GPT-5 Drops Today: We Re-Ran Our 9 Production Workflows on It (And One Already Got Cheaper)

OpenAI released GPT-5 today, August 7, 2025. Instead of writing a hot take, we did what we tell clients to do: we re-ran our 9 live production workflows on it the same afternoon and measured. One workflow — invoice-field extraction for a procurement client — got both cheaper and more accurate, so we switched it before dinner. Three others we moved. Four we deliberately kept on gpt-4o-mini. Here's the cost-per-task data and the decision behind each call. ## The 60-Word Answer GPT-5 launched today at $1.25 per 1M input tokens and $10 per 1M output, with a 400K context window. The cheaper tiers — gpt-5-mini ($0.25/$2) and gpt-5-nano ($0.05/$0.40) — matter more for production than the flagship. We moved 3 of our 9 workflows: one to gpt-5-mini for a real cost drop, two to GPT-5 for accuracy. We kept 4 on gpt-4o-mini where it was already good enough. ## Why This Matters Today GPT-5 shipped today across ChatGPT and the API. The headline that actually changes our cost math isn't the flagship — it's the family. GPT-5 comes in three API sizes: gpt-5, gpt-5-mini, and gpt-5-nano, all with a 400K-token context window (per OpenAI's pricing page). For us, the question on launch day is never "is it smarter" — it's "does it change the cost-per-task on a workflow we're already running, or make one we'd parked viable." So we measured. ## The Cost-Per-Task Chart We price our automations per task, not per token, because that's how clients think. Here's the per-task cost for four representative workflows, before and after the GPT-5 switch decision. The invoice-extraction workflow is the one that got cheaper — moving from gpt-4o-mini to gpt-5-mini cut errors enough that we removed a human review step, which is where the real saving came from. ## The 3 We Switched (and Why) ## The 4 We Kept on GPT-4o-Mini This is the unfashionable half of launch day. Four workflows stayed exactly where they were, and that was the right call. ## How We Re-Ran Everything in One Afternoon The reason we could measure on launch day, not launch month, is that every workflow is built to swap models with a one-line config change. If yours aren't, that's the real lesson here. ## The Wrong Reactions We're Seeing Today ## Our Take GPT-5 is a real step up on reasoning and long-context work, and the mini/nano tiers are where most teams will feel it. But launch day is an audit trigger, not a migration mandate. The teams that win today aren't the ones who switched fastest — they're the ones who could measure fastest, because their workflows treat the model as swappable config with a golden eval set behind it. This is the same discipline our founder Vivek Singh writes about on model churn: build for the model to change, because it will, every few months. Our AI automation team ships every client workflow with model-as-config and an eval set from day one — see how we re-ran our 9 most-used n8n workflows for 2025 and how we handled the Claude Haiku 4.5 launch the same way. Community pulse: the launch-day Hacker News thread split predictably between "it's a huge leap" and "incremental, mostly cheaper" — both true depending on whether your work is reasoning-heavy or routing-heavy. ## Frequently Asked Questions ### How much does GPT-5 cost via the API? At launch on August 7, 2025, GPT-5 is $1.25 per 1M input tokens and $10 per 1M output tokens. The cheaper tiers are gpt-5-mini ($0.25 / $2.00) and gpt-5-nano ($0.05 / $0.40). All three offer a 400K-token context window per OpenAI's pricing page. ### Should I switch all my workflows to GPT-5? No. Switch only where the new model removes a human review step, fits content that used to need chunking, or fixes a reliability bug. Classification, routing, and simple extraction tasks usually run fine and cheaper on a smaller model. Re-measure each workflow before deciding. ### What's the difference between gpt-5, gpt-5-mini, and gpt-5-nano? They trade capability for cost. gpt-5 is the most capable and most expensive; gpt-5-mini is a mid tier that often hits the sweet spot for production; gpt-5-nano is the cheapest for high-volume, simpler tasks. All share the 400K context window at launch. ### How do I test a new model on my own workflows quickly? Keep the model name in config, not hardcoded, and maintain a golden eval set of 20–50 known-correct input/output pairs per workflow. On launch day you change one config value, re-run the eval, and compare accuracy and cost against your recorded baseline. ### Did GPT-5 make any workflow cheaper for you? Yes — invoice-field extraction. Moving from gpt-4o-mini to gpt-5-mini raised field accuracy enough that we dropped the human spot-check on 70% of invoices. Net cost per task fell from ₹2.10 to ₹1.40, with the saving coming from removed labour rather than tokens. ### Is a working chatbot built on an older model now obsolete? Not because a new model launched. A bot is obsolete only when you can measure an accuracy or cost gap that matters to your users. Treat a launch as a reason to re-run your eval set, not as automatic proof your current build needs replacing. As Hrishikesh, our CTO, said when the model dropped: "Don't ask if it's better. Ask which of your bills it changes — and have the eval set to answer by tonight."