Research Period: 28 September to 5 October 2026
1. Claude Sonnet 5.5: faster and more capable, same price
What changed
Anthropic released Claude Sonnet 5.5 on 28 September. It’s more than 30% faster than the previous Sonnet and scores 70.6% on Terminal-Bench 4.0, a test of how well a model handles long, multi-step tasks on a computer. The price hasn’t moved: $2 per million input tokens and $10 per million output. It’s available in Claude Pro, Max and Teams, through the API, and on AWS Bedrock, Google Cloud Vertex and Azure.
Why it matters
Sonnet is the model most people use day to day, so this lands on work you’re already doing. In the app, you don’t need to do anything. On the API, it’s a drop-in replacement: change the model name and get faster, more capable output for the same bill.
Use cases
- Swap the model name in existing API automations and compare speed over a week
- Retry a multi-step task that stalled on the old Sonnet, such as reconciling two spreadsheets
- Use the faster replies for customer drafts where waiting slows your team
- Hand longer data-cleaning jobs to Claude Code and see how far it gets unaided
- Test it on your highest-volume prompt to measure the speed gain
Sources: Anthropic | MarkTechPost
2. GPT-6.1 Sol: near-Astra quality at a fifth of the price
What changed
At DevDay on 29 September, OpenAI released GPT-6.1 Sol, a week after we covered GPT-6 Sol. On OpenAI’s figures it matches GPT-6 Astra, the flagship, on the DeepSWE coding benchmark at roughly one-fifth of the cost, and comes within about two points of Astra on computer-use tasks at around a seventh of the cost per task. The price is unchanged from GPT-6 Sol at $2 per million input tokens and $10 per million output, against $10 and $50 for Astra. It can take in about 1.05 million tokens at once, though anything over 272,000 tokens is charged at a higher rate for the whole request. It’s available in ChatGPT Work, Codex and the API.
Why it matters
Last week the story was cheaper models. This week the cheaper model got a lot better without the price moving. For most routine work, including drafting, analysis and multi-step tasks that run on their own, you no longer need to pay flagship prices to get close to flagship results. Keep an eye on the long-document surcharge: feeding it a very large pack of files costs noticeably more per request.
Use cases
- Move API automations from Astra to GPT-6.1 Sol and compare a month’s bill
- Use it in Codex for coding jobs where Astra’s price was hard to justify
- Run one real task through GPT-6.1 Sol and Claude Sonnet 5.5 side by side before choosing
- Keep prompts under 272,000 tokens where you can to avoid the surcharge
- Revisit any agent workflow you parked because the flagship model made it too expensive
3. Microsoft 365 Copilot: MCP Apps and pay-as-you-go billing
What changed
Following last week’s Home, Code and Autopilot announcement, Microsoft has opened a preview of MCP Apps, bringing rich, interactive agent screens inside Microsoft 365. Home pulls Word, Excel and PowerPoint into one starting point, and Code builds simple apps and dashboards from a plain-English description. There’s also new usage-based billing (Credits) alongside the flat rate, so you pay per use for the most advanced models. It’s available to Microsoft 365 Copilot subscribers, including in the UK and Europe.
Why it matters
If your business runs on Microsoft 365, agents can now work inside the place you already work rather than another tab. The billing change helps smaller teams too: pay for heavy use only when you need it.
Use cases
- Check your Microsoft 365 admin centre for the MCP Apps preview
- Ask Code to build a sales dashboard from an existing Excel tracker
- Start a client proposal in Home, with Word and PowerPoint in one place
- Compare Credits with flat-rate pricing for occasional users of advanced models
- List the two or three tools you’d most like working inside Microsoft 365
Sources: Microsoft Blog | Microsoft TechCommunity
Additional Noteworthy
Skills in Gemini: Google is replacing Gems with Skills, reusable instructions you call with a forward slash and can stack, such as a writing-style skill plus a brand-guidelines skill. They use the same open SKILL.md format as Claude’s skills, so a skill written for one should copy across. Rolling out to Google Workspace business and enterprise editions from 5 October and to the Gemini app from 13 October, both complete by mid-November; Gems keep working for business accounts until at least March 2027. | Google Workspace Updates | Google
Claude Code Mods: TypeScript plugins that change how Claude Code behaves, such as blocking unsafe commands or redacting sensitive data; one for teams with a developer. | Anthropic
Gemini 4 Argon (watch): Google’s first Gemini 4 model, announced 30 September, leads GPT-6 Astra and Claude Opus 5.5 on several of Google’s own coding and knowledge-work benchmarks and can write up to 1 million tokens in one response. It is limited to cyber-defence partners for now, with paid API customers and Google AI Ultra subscribers next; no dates yet. Introductory API price $2/$10 per million tokens. | Google | MarkTechPost
OpenAI Dots (watch): Always-on ChatGPT agents with their own cloud computer and memory, announced at DevDay. Not available on the Pro plan in the UK, EEA or Switzerland, with no date given; Business Premium subscribers can use them in all supported regions. | OpenAI | Mixed News
Perplexity Business Skills (watch): Ready-made workflows for tax, cash flow and marketing, US-only for now via American Express; watch for a UK launch. | Unite.AI
Strategic Awareness
This week’s three reads belong together now that agents increasingly run on their own. Ethan Mollick finds that once several agents work together, your job moves from managing each step to setting the goal and checking the outcome. Christian Catalini in HBR goes further: if AI makes doing the work cheap, verifying it becomes the scarce skill, and businesses with a reliable checking routine and their own trustworthy data will pull ahead. Simon Willison adds the practical safeguard: put a hard spending cap on the account before any agent runs on its own, so a runaway task can’t produce a nasty bill. The common thread is to decide how you’ll check and limit the work before you hand it over.
Sources: One Useful Thing | Harvard Business Review | Simon Willison’s Weblog
Looking Ahead
Capable models keep getting cheaper, Google’s Gemini 4 generation is on its way, and always-on agents are arriving, though OpenAI’s Dots isn’t yet open to Pro users in the UK or Europe. Pick one routine task this month, write down how you’d check the result, set a spending cap, and try it.
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.