Meet GPT-5

ChatGPT goes to Washington

August 08, 2025

Welcome back! When school’s out, a few things change: The line is longer at the local ice cream shop, the parents go a little more stir crazy, and…ChatGPT usage absolutely tanks, as X users pointed out this week.

What’s your take—is the drop-off in usage a side effect of summer break, or did token processing decline because the competition got more intense? And with back to school around the corner, are we about to see another spike? Hit reply and share your theory. In the meantime, let’s dive into a slew of major OpenAI headlines.

OpenAI just dropped the model of the Summer

OpenAI releases its smartest, fastest, and most useful model yet, with thinking built in / Via OpenAI

OpenAI dropped GPT-5 yesterday, what it says is its best model yet. It’s designed to do way more than chat, and for the first time, it’s available to everyone (even free users). Let’s dive into the details. ⬇️ 

What’s new: 

  • GPT-5 is OpenAI’s first “unified” model, combining the speed of GPT-4o with the deep reasoning of its o-series models. 

  • It takes action on your behalf with stronger reasoning, fewer hallucinations, and more agent-like behavior. You get smart answers fast—but also a model that can build apps, generate research briefs, or navigate your calendar. 

  • The model includes a new routing system that decides in real time how best to respond, whether with speed or deeper thinking.

How it performs: 

  • On coding benchmarks, GPT-5 outpaces Claude Opus 4.1 and Gemini 2.5 Pro, spinning up entire software apps in one go. 

  • It’s also better at avoiding health misinformation, hallucinating just 1.6% of the time on medical questions—down from ~13–16% in earlier models.

Where it falls short: On simulated online tasks like navigating retail sites, it slightly trails Claude. And on tough academic tests like “Humanity’s Last Exam,” GPT-5 Pro performs well—but not top of class.

What it means for users: ChatGPT users now get access to GPT-5 by default, with Plus and Pro subscribers getting more usage and access to GPT-5 Pro. Developers can use three versions (nano, mini, and base), tweak response length, and pay $1.25 per million tokens (about 750K words, longer than the entire Lord of the Rings series).

Why it matters: OpenAI says this is a step closer to AGI—AI that can do most economically valuable work better than humans. But even Altman admits that GPT-5 is impressive, yet still imperfect. The real test is whether developers and users adopt it…and what OpenAI builds next.

US government gets ChatGPT Enterprise for $1 per agency

OpenAI just inked a major deal with the US government. For the next year, federal agencies will get access to ChatGPT Enterprise at just $1 per agency, thanks to a new partnership between OpenAI and the General Services Administration (GSA). It’s a big move that signals how seriously the US is leaning into AI across the government.

The goal: This partnership aims to save federal workers time on repetitive tasks like paperwork, research, and drafting emails so they can focus on more impactful work. In early pilots, employees in Pennsylvania saved over 90 minutes a day using ChatGPT. North Carolina’s state treasury department reported an 85% satisfaction rate in its own pilot.

Security comes first: Handling sensitive government data means the security bar is high. OpenAI says none of the data entered into ChatGPT Enterprise will be used to train its models. The GSA has formally approved the tool, and the deployment comes with full compliance and privacy guardrails.

Why it matters: By giving civil servants AI tools at scale, the US is showing what national-level AI adoption might look like—at least when it’s done securely. Other governments are likely watching closely. The question now: Will they follow suit?

Microsoft brings OpenAI’s open model to Windows devices

OpenAI keeps on stealing the show. Its lightweight open-weight AI model—gpt‑oss‑20b—is now available directly on Windows 11 courtesy of Microsoft’s Windows AI Foundry. This move lets users tap into AI locally without needing cloud access.

Built for agents, optimized for speed: Microsoft optimized gpt‑oss‑20b for tasks like code execution and tool use—even in low-bandwidth situations. All you need is a Windows PC or laptop with a modern GPU and at least 16GB of VRAM. 

Where else you’ll find it: Windows users get first access, but Microsoft promises macOS support soon. The model is also available via Azure AI Foundry and AWS, along with the more powerful gpt‑oss‑120b.

Why it matters: This is part of the growing shift toward on-device AI. Instead of relying on server-heavy infrastructure, companies are racing to bring capable models directly to laptops and desktops.

AI review that catches what others miss

Recurse

Recurse is an AI-powered code review assistant that scans your pull requests for bugs and breaking changes before you merge. It works across major languages with no extra setup.

How you can use it:

  • Get inline suggestions with fix ideas and documentation

  • Catch logic errors that slip past traditional tests

  • Maintain high code quality in fast-moving teams

  • Reduce time spent on manual code reviews

Pricing: $25 per month per user 

Translate your content for the world

Aview

Aview uses AI to dub, translate, and localize videos into 50+ languages—while keeping your voice and style intact.

How you can use it:

  • Expand your YouTube reach with culturally tuned AI voiceovers

  • Localize livestreams and courses without hiring a translator

  • Forecast performance by region with built-in analytics

Pricing: Free plan available; paid starts at $49 per month 

Jobs, announcements, and big ideas

  • Midjourney debuts HD video for Pro and Mega users.

  • Cursor 1.4 update improves AI performance on big codebases.

  • Google adds Guided Learning to Gemini for deeper understanding.

  • Google’s new agent Jules brings asynchronous coding to the public.

  • Anthropic rolls out Claude Code for automating AI security reviews.

  • Genie 3 turns videos into interactive experiences with AI.

  • Microsoft previews AI-first Windows with voice and sensory features by 2030.

GPT-5 is here. I break down what it can do, how it performs, and why it’s OpenAI’s most capable model yet.

That’s a wrap! Earlier this week, many of you weighed in on who you think will win the AI platform race in the long run. Google took the lead, followed by OpenAI. We’ll see if the next round of model releases shakes up those numbers.

—Matt (FutureTools.io)

P.S. This newsletter is 100% written by a human. Okay, maybe 96%.