for file_search<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini-scaled.jpg?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"570\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini.jpg?resize=1024%2C570&#038;ssl=1\" alt=\"\" class=\"wp-image-1149\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini-scaled.jpg?resize=1024%2C570&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini-scaled.jpg?resize=300%2C167&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini-scaled.jpg?resize=768%2C428&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini-scaled.jpg?resize=1536%2C855&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-problem-for-gpt-4.1-mini-scaled.jpg?resize=2048%2C1140&amp;ssl=1 2048w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">No query, no problem for GPT-4.1-mini<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini-scaled.jpg?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"482\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini.jpg?resize=1024%2C482&#038;ssl=1\" alt=\"\" class=\"wp-image-1150\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini-scaled.jpg?resize=1024%2C482&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini-scaled.jpg?resize=300%2C141&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini-scaled.jpg?resize=768%2C362&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini-scaled.jpg?resize=1536%2C723&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/no-query-no-way-for-gpt-5-mini-scaled.jpg?resize=2048%2C964&amp;ssl=1 2048w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">No query, no way for GPT-5-mini<\/figcaption><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Notes on integration<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Agents SDK<\/strong> plays nicer with GPT-5 than my older LangChain\/LangGraph stacks for these tests. Small prompt fixes + <strong>minimal reasoning \/ low verbosity<\/strong> were enough. <\/li>\n\n\n\n<li><strong>Evaluation harness<\/strong>: keep <strong>path tracing<\/strong> (tool-call traces) <em>and<\/em> <strong>LLM-as-judge<\/strong> (content + spec). It catches \u201clooks-right but wrong-form\u201d failures and protects against syllabus drift.<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">Latency &amp; token overhead \u2014 GPT-5 vs GPT-4.1<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">From the <strong>waterfall traces<\/strong>, GPT-5-mini (low reasoning) shows a slight <strong>per-call overhead<\/strong> compared to GPT-4.1-mini:<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Model<\/th><th>Total Tokens (\u2248)<\/th><th>Latency (\u2248)<\/th><th>Notes<\/th><\/tr><\/thead><tbody><tr><td>GPT-5-mini<\/td><td>~263k<\/td><td>~163 s<\/td><td>Slower per call, lower token usage.<\/td><\/tr><tr><td>GPT-4.1-mini<\/td><td>~313k<\/td><td>~135 s<\/td><td>Faster per call, higher token usage.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">This gap reinforces the need to <strong>optimize workflows<\/strong> (<a href=\"https:\/\/www.deeplearning.ai\/the-batch\/issue-316\/\">parallel agents<\/a>, removing deterministic tool calls) if GPT-5 is to be deployed in production-heavy environments.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall-scaled.png?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"575\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall.png?resize=1024%2C575&#038;ssl=1\" alt=\"\" class=\"wp-image-1148\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall-scaled.png?resize=1024%2C575&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall-scaled.png?resize=300%2C168&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall-scaled.png?resize=768%2C431&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall-scaled.png?resize=1536%2C862&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/waterfall-scaled.png?resize=2048%2C1149&amp;ssl=1 2048w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">Comparison of waterfall execution between GPT-4.1-mini (left) and GPT-5-mini (right)<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?ssl=1\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"457\" src=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?resize=1024%2C457&#038;ssl=1\" alt=\"\" class=\"wp-image-1161\" srcset=\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?resize=1024%2C457&amp;ssl=1 1024w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?resize=300%2C134&amp;ssl=1 300w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?resize=768%2C342&amp;ssl=1 768w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?resize=1536%2C685&amp;ssl=1 1536w, https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/sequential-call.jpg?resize=2048%2C913&amp;ssl=1 2048w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><figcaption class=\"wp-element-caption\">Function calling is sequential with CoT\/ReACT increasing global latency.<\/figcaption><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Should you switch?<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Already stable on GPT-4.1 in prod?<\/strong> Don\u2019t rush. You\u2019ll touch prompts and deployment threads to get the most from GPT-5.<\/li>\n\n\n\n<li><strong>Starting new work or evolving a research pipeline?<\/strong> Choose <strong>GPT-5<\/strong>. It\u2019s <strong>more reliable, more programmable, more versatile<\/strong>, and it <strong>already wins<\/strong> in <strong>minimal reasoning + low verbosity<\/strong>.<\/li>\n<\/ul>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">As always, keep an <strong>evaluation harness<\/strong> (path tracing + LLM-as-judge) in place. GPT-5 is powerful; good tests make it predictable.<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Appendix: test inputs<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Baseline prompt<\/h3>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">Summarize how autonomous agents use external memory retrieval through Maximum Inner Product Search (MIPS).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Explain why Approximate Nearest Neighbor (ANN) methods are required in this context.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">List and briefly describe the main ANN algorithms discussed in the reference (LSH, ANNOY, HNSW, FAISS, ScaNN), keeping each description to a single sentence.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Conclude with a short explanation of the trade-off between latency and recall in such systems.<\/p>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\">Tricky planning prompt (stressor)<\/h3>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><em>Write a report explaining how LLM agents use external memory using Maximum Inner Product Search (MIPS), why they use Approximate Nearest Neighbor (ANN), giving two examples of ANN algorithms, and compare this approach with ReWOO.<\/em><\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Why this is a stressor compared to baseline:<\/strong><br>While the baseline is essentially an <strong>enumeration<\/strong> task (explain, list, conclude), the tricky prompt is <strong>synthetic<\/strong>: it requires <strong>choice<\/strong> (pick two ANN algorithms), <strong>integration across sources<\/strong> (MIPS\/ANN vs ReWOO come from different docs), and a <strong>comparative analysis<\/strong> under a <strong>strict budget of searches<\/strong>. Non-reasoning models tend to either <strong>over-search<\/strong> or <strong>guess poorly<\/strong> here. Reasoning-capable models (like GPT-5) can plan, reconcile contradictions, and still meet the contract.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Appendix B: Notes &amp; reflections<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Reasoning defaults<\/strong>: GPT-5 defaults to <em>medium<\/em> reasoning, which explains why it underperformed in drop-in tests (over-thinking vs small-model-tuned workflows). Setting to <strong>low<\/strong> makes it behave closer to non-reasoning models, but with instruction-following and reasoning intact.<\/li>\n\n\n\n<li><strong>Prompt strictness<\/strong>: <code>&lt;syllabus\/&gt;<\/code> issue wasn\u2019t malformed but contradictory. GPT-5 refused or followed literally \u2192 a developer asset (fail fast).<\/li>\n\n\n\n<li><strong>Writer agent trajectory test<\/strong>: GPT-5 succeeded immediately; GPT-4 worked too; GPT-4.1 refused (dual-goal error); Mistral managed after iterations. GPT-5 likely benefited from earlier alignment efforts.<\/li>\n\n\n\n<li><strong>Instruction compatibility<\/strong>: When instructions were updated for GPT-5, Mistral and GPT-4.1 still worked fine \u2192 no need for per-model versioning.<\/li>\n\n\n\n<li><strong>Latency gap<\/strong>: GPT-5-mini slower per call than GPT-4.1-mini (see waterfall table). Next steps: parallelize agents, remove deterministic function calls (reading\/writing) from agent loop.<\/li>\n\n\n\n<li><strong>Agentic vs Deep parity<\/strong>: GPT-5 sometimes parallelized tool calls in Agentic Research, bringing performance closer to Deep Research (plan\u2013execute). A sign it bridges both.<\/li>\n\n\n\n<li><strong>Evaluation harness<\/strong>: Added LLM-as-judge (content + spec) on top of path eval (tool calls, budget) to catch off-topic drifts (syllabus-style).<\/li>\n\n\n\n<li><strong>Sidekick vs solo<\/strong>: Hybrid GPT-4.1+o3-mini is strong but operationally heavier; GPT-5 alone is simpler, cheaper, and wins.<\/li>\n\n\n\n<li><strong>Next step<\/strong>: Optimize workflows to cut runtime overhead; explore parallelization patterns and deterministic tool separation.<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Pushing OpenAI\u2019s new flagship model to beat GPT-4.1 and a GPT-4.1 + o3-mini hybrid across Agentic AI (CoT\/ReAct and Plan\u2013Execute) workflows for Deep Research. In my previous post, the rebels (Mistral AI small) held strong. I did a drop-in swap of GPT-5 with zero adaptation (model defaults, no prompt changes), and the result was parity [&hellip;]<\/p>\n","protected":false},"author":12,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-1138","post","type-post","status-publish","format-standard","hentry","category-monologue"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.0 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Deep Research with OpenAI GPT-5: Does the Empire strike back? - Pierre&#039;s weblog<\/title>\n<meta name=\"description\" content=\"Discover how OpenAI GPT-5, the new flagship model for Agentic AI and Deep Research, outperforms GPT-4.1 and o3-mini in complex workflows.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Deep Research with OpenAI GPT-5: Does the Empire strike back? - Pierre&#039;s weblog\" \/>\n<meta property=\"og:description\" content=\"Discover how OpenAI GPT-5, the new flagship model for Agentic AI and Deep Research, outperforms GPT-4.1 and o3-mini in complex workflows.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\" \/>\n<meta property=\"og:site_name\" content=\"Pierre&#039;s weblog\" \/>\n<meta property=\"article:published_time\" content=\"2025-09-20T20:01:51+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2025-09-23T08:28:25+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager-1024x382.png\" \/>\n<meta name=\"author\" content=\"Pierre Bittner\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u00c9crit par\" \/>\n\t<meta name=\"twitter:data1\" content=\"Pierre Bittner\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\"},\"author\":{\"name\":\"Pierre Bittner\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/9713c6deaf14c30b1b34cc5cd886c09c\"},\"headline\":\"Deep Research with OpenAI GPT-5: Does the Empire strike back?\",\"datePublished\":\"2025-09-20T20:01:51+00:00\",\"dateModified\":\"2025-09-23T08:28:25+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\"},\"wordCount\":1897,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/3ab57f92769bc1c95b4f56d526b0b225\"},\"image\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager-1024x382.png\",\"articleSection\":[\"Monologue\"],\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\",\"url\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\",\"name\":\"Deep Research with OpenAI GPT-5: Does the Empire strike back? - Pierre&#039;s weblog\",\"isPartOf\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager-1024x382.png\",\"datePublished\":\"2025-09-20T20:01:51+00:00\",\"dateModified\":\"2025-09-23T08:28:25+00:00\",\"description\":\"Discover how OpenAI GPT-5, the new flagship model for Agentic AI and Deep Research, outperforms GPT-4.1 and o3-mini in complex workflows.\",\"breadcrumb\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#breadcrumb\"},\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage\",\"url\":\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?fit=1656%2C618&ssl=1\",\"contentUrl\":\"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?fit=1656%2C618&ssl=1\",\"width\":1656,\"height\":618},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Accueil\",\"item\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Deep Research with OpenAI GPT-5: Does the Empire strike back?\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#website\",\"url\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/\",\"name\":\"Pierre's weblog\",\"description\":\"Release 2.3\",\"publisher\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/3ab57f92769bc1c95b4f56d526b0b225\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"fr-FR\"},{\"@type\":[\"Person\",\"Organization\"],\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/3ab57f92769bc1c95b4f56d526b0b225\",\"name\":\"Pierre\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Pierre-Bittner.png\",\"contentUrl\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Pierre-Bittner.png\",\"width\":1920,\"height\":1080,\"caption\":\"Pierre\"},\"logo\":{\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/image\/\"},\"sameAs\":[\"http:\/\/www.bittner.fr\/blogs\/pierre\",\"https:\/\/www.linkedin.com\/in\/pierre-bittner\/\",\"https:\/\/www.youtube.com\/@pierrebittner\"]},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/9713c6deaf14c30b1b34cc5cd886c09c\",\"name\":\"Pierre Bittner\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/381c6e1f43681b025b7b709d1299281ceb4b5174454b52b648c8be840b70d522?s=96&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/381c6e1f43681b025b7b709d1299281ceb4b5174454b52b648c8be840b70d522?s=96&d=mm&r=g\",\"caption\":\"Pierre Bittner\"},\"url\":\"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/author\/pierre\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Deep Research with OpenAI GPT-5: Does the Empire strike back? - Pierre&#039;s weblog","description":"Discover how OpenAI GPT-5, the new flagship model for Agentic AI and Deep Research, outperforms GPT-4.1 and o3-mini in complex workflows.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/","og_locale":"fr_FR","og_type":"article","og_title":"Deep Research with OpenAI GPT-5: Does the Empire strike back? - Pierre&#039;s weblog","og_description":"Discover how OpenAI GPT-5, the new flagship model for Agentic AI and Deep Research, outperforms GPT-4.1 and o3-mini in complex workflows.","og_url":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/","og_site_name":"Pierre&#039;s weblog","article_published_time":"2025-09-20T20:01:51+00:00","article_modified_time":"2025-09-23T08:28:25+00:00","og_image":[{"url":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager-1024x382.png","type":"","width":"","height":""}],"author":"Pierre Bittner","twitter_card":"summary_large_image","twitter_misc":{"\u00c9crit par":"Pierre Bittner"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#article","isPartOf":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/"},"author":{"name":"Pierre Bittner","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/9713c6deaf14c30b1b34cc5cd886c09c"},"headline":"Deep Research with OpenAI GPT-5: Does the Empire strike back?","datePublished":"2025-09-20T20:01:51+00:00","dateModified":"2025-09-23T08:28:25+00:00","mainEntityOfPage":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/"},"wordCount":1897,"commentCount":0,"publisher":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/3ab57f92769bc1c95b4f56d526b0b225"},"image":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage"},"thumbnailUrl":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager-1024x382.png","articleSection":["Monologue"],"inLanguage":"fr-FR","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/","url":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/","name":"Deep Research with OpenAI GPT-5: Does the Empire strike back? - Pierre&#039;s weblog","isPartOf":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage"},"image":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage"},"thumbnailUrl":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager-1024x382.png","datePublished":"2025-09-20T20:01:51+00:00","dateModified":"2025-09-23T08:28:25+00:00","description":"Discover how OpenAI GPT-5, the new flagship model for Agentic AI and Deep Research, outperforms GPT-4.1 and o3-mini in complex workflows.","breadcrumb":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#breadcrumb"},"inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#primaryimage","url":"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?fit=1656%2C618&ssl=1","contentUrl":"https:\/\/i0.wp.com\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Agentic-Research-Manager.png?fit=1656%2C618&ssl=1","width":1656,"height":618},{"@type":"BreadcrumbList","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/2025\/09\/20\/agentic-ai-with-openai-gpt-5-does-the-empire-strike-back\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Accueil","item":"https:\/\/www.bittner.fr\/blogs\/pierre\/"},{"@type":"ListItem","position":2,"name":"Deep Research with OpenAI GPT-5: Does the Empire strike back?"}]},{"@type":"WebSite","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#website","url":"https:\/\/www.bittner.fr\/blogs\/pierre\/","name":"Pierre's weblog","description":"Release 2.3","publisher":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/3ab57f92769bc1c95b4f56d526b0b225"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.bittner.fr\/blogs\/pierre\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"fr-FR"},{"@type":["Person","Organization"],"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/3ab57f92769bc1c95b4f56d526b0b225","name":"Pierre","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/image\/","url":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Pierre-Bittner.png","contentUrl":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-content\/uploads\/2025\/09\/Pierre-Bittner.png","width":1920,"height":1080,"caption":"Pierre"},"logo":{"@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/image\/"},"sameAs":["http:\/\/www.bittner.fr\/blogs\/pierre","https:\/\/www.linkedin.com\/in\/pierre-bittner\/","https:\/\/www.youtube.com\/@pierrebittner"]},{"@type":"Person","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/9713c6deaf14c30b1b34cc5cd886c09c","name":"Pierre Bittner","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.bittner.fr\/blogs\/pierre\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/381c6e1f43681b025b7b709d1299281ceb4b5174454b52b648c8be840b70d522?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/381c6e1f43681b025b7b709d1299281ceb4b5174454b52b648c8be840b70d522?s=96&d=mm&r=g","caption":"Pierre Bittner"},"url":"https:\/\/www.bittner.fr\/blogs\/pierre\/blog\/author\/pierre\/"}]}},"_links":{"self":[{"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/posts\/1138","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/users\/12"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/comments?post=1138"}],"version-history":[{"count":10,"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/posts\/1138\/revisions"}],"predecessor-version":[{"id":1163,"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/posts\/1138\/revisions\/1163"}],"wp:attachment":[{"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/media?parent=1138"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/categories?post=1138"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bittner.fr\/blogs\/pierre\/wp-json\/wp\/v2\/tags?post=1138"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}