{"slug": "what-actually-mattered-sep-21-24", "title": "What actually mattered: Sep 21–24", "summary": "Anthropic released Claude Opus 5.5 on Tuesday and OpenAI released two new models, GPT-6 Sol and the cheaper Luna, with Anthropic claiming Opus 5.5 costs less to run than Opus 5. On Wednesday, Anthropic said roughly 950 Claude agents searched DNA databases for 21 hours, gathered more than 200,000 enzyme examples, narrowed them to 20 candidates, and found one enzyme system never previously described, which its scientists confirmed in the lab. Australia's prime minister said Thursday that an OpenAI research agent, asked in June to research medicine spending, broke into a Services Australia portal it was not allowed to access, and OpenAI did not tell the government until September 10.", "body_md": "Big week. Three things matter for you, and the first two are good news.\n\n## Big story 1: AI just got cheaper. Here's what that opens up.\n\n**On Tuesday, Anthropic released [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5), and OpenAI released [two new models](https://openai.com/index/introducing-gpt-6-sol-and-luna/) of its own.**\n\nThe best AI models are very expensive to run. That's why a lot of jobs never got handed to AI. This week, both Anthropic and OpenAI released new models that cost far less to use than the ones they replace.\n\nIn particular, I've been incredibly impressed with Claude Opus 5.5. In my initial tests, it's at least on par with Claude Fable 5.1, maybe even better, and it's dramatically cheaper and faster, which means you can use it on more things without paying more. Anthropic says it costs less to run than Opus 5, and that feels spot on from my early use.\n\nOpenAI's new models, Sol and the cheaper Luna, are a different story for me. I got early access and have been testing them for a while, and honestly, I didn't have a great experience. I stopped using them partway through my testing period. They're in ChatGPT Work and Codex (OpenAI's tools for work and for coding) [on most paid plans](https://openai.com/index/introducing-gpt-6-sol-and-luna/) if you want to try them yourself, but Opus 5.5 is the one I'd point you to.\n\n**What to do:** If you're on a paid Claude plan, open Claude and make sure Opus 5.5 is the model selected. You'll notice a nice step up in intelligence, and [on most paid plans](https://www.anthropic.com/claude-opus-5-5) you can now do more in a five-hour stretch before you hit your limit. Then give it a job you'd have thought was too big for AI. My full review of Opus 5.5 is coming soon.\n\n## Big story 2: AI agents just discovered something new in biology\n\n**On Wednesday, Anthropic said a swarm of Claude agents [discovered a new enzyme system](https://www.anthropic.com/news/claude-discovers-novel-enzyme-system), and its scientists confirmed it in the lab.**\n\nHere's what happened, in plain terms. Anthropic's scientists gave about 950 Claude agents one prompt and let them search huge DNA databases for 21 hours, with the goal of finding new versions of one particular enzyme (a molecular tool that copies genetic code). The agents gathered more than 200,000 examples of it and narrowed them down to 20 worth testing. One of those turned out to be a kind of enzyme system nobody had described before. The scientists ran the experiments and confirmed it was real.\n\nNobody knows yet what this new system does. But it has the same set of features as CRISPR, the tool behind gene editing, and a handful of other systems that can cut, copy and paste DNA. Finds like that are where gene editing, and a lot of modern medicine, came from.\n\nNow zoom out. Finding things like this is the slowest part of biology: years of careful searching by small teams. If this works at scale, that search runs around the clock, on every disease we care about. For now, people decide what's worth testing and check the results. I don't think that lasts long. Pretty soon the AI will decide what's worth testing too. That's what I meant on Wednesday when I said AI going well looks like millions of tireless researchers. This is the first small version of it.\n\n**Why it matters:** This is what AI in science is going to look like. Thousands of tireless assistants doing the searching that used to eat up years of a team's time.\n\n## Also worth knowing\n\n**An AI agent broke into a government system.** Australia's prime minister said Thursday that an OpenAI research agent, asked in June to research medicine spending, ended up breaking into a Services Australia portal it wasn't allowed in. No personal data is believed to have been taken so far, but OpenAI didn't tell the government until September 10.\n\n**Why it matters:** if your team runs AI agents on the open web, you own what they touch. Ask what your agents are allowed to touch, and what stops them from doing something harmful.\n\n**Amazon sellers can now run their store from inside Claude.** Amazon [launched a plugin](https://www.aboutamazon.com/news/innovation-at-amazon/seller-assistant-plugin-amazon-quick-claude) that connects a seller's listings, inventory and sales data to Claude (and to Amazon's own assistant), which can then suggest or carry out actions like adjusting prices or refreshing listings, with the seller approving them before they happen.\n\n**Why it matters:** a big platform just opened its back office to an outside AI assistant. Expect more of your tools to do the same.\n\n## Quick hits\n\nGoogle's new [Gemini voice models](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/) can copy a voice from 30 seconds of audio. You might want to agree on a way for your family, colleagues and friends to confirm that it's really you on a call.\n\nChatGPT's voice mode can now [use your connected apps and plugins](https://help.openai.com/en/articles/6825453-chatgpt-release-notes) mid-conversation, and if you have ChatGPT Work, start tasks like documents or spreadsheets by voice.\n\nIf there's a topic you want me to cover, or a question you've got, hit reply. I can't cover everything or reply to everyone, but I read every single one and answer as many as I can.\n\n— Matt", "url": "https://wpnews.pro/news/what-actually-mattered-sep-21-24", "canonical_source": "https://somethingbig.ai/what-actually-mattered-sep-21-24", "published_at": "2026-09-25 00:11:39+00:00", "updated_at": "2026-09-25 00:31:06.220975+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-research", "ai-safety"], "entities": ["Anthropic", "Claude Opus 5.5", "OpenAI", "GPT-6 Sol", "Luna", "Claude", "Services Australia", "Australia"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/what-actually-mattered-sep-21-24", "markdown": "https://wpnews.pro/news/what-actually-mattered-sep-21-24.md", "text": "https://wpnews.pro/news/what-actually-mattered-sep-21-24.txt", "jsonld": "https://wpnews.pro/news/what-actually-mattered-sep-21-24.jsonld"}}