😺 Everything to know about GPT-5

Welcome, humans.

Thanks to everyone who joined us live yesterday as we wrangled our way through our first ever LIVE Youtube video. OpenAI CEO Sam Altman tried his own hand at live-tweeting through the launch, and found out mid-stream when the internet roasted his charts. This graph sums up the feeling people had:

Did we just move from “benchmark” maxing to “benchmark chart maxing”? lol.

On our side, we were pretty impressed by how, after we got access, we really didn’t have any problems with latency or getting our prompts through (no doubt thanks to the murderers’ row of data center providers backing the launch). Sam said they were cranking out 2B tokens per minute at one point, and gave a much deserved shout out to the eng and infra teams for pulling that off (even if he loves the smell of burning silicon in the morning)…

Also, it looks like Google and OpenAI worked together to make this launch a success… which means whatever new model Google is cooking up will probably be coming next week at earliest.

The ultimate winners from today’s launch? The ~1K OpenAI staff members who got anywhere from ~$200K to $1.5M in bonuses for sticking around and not bailing to Meta (or others). OpenAI will also soon let them cash out their shares (if interested) in a share sale that values the company at $500B before any potential IPO.

Well hey, good for them. As Kyla writes, the U.S. economy is basically just AI, healthcare, and Labubu’s at this point…

Here’s what happened in AI today:

• OpenAI finally released ChatGPT-5…here’s what you need to know.
• xAI will put ads inside Grok responses.
• n8n could be about to raise a new round at $2.3B.
• Meta acquired an AI startup that makes emotional voice AI.

Advertise in The Neuron here

GPT-5 is here… here’s everything you need to know (so far…).

DEEP DIVE: Everything you need to know about GPT-5.

So, we wrote a deep dive with every interesting link and demo and fact we could sniff up about GPT-5, which you can read via the above link. Below, we had GPT-5 summarize it “with 100% fidelity” (our favorite trick to get AI to not skip anything when summarizing) and pasted it below with no edits. We recommend you read both versions to see how GPT-5 did!

OpenAI has launched GPT-5, its most capable model yet, now in ChatGPT and via API, with faster performance, stronger reasoning, major safety upgrades—and, per Sam Altman, still short of AGI. It unifies past models (4o, o3) into a single system that automatically switches between a fast “gpt-5-main” and a powerful “gpt-5-thinking” model, guided by a smart router. This means users get quick answers for simple queries and deep analysis for complex ones—without toggling settings.

Performance gains:
• Coding: Sets a new benchmark with 74.9% on SWE-bench Verified, excels in front-end design, and reduces code-editing errors by a third.
• Health: On the new HealthBench, built with 250+ physicians, GPT-5 scored 46.2% vs GPT-4o’s 0%, acting as an “active thought partner” for medical queries.
• Writing: Produces more emotionally resonant, constraint-driven creative work.
• Truthfulness: 45% fewer factual errors than GPT-4o, better at admitting limits.

Notable demos: Integrated Gmail and Google Calendar access lets it plan events, draft replies, and generate packing lists. In live tests, it built full apps—like a French-learning site and financial dashboards—in minutes, autonomously scaffolding code, fixing bugs, and deploying. The GPT site showcases playable demos from games to design tools.

Developer features:
• New parameters like reasoning_effort and verbosity for finer control.
• Custom tools definable beyond JSON.
• Massive context windows—272K input, 128K output tokens.
• API pricing: $1.25 input/$10 output for full model, with cheaper mini and nano tiers.

Safety: Multi-layered system cuts hallucinations, mitigates deception, and uses “safe completions” for sensitive topics. High-risk domains trigger constant monitoring, built with global AI safety institutes.

Prompting tips: OpenAI advises using GPT-5 as a meta-prompter to improve your own prompts, breaking complex tasks into turns, and removing contradictory instructions.

Why it matters: GPT-5 embodies a “team of specialists” approach—fast small models for most tasks, powerful ones for hard problems—reflecting NVIDIA’s “heterogeneous agentic system” vision. This could evolve into orchestration across dozens of specialized models, mirroring human collective intelligence.

Bottom line: GPT-5 isn’t AGI, but it’s a leap in usability, reliability, and breadth—pushing ChatGPT toward being a truly personal, expert assistant.

Now go skim the full deep dive and compare the two drafts to see how GPT-5 did!

FROM OUR PARTNERS

Your keyboard called—it’s taking a long weekend ⌨️✈️

We’ve been banging away on keyboards for 150 years. Until now, voice dictation hasn’t been reliable enough to change that.

Wispr Flow finally delivers the no-edit confidence we’ve all been waiting for:

• 4× quicker than typing. Dictate emails, docs, and DMs in real time and save precious hours every week.
• AI auto-edits on the fly. Flow cleans filler words, fixes grammar, and formats perfectly as you speak.
• Works inside every app with no setup. Fly through Slack notifications, give more context to ChatGPT, or brain dump into Notion.
• Use it at your desk or on the go. Available on Mac, Windows, and iPhone.

“This is the best AI product I’ve used since ChatGPT.”
— Rahul Vohra, CEO, Superhuman

Give your hands a break ➜ start flowing for free today.

Prompt Tip of the Day.

As part of the GPT-5 release, OpenAI also released a giant guide to help you learn how to use it (as pointed out by Elvis Saravia).

For example, they released this prompting guide along with a prompt optimizer tool you can use to run your prompts through.

The TL;DR = OpenAI’s team found that early testers had huge success using GPT-5 as a “meta-prompter” …feeding it unsuccessful prompts and asking what specific phrases to add or remove to get better results.

Here's their template:

Several users discovered contradictions and ambiguities in their core prompt libraries just by running this exercise. Removing those conflicts “drastically streamlined and improved their GPT-5 performance.”

Treats to Try.

*Asterisk = from our partners. Advertise in The Neuron here.

1. *Spinach AI records, transcribes & summarizes your meetings, then automatically updates your CRM and project tools.
2. Rork turns your app ideas into working iOS and Android apps, and now has access to GPT-5
3. OpenAI's Codex CLI just shipped major upgrades today, including GPT-5 integration and free usage with your ChatGPT plan, so you can code from the command line without needing an API key (Github).
4. MagicPath turns your design sketches into working code on an infinite canvas—free to try (from Pietro Schirano, featured in this GPT-5 video!).
5. Cursor lets you code with AI in both your editor and your terminal now, so you can seamlessly switch between the two.
6. Local Browser use is a a 600-line Python script (built in ~3 hours at a Hackathon) that automates your browser completely offline.
7. Decart makes the genAI video tool Mirage, which warps your live video in real-time and generates interactive worlds (raised $100M).
8. Dreamcore creates mobile games from text prompts that you can share in a vertical feed for others to discover and play—free to try.
9. Anything turns your app idea into a working iOS/Android app in minutes - just describe it and deploy.
10. GPT-5… go try it out. No excuses!

Around the Horn.

What you’re looking at here is the accuracy of AI models at completing software engineering tasks vs the time it takes a human to do the same task… GPT-5 is now at over 2 hours.

• n8n is supposedly about to raise “hundreds of millions” in a new round that could value it at $2.3B (up from $350M four months ago).
• xAI will soon put ads into Grok’s responses, which will let marketers pay to appear in Grok’s suggestions to help solve a user’s problem.
• Meta acquired WaveForms AI, who create a series of audio AI models that can recreate (and respond to) human emotions.
• Oh, and apparently Meta’s poached researchers are now working as TBD Lab to develop the next generation of Llama (and other projects).

Intelligent Insights

1. Are standalone AI coding startups like Cursor and Windsurf money-losing businesses? Ed Zitron certainly thinks so, and sees startups like Cursor as a systemic risk to the AI industry.
2. Remember how Google indexed 4K chats that were “shared” publicly? It turns out the number was more like 96K (130K including Grok and Claude chats), and of those, the WSJ analyzed “at least dozens” of long chats where ChatGPT made delusional claims.
3. Chris Olah of Anthropic wrote about how the tools scientists use to understand how AI systems work can learn shortcuts and memorization tricks instead of copying the AI's actual problem-solving methods, potentially giving researchers completely false explanations about how AI really operates—though his “Jacobian matching” technique can catch these deceptive interpretations by forcing the tools to match the mathematical fingerprint of the original AI's computation, building on recent advances in attribution graphs and attention mechanisms that show the potential of these interpretability methods.

A Cat’s Commentary.