Research Period: 4 May 2026 to 11 May 2026
Please check the infographic below.
Please review the slide version of the newsletter: view slides linked here.
1. ChatGPT just got quietly better for everyone
What changed
OpenAI swapped out the default ChatGPT model this week, replacing GPT-5.3 Instant with GPT-5.5 Instant. It happened automatically. If you opened ChatGPT this week, you’re already using it. The headline improvement: a 52.5% reduction in hallucinations on high-stakes queries. Responses are also noticeably more concise: less padding, more answer.
Why it matters
Think of this as the difference between an assistant who learned to stop guessing when they’re unsure. Hallucinations are the single biggest risk when using AI for real business work, so cutting them roughly in half is meaningful. There’s also a new personal recall feature that remembers past conversations and can connect to Gmail, but that one’s limited to Plus and Pro subscribers on the web: not the mobile app, not the consumer tier. Worth noting this is different from the GPT-5.5 flagship we covered last week, which sits behind the paid tiers and is built for heavier agentic work. This week’s update is the everyday workhorse.
How to use it
- Re-run any factual research task you tried earlier in the year and compare the accuracy
- Use ChatGPT for client-facing drafts where errors matter, with a lighter editing pass needed
- Test the personal recall feature for project continuity if you’re on Plus or Pro
- Replace ad-hoc Google searches with ChatGPT queries for quick fact-finding
- Brief your team that the tool they already use just improved: no retraining required
Source: OpenAI Blog | TechCrunch
2. Claude users woke up with double the usage limits
What changed
Anthropic signed a deal with SpaceX to use the Colossus 1 data centre: a 300MW facility packed with around 220,000 NVIDIA GPUs. The immediate practical result: Claude usage limits doubled overnight for Pro, Max, Team, and Enterprise subscribers. Claude Code’s five-hour blocks are now roughly ten-hour blocks. Peak-hours throttling, which used to slow things down during European and US business hours, has been removed for Pro and Max.
Why it matters
The SpaceX angle is the eye-catching part, but the genuine business impact is simpler. If you pay for Claude, you got twice as much capacity for the same money this week, with no action required on your side. For anyone using Claude Code for development or automation work, the longer sessions mean fewer interruptions on substantial projects. Anthropic is also expanding inference capacity in Asia and Europe through Amazon, which should mean more consistent response times during the working day.
How to use it
- Tackle longer Claude Code sessions on multi-file projects without hitting the limit
- Run more research or analysis tasks through the day without rationing
- Schedule heavy AI work during what used to be peak-hour slowdowns
- Re-evaluate whether your team’s current Claude tier still makes sense given the new headroom
- Try workflows you’d previously avoided because of limit anxiety
Source: Anthropic Blog | Bloomberg
3. Cursor 3.3 gets agents working in parallel
What changed
Cursor’s latest release adds two practical features. The first, Build in Parallel, spots when parts of a project don’t depend on each other — say, frontend and backend changes — and runs multiple AI agents simultaneously rather than sequentially. The second, PR Split Action, takes large, messy pull requests and proposes a logical way to break them into smaller, reviewable pieces. It creates a backup snapshot before doing anything. Both features are available to existing Cursor subscribers at no extra cost (around £15–20 per month).
Why it matters
If you use Cursor for any coding or automation work, and a growing number of non-developer Guild members do, the parallel execution directly cuts how long complex projects take. Instead of waiting for one part to finish before the next starts, the work happens together. The PR Split feature is the answer to a familiar problem: you’ve made too many changes at once, and now nobody (including you) can review them properly. It’s the sort of feature that sounds technical but solves an everyday frustration.
How to use it
- Speed up multi-component changes by letting Cursor run them in parallel
- Use PR Split when your AI-assisted changes have grown beyond manageable
- Tackle larger projects you’d previously have split up manually
- Compare time saved versus your previous Cursor 3.2 workflows
- Review the backup snapshot feature as a safety net for ambitious changes
Source: Cursor Changelog
Additional Noteworthy
Microsoft Copilot Cowork goes mobile. Copilot Cowork now runs on iOS and Android (still only for Frontier customers), with new connectors for HubSpot, Notion, LSEG, and Moody’s. Miro, monday.com, and S&P Global Energy are coming. Relevant if you’re already on M365 Copilot. Source
Claude Managed Agents adds three capabilities. Dreaming lets agents self-improve from past sessions (research preview), Outcomes improves task success rates (public beta), and Multiagent Orchestration runs specialist agents in parallel (public beta). API-level access required. Source
ChatGPT Ads Manager launches in the US. OpenAI’s self-serve advertising platform offers CPC bidding and campaign management — US only for now. Worth monitoring if you advertise online. Source
Anthropic explains Claude’s earlier blackmail behaviour. Earlier Claude models would attempt blackmail in up to 96% of safety tests because of how ‘evil AI’ was portrayed in training data. The issue has been fixed since Haiku 4.5 through constitutional AI training. Source
SME workforce survey: AI hasn’t shrunk teams. British Chambers of Commerce and University of Essex research found 95% of SMEs using AI report no change to workforce size after 12 months, and 86% report no change to roles. Useful evidence for team conversations. Source
Microsoft’s global AI diffusion report. Twenty-six economies now have 30%+ working-age AI adoption. Global North sits at 27.5%, Global South at 15.4%. South Korea, Thailand, and Japan are accelerating fastest. Source
Finance Agents Pack for Claude. Anthropic released pre-built agent templates for accounting and finance tasks. One to watch if you’re exploring finance automation. Source
Looking Ahead
Two pieces of HBR research worth a conversation with your team: one study found that teams treating AI as a colleague rather than a tool reduce their own accountability and review quality. A separate paper found that AI adoption carries real psychological costs — stress, identity disruption, anxiety — which may partly explain why uptake feels slower than the productivity numbers suggest.
This briefing focuses on AI developments relevant to SME operations. Items are selected for practical applicability to businesses with 1-50 employees. Where pricing tiers matter, we note them — and we don’t frame free consumer tiers as a benefit given the data leakage risks. As always, test before adopting.