Four things happened in AI over the last week that change what you can actually do at work. Not the funding rounds and not the executive drama: the releases that alter the tools on your desk and the argument you can make in your own performance review.
Here is what changed, what it means for your job, and what I would do about each one.
| What changed | A new frontier model, a protocol rewrite, a free near-frontier model, and the first solid data on AI and hiring |
|---|---|
| Who it affects | Anyone using AI at work, anyone building internal tools, and anyone deciding which employer to bet on |
| Time to act | This week for the model switch, this month for the protocol change |
| Cost to you | Nothing. Every change below lowers price or effort rather than raising it |
Frontier capability got cheaper, again
Anthropic released Claude Opus 5 on 24 July, doubling the previous model’s coding benchmark score at the same price, and added an effort dial that lets you trade cost against quality on a per-task basis.
The dial is the part worth your attention. Most people run every task at maximum quality, which is like couriering a birthday card. Cheap effort for drafting and sorting, high effort for the reasoning that actually matters, and your cost per useful output drops sharply.
Small and midsize businesses using AI report roughly four times more hiring than layoffs, according to Intuit’s 2026 AI Impact Report.
The jobs data finally has a shape
The most useful number this week was not about models at all. Intuit’s 2026 AI Impact Report found that small and midsize businesses using AI report about four times more hiring than layoffs, with 78% saying it lifted productivity.
Be careful how you read that. It is self-reported, it covers small and midsize businesses rather than large enterprises, and companies that adopt AI early may simply be the ones already growing. It is not proof that AI creates jobs. What it does undercut is the flat assumption that adopting AI means cutting headcount, at least at this size of company.
MCP went stateless, and some things will break
The Model Context Protocol, the standard that lets an assistant reach your real files and tools, shipped its largest spec update yet on 28 July, moving to a stateless design.
Two consequences. If your company runs in-house MCP servers, older ones may stop working with newer clients, so somebody needs to own that upgrade. And deploying agent tooling gets meaningfully simpler, which lowers the barrier for you to propose it. If you have been waiting for the moment to suggest connecting an assistant to a real internal system, the plumbing just got easier to defend.
Near-frontier models are now free to self-host
Moonshot released open weights for Kimi K3, a 2.8 trillion parameter model that performs close to frontier systems while being substantially easier to run, according to Tom’s Hardware.
This matters most in regulated industries. I have spent fifteen years in cybersecurity, and the single most common reason a good AI idea dies inside a bank, a hospital or an insurer is that the data cannot leave the building. That objection is now weaker. “We cannot send this to an API” no longer means “we cannot use AI”.
Prompts you can use
Paste these straight in. Change the parts in square brackets and nothing else.
Help me decide which of my recurring tasks deserve expensive AI reasoning and which do not. My recurring tasks: [list 8 to 10 things you use AI for in a normal week] For each one, tell me: - Cheap or expensive setting, and why in one line - What would actually go wrong if I used the cheap setting - Whether the task should be handed over entirely rather than assisted Then give me the one task where spending more would change my output the most.
This is the judgement that is now worth more than prompt-writing skill. Redo it every few months as your work changes.
You are helping me write a short internal note about a change in AI tooling. The change: [paste the announcement or link] My audience: [engineers / marketers / leadership / mixed] What I want them to do about it: [nothing yet / evaluate / act this month] Write 150 words maximum. Lead with what changes for them, not what the vendor announced. Include one sentence on what could go wrong. No hype, no adjectives doing the work of evidence. If the announcement is a vendor claim rather than verified, say so.
Change the audience line and the same prompt produces a very different, and better, note.
Takeaways
- Stop running every task at maximum quality. Use the cheap setting for drafting and sorting, the expensive one for judgement.
- If your company runs its own MCP servers, someone needs to own the stateless upgrade before something quietly breaks.
- The data-cannot-leave-the-building objection is now answerable. If that killed your idea last year, raise it again.
- Treat the hiring numbers as a signal, not a proof. Self-reported, small and midsize businesses only.
What happens next
Expect the effort dial to become normal across every major model, which shifts the skill from writing clever prompts to knowing which tasks deserve expensive thinking. That judgement is harder to copy than any prompt.
And watch whether the hiring pattern holds at enterprise scale. Small companies adopting AI to grow is a very different story from large ones adopting it to cut, and the next few quarters of data will tell us which one we are actually in.
I write this up every week: what changed, what it means for your career, and one thing worth trying before Monday.
Questions people actually ask
What is the effort dial in Claude Opus 5 and why does it matter?
Claude Opus 5, released by Anthropic on July 24, 2026, adds adjustable effort settings so you can choose between fast, cheap output and maximum-quality reasoning on a per-task basis, at the same price as Opus 4.8. Practically, this means routing routine drafting to a cheap setting and saving expensive reasoning for decisions that actually need it, cutting your cost per useful output.
Will the new MCP spec break the AI tools my company already has?
Possibly. The Model Context Protocol’s July 28, 2026 update removes session based connections in favor of a stateless design, which simplifies scaling but can break older self hosted MCP servers when paired with newer clients. If your company runs in house MCP infrastructure, someone should test compatibility and plan the upgrade rather than assuming it will keep working unchanged.
Can I actually run Kimi K3 without sending data to an outside API?
Yes. Moonshot released Kimi K3’s weights openly on July 26, 2026, a 2.8 trillion parameter model reported to perform close to top frontier systems while being substantially easier to self host than earlier large open models. For regulated industries where data cannot leave the building, that removes a common blocker to trying AI at all, though hosting still needs real infrastructure.
Does using AI actually lead to more hiring or more layoffs?
Intuit’s 2026 AI Impact Report found small and midsize businesses using AI report roughly four times more hiring than layoffs, with 78 percent saying it improved productivity. That is self reported data limited to smaller companies, so treat it as an encouraging signal rather than proof, since businesses already growing may simply be the ones adopting AI first.
What’s the one thing I should change about how I use AI this week?
Stop running every task through your AI tool at maximum quality. Use the cheapest available setting for drafting, sorting, and routine tasks, and reserve expensive, high effort reasoning for decisions where getting it wrong actually costs you something. That judgment call, not prompt writing skill, is becoming the more valuable habit to build.
Sources
- Claude Opus 5anthropic.com
- shipped its largest spec update yet on 28 Julyblog.modelcontextprotocol.io
- according to Tom’s Hardwaretomshardware.com