💸 GPT-6 got cheaper and Claude's limits went up
What last week's GPT-6 Sol and Luna launch means for your plan, plus more room on Claude Pro and Team.
Three-minute read. Today’s lead: what last week’s GPT-6 Sol and Luna launch means for the plan you already pay for.
PRICING
GPT-6 Sol and Luna reach ChatGPT Work at half the old API price
What changed. On 22 September OpenAI released GPT-6 Sol and GPT-6 Luna, two cheaper siblings of its top model, Astra. They run in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and Free and Go users can try Luna in the desktop app. OpenAI cut API prices by 50%: Sol now costs $2 per million input tokens and $10 per million output, and Luna costs $0.10 and $0.50. Make added both models on 23 September, in its AI Toolkit, its OpenAI app and its AI Agents.
What it means at your size. If you pay for Plus or Business, you already have both models in Work, with no plan change. If a Make scenario or other automation of yours calls OpenAI with your own API key, each run on these models costs half what the GPT-5.6 versions did. OpenAI pitches Luna for high-volume jobs such as sorting and routing requests, which is the kind of step most small-team automations run all day.
The first step. This week, open the one automation that calls OpenAI most often, switch its model to GPT-6 Luna, and compare ten outputs against last week’s before you leave it running.
The limitation. The launch is six days old, and the new models sit in Work and Codex only. OpenAI says they are not in regular Chat yet. The benchmark wins in the post are OpenAI’s own numbers.
OpenAI on GPT-6 Sol and Luna (openai.com)
CAPACITY
Claude raises five-hour usage limits on Pro, Max and Team
On 22 September Anthropic said it is raising the five-hour usage limits on Claude’s Pro, Max and Team plans. Subscribers also get one rate-limit reset they can save and spend when they choose. For a team of two or three on Pro or Team, that means fewer afternoons stopped by a limit, and one spare reset to keep for a deadline day. Anthropic did not publish the new numbers.
Claude on higher five-hour limits and a saveable reset (x.com)
The rest of the day
🪟 Microsoft rebuilt its Copilot app around Home, where chat and delegated work sit together with Word, Excel and PowerPoint built in; Code, which builds small apps and dashboards from a description; and Autopilot, an agent that keeps working on a goal you set. Most of it is in testing, Code reaches Microsoft 365 Premium and Pro subscribers in preview later this year, and Code and Autopilot bill by usage, so check the meter before you switch them on. (blogs.microsoft.com)
🔌 Google added 12 connected apps to Gemini on 23 September, including Airtable, monday.com, Zoho, PandaDoc, Squarespace, Webflow and Adobe. You connect them in Gemini settings or @mention them in a chat. Google did not say which plans get them. (blog.google)
🔐 ChatGPT now shows a Security history of sign-ins, sign-outs and security-setting changes, with time, location and device, under Settings > Security and login. Check yours this week. Pair it with Nate Herk’s rules for any agent you connect to an account: read access first, drafts before sends, and a human sign-off on money, deletions and publishing. OpenAI’s ChatGPT release notes on Security history and (youtube.com)
🎙️ Voice now works inside ChatGPT Work, where you can ask it out loud to create documents, presentations and spreadsheets, and an unfinished task carries on in text when you hang up. (help.openai.com)
🛒 Anthropic opened Claude Marketplace, where you can add connectors and plugins such as Slack and Notion, and buy agents from companies like Cursor and CrowdStrike. (x.com)
⚙️ Zapier’s connector lets Claude act across more than 9,000 apps, including Gmail, Google Sheets, Clio and Jobber. You choose what Claude can touch, and Zapier says your credentials stay out of the chat. (x.com)
Too early
Anthropic is working through the waitlist for Projects in Claude Code, where you describe a job once and Claude splits it into threads that run side by side while you are away. On 24 September it opened access to more Pro and Max users and added a way to run those threads on your own machine. Team plans come later, so most small teams can only join the waitlist for now. (x.com)
🪑 One for the road
Epoch AI built a test for the question every flat-pack owner has asked: can AI tell you built the IKEA furniture wrong? Models get the manual and a photo of a half-built piece and have to spot the mistake. Ten months ago the best one caught 28%. Now it catches 80%. (x.com)