Research Period: 21 September to 28 September 2026
1. Claude Marketplace: one catalogue for the tools Claude can work with
What changed
This week Anthropic opened the Claude Marketplace, a single catalogue inside Claude with three sections: connectors and plugins, ready-made agents and products, and service partners who can help you set things up. It launched with more than 2,000 connectors and plugins, including Atlassian, Google, Microsoft, Notion and Salesforce. Larger customers with a committed spend agreement can also put part of that budget towards Claude-powered software from companies such as Cursor, Harvey and Snowflake.
Why it matters
Until now, finding out whether Claude could work with your other software meant hunting through documentation or asking someone technical. The Marketplace puts it all in one place, a bit like an app store for Claude. For a small business the connectors are the part to look at first: if the tools you use every day are listed, Claude can read from and work with them rather than you copying and pasting between windows. The service partner listings are also worth a glance if you want help but don’t have anyone in-house to do the setup.
Use cases
- Check whether your accounting, CRM and project tools are listed before your next software renewal
- Connect Google Drive so Claude can summarise last quarter’s proposals without you copying anything across
- Link Notion or Atlassian and ask Claude to pull this week’s project updates into one note
- Connect Salesforce and ask for a list of deals that haven’t been touched in a fortnight
- Browse the service partners if you want a hand wiring Claude into a specific workflow
Sources: BleepingComputer | gHacks
2. Claude Opus 5.5: top-tier quality for around 40% less
What changed
Anthropic released Claude Opus 5.5 on 22 September. Anthropic reports it matches or beats Fable 5.1 on agentic coding and knowledge-work benchmarks. It costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, and Anthropic estimates typical workloads will cost around 40% less once cheaper caching and faster output are counted. It’s available through the API and on AWS, Google Cloud and Microsoft Azure.
Why it matters
For anyone paying for Claude through the API, this is a price cut on work you’re already doing. Opus 5.5 is a drop-in swap for Opus 5: the same quality of thinking or better, and on Anthropic’s own estimate a job that cost £10 last month should now cost roughly £6. Check it against your own bill rather than taking the estimate on trust. To put a million tokens in context, that’s around eight or nine full-length novels of text. If you’ve held back on automating a long document process because of running costs, it’s worth doing the sums again.
Use cases
- Swap Opus 5 for Opus 5.5 in existing automations and compare next month’s API bill
- Rerun cost estimates for batch jobs such as reviewing a year of supplier contracts
- Use it for long reports where accuracy matters more than an instant answer
- Hand coding tasks in Claude Code to a model priced below the previous flagship
- Revisit shelved automation ideas that looked too expensive to run every week
Sources: Anthropic | MacRumors
3. GPT-6 Sol and Luna: Astra-class quality at half the price
What changed
On 22 September OpenAI released GPT-6 Sol and GPT-6 Luna, two cheaper models that sit beneath GPT-6 Astra, the flagship launched earlier in the month. Sol is for everyday and complex work such as coding; Luna is for fast, high-volume jobs like summarising documents and pulling out information. OpenAI says Sol makes about half as many mistakes as its predecessor. API prices have halved: Sol costs $2 per million input tokens and $10 per million output, and Luna $0.10 and $0.50. Both are rolling out in ChatGPT Work and Codex for paid plans, and Luna is also available to users on the desktop app.
Why it matters
This is the second big price cut of the week, landing 90 minutes after Anthropic’s Opus 5.5. For small businesses the practical effect is that capable AI is getting cheaper to run on routine work. Luna in particular is priced for jobs you’d run hundreds of times a month, such as tagging enquiries or summarising call notes, where the running cost used to be the thing that stopped people automating.
Use cases
- Use Luna for routine drafting and triage where speed matters more than depth
- Rerun the sums on any automation you shelved because the per-task cost looked too high
- Compare Sol against Claude Opus 5.5 on one real task before choosing where to spend
- Ask Sol to check a contract or proposal for gaps now that it makes fewer errors
- Try Luna on the desktop app before paying for anything
Sources: OpenAI | TechCrunch | MacRumors
Additional Noteworthy
ChatGPT Live voice gains plugin support: You can now query connected apps such as calendars and CRMs mid-conversation by voice on web, iOS and Android, though files and screen sharing aren’t supported yet. | Unite.AI
Microsoft rebuilds Copilot around Home, Code and Autopilot: Home brings Chat, Cowork and Word, Excel and PowerPoint into one place; Code lets non-developers build dashboards and small internal apps by describing them; Autopilot is an always-on agent that watches channels and runs recurring work. Home and Code reach the Frontier programme first and Autopilot enters private preview, with agent work billed by usage. | Microsoft Blog
Claude Tag personal connectors: Teams plan users can now bring their own connected tools, such as a personal calendar, into shared channels without exposing them to colleagues. | Claude Blog
Perplexity Computer adds Effort Mode and a Skills Marketplace: Pro and Max users can choose from four effort levels, Light to Ultra, to match compute to the task, and install reusable workflows; sensitive tasks can run locally. The changelog blocks automated checks, so verify details on the page. | Perplexity Changelog
Strategic Awareness
Simon Willison’s year-to-date review is the best single read this week if you’re wondering what has actually changed in 2026. His central argument is that coding agents crossed from ‘often makes mistakes’ to reliable daily tools this year. He’s candid about the costs too: security incidents where agents caused real damage, and a sense of ‘AI-induced ennui’ among engineers whose routine work has been handed over. His conclusion is the one to hold onto: judgement, taste and direction remain firmly human jobs. Meanwhile, Google DeepMind has confirmed Gemini 4 is in post-training, the final tuning and safety stage, and hopes to release it much earlier than the end of 2026. No features or dates yet, so there’s nothing to act on, but expect another round of model comparisons before Christmas.
Sources: 2026 in LLMs so far (Simon Willison) | 9to5Google
Looking Ahead
Connecting AI to your existing tools is becoming a two-minute job rather than a project. The next few weeks are a good time to pick one workflow and wire it up properly, before Gemini 4 reshapes the options again.
This briefing focuses on AI developments relevant to SME operations, emphasising practical applications in productivity, cost efficiency, and business capability. For questions or to suggest sources reach out on the group WhatsApp.