Transcription
If you're using GPT 5.5 the same way you used 5.4, you're barely scratching the surface of what this model can do. The truth is, GPT 5.5 is faster, produces smarter outputs, and burns through fewer tokens than any model OpenAI has released before. But only if you change two things about how you set it up and how you prompt it. So, I'm going to show you exactly how to get better results than 99% of users with specific examples on screen for every technique.
The first thing to know about GPT 5.5 is that the way you access it inside ChatGPT is different from how previous models worked. And getting this wrong means you're either using the wrong model without knowing it, or you're burning through usage on settings you don't need. I'll head over to chat.openai.com and click the drop down at the right side of the text box. You'll see two options: instant and thinking, and underneath those there's a configure button. Instant is GPT 5.5 in fast mode for everyday answers. Thinking is GPT 5.5 with deeper reasoning for complex work, and pro is GPT 5.5 pro for the hardest tasks. The mistake I see constantly is people staying on instant and assuming they're getting the full power of 5.5 on every prompt.
Instant and thinking are both running GPT 5.5, but they use it completely differently. Instant is built for speed, so it gives you fast answers without spending time on deep reasoning. Thinking is built for depth, so it plans its approach, checks its own work, and reasons through complex problems step-by-step before responding. Instant can auto switch to thinking on complex prompts, but for simpler questions it stays in fast mode, and the interface doesn't always make it obvious which mode just handled your prompt. So, if you want GPT 5.5's full reasoning on every conversation, you need to manually select thinking from the picker.
To show you why that matters, I'll run the same prompt twice, once on instant and once on thinking. I'll type, "Give me a breakdown of the competitive landscape in the enterprise AI market, including valuations, revenue, and which companies are best positioned for the next 12 months." On instant, the response comes back fast but surface level, with hedged numbers and consensus framings that read like a summary stitched together from three or four tech articles. On thinking, the same prompt produces a noticeably different output. The numbers are more specific and sourced. The analysis works through layers to explain why each company is positioned the way it is, and the response is structured around rankings and tradeoffs rather than just information. That's the gap between fast mode and deep reasoning on the same model. And if you stay on instant, you're missing the part of 5.5 that makes it worth the upgrade.
And the second part of the setup is the thing that saved me the most usage this week. When you select thinking, you'll see a thinking effort option with two levels, standard and extended. Standard is the default and handles the majority of tasks well, while extended give the model more time to reason through harder problems. The instinct is to push the effort up to extended on everything because more thinking sounds like it should mean better output. On GPT 5.5, that's no longer true. The standard setting on 5.5 matches or beats what 5.4 produced on its highest settings for the majority of tasks because 5.5 is built on a larger base model where each token carries more intelligence than before. In a full week of testing, I only pushed the effort above standard a handful of times on tasks that were particularly complex. And the rest of the time standard produced output just as good while using significantly fewer tokens and finishing much faster.
This one change, selecting thinking manually and leaving the effort on standard instead of cranking it up on everything, is the single biggest efficiency gain you can make on GPT 5.5 because you're getting full reasoning depth at near instant speed without burning through your weekly allowance. And once you have that setup right, the next step is changing how you prompt it because 5.5 handles prompts differently than any model before it.
The biggest shift in how GPT 5.5 processes prompts is that it handles multi-step tasks in a single pass better than any previous model. And that changes how you should be structuring your requests. On older models, the best practice was to break complex work into small pieces. Upload a document in one message, ask for a summary in the next, then ask for the analysis, then ask for the output you actually need, which is four separate prompts for one task. On GPT 5.5, you can skip all of that and give it the entire job in one prompt because the model is better at planning its own approach, holding context between steps, and carrying the output from step one into step two without losing anything.
I'll show you what this looks like with a document task. I'll upload a 20-page quarterly report and paste in this prompt. On 5.4, this kind of stacked prompt would often result in the model forgetting the earlier parts by the time it reached the email draft or producing a summary that didn't connect to the email's recommendations. On 5.5, the entire chain executes cleanly. It reads across the report, identifies distinct risks, pulls and computes specific numbers, and drafts the email with the same numbers embedded, all in one response without me sending a single follow-up. GPT 5.5 thinking has a significantly larger context window than previous models, which means it can hold documents that are hundreds of pages long without losing track of details buried in the middle. So, the rule of thumb with 5.5 is stop breaking your prompts into small steps. Give it the whole document, describe the full task including what you want the final output to look like, and let the model figure out the steps on its own. That approach uses fewer total tokens than the four-message drip feed because the model doesn't need to reread context on each follow-up, and it produces more consistent output because everything stays in one pass.
I'll run one more prompt to show this on a data analysis task. I'll upload a spreadsheet and paste in this prompt. GPT 5.5 reads the full spreadsheet, runs the analysis, and produces the output I asked for in one response. The charts are clean, the numbers are accurate, and the summary highlights the specific patterns I asked about without me needing to point them out. OpenAI's own teams reported saving 5 to 10 hours per week by working with 5.5 this way, and for my testing that number feels about right for anyone who spends a lot of time on document or data-heavy work. And the same principle works for research prompts. Instead of asking one question and then following up with three more, I'll stack the entire research brief into a single prompt. The response that comes back is a structured report with cited sources organized by the categories I asked for with the level of depth that would normally take opening 15 tabs and reading through each source myself. The key with research prompts on 5.5 is to be specific about the output format you want because the model is good enough to match it. If you tell it you want a two-page brief with recommendations at the end, it structures the brief around those recommendations rather than just dumping information. This approach works for everything from competitive analysis to client research to market sizing, and the output is consistent enough that I've stopped doing the first pass of research manually on anything.
[music] The same single shot principle applies to image generation, but the workflow there has its own technique that's worth showing separately. OpenAI released ChatGPT Images 2.0 alongside GPT 5.5, and the quality jump is immediately visible, but the way you should be using it is different from how image generation worked before. So, I'll paste in an image prompt. The image that comes back looks commercial grade. The text renders correctly, the speaker is placed exactly where I described it to be, and the composition looks like something from a product shoot. But, the technique that saves the most time works significantly better on 5.5 because of how the new image model plans compositions before generating. On older models, if the image wasn't right, you'd usually rewrite the entire prompt from scratch and try again because the results were inconsistent enough that iterating on specific parts rarely worked well. On GPT 5.5, the image generation happens inside the same conversation, which means the model remembers exactly what it created and can iterate on specific parts of it. So, instead of rewriting the whole prompt, I'll type, "Move the speaker slightly to the left, make the lighting warmer, and add a subtle shadow underneath." It moved it to the left properly, and GPT 5.5 followed the edit while adjusting the existing image rather than generating a completely new [music] one from scratch. That conversation of generating, giving specific feedback, and watching the model refine the same image takes about 2 or 3 minutes compared to the 10 or 15 minutes of rewriting full prompts from scratch [music] that older models required. And because Images 2.0 has thinking built in, it plans the composition before generating, which means the first attempt is usually closer to what you wanted than anything previous image models produced on the first try. [music] I'll paste in one more prompt to show what the world building looks like. The level of detail in the output is on a different level from anything OpenAI's previous image models produced, and for anyone who creates content, builds presentations, or designs marketing materials, >> [music] >> this is the upgrade that saves the most time because the output is usable without needing to open image editing tools. This iteration workflow also works for mock-ups and design work. If you're on Claude Pro or Max, you have access to Claude Design for ideation and layout, and then you can bring those ideas into Chat GPT images 2.0 for the final polish, which gives you the best of both tools without paying for a separate design subscription.
The image generation is the flashiest upgrade in 5.5, but the one that saves me the most time every day doesn't produce anything visual. The last area where using GPT 5.5 properly makes a noticeable is in how you approach writing and research prompts, because the model handles both differently than 5.4 did. I'll paste in the same prompt twice, starting with the 5.4 result then the 5.5 result, so you can see the different. The 5.4 version still answers the prompt, but it feels noticeably weaker. The writing is safer, flatter, and more generic. It explains the same idea clearly, but it does not have the same point of view or rhythm, so it reads more like a clean AI summary than something a real founder would post. The 5.5 version feels much closer to the prompt. It sounds more like a founder making a clear argument, not just explaining a topic. The pacing has more bite, the examples feel more specific, and the phrasing feels sharper without becoming preachy.
The technique with writing on 5.5 is to be specific about what you don't want rather than just describing what you do want, because the model is now good enough to follow exclusion rules reliably. If you say don't use buzzwords, don't open with a question, don't start two paragraphs the same way, 5.5 actually follows those constraints where older models would ignore half of them by the third paragraph. And this is where memory makes a big difference because GPT 5.5 thinking supports Chat GPT's memory feature, which means you can save your writing preferences once and the model applies them to every future conversation without you needing to re-prompt. So, instead of pasting the same style instructions into every new chat, I can tell GPT 5.5 something like, "Whenever I ask you to write something, never use buzzwords, always vary sentence length, and match the tone of a founder rather than a marketer." And it remembers that across every conversation going forward. That one setup step means every writing prompt I send from now on already has my style preferences built in, which saves me from repeating the same instructions dozens of times per week and produces more consistent output because the model isn't starting from zero each time.
On the coding side, GPT 5.5 has also made significant improvements, particularly in agentic coding where the model writes tests and debugs across multiple files. OpenAI's benchmarks show it outperforming both 5.4 and Opus 4.7 on coding tasks [music] with faster completion times and fewer errors across the board. Where GPT 5.5 falls short compared to Claude is on the application side. Claude still has a more integrated experience where chat, projects, skills, code, and automation all live in one window. ChatGPT splits its best features across the main app and Codex. And for long-form writing where maintaining a specific voice across a long document matters, Opus 4.7 still follows custom instructions more reliably. But GPT 5.5 is faster, more token efficient, produces significantly better images, and handles multi-step tasks in a single prompt without losing context halfway through.
The honest answer is that if you're already on ChatGPT Plus, you should switch to Thinking immediately and start using the techniques from this video because the default 5.5 experience with the right setup outperforms what 5.4 delivered on maximum settings. If you're on Claude and you do mostly writing and project-based work, Opus 4.7 is still the better fit for that workflow. And if you use AI heavily enough that both matter, $37 a month for both subscriptions is the setup that gets you the best of each. And if you're trying to figure out which AI tool is the best fit for how you work, I tested Gemini, Claude, and ChatGPT head-to-head in this next video right here. Thank you for watching and I'll see you in the next one.