The tool is not the model: what your AI can do keeps changing under you

Overview

On 30 June 2026, a large share of Claude’s Free and Pro users woke up talking to a different, more capable model than the day before. Anthropic had made Claude Sonnet 5 “the default model for Free and Pro plans” — a model “built to be the most agentic Sonnet model yet”, able to “make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models”. Nobody clicked upgrade. The brand on the tab stayed the same; the thing behind it changed.

By the end you’ll hold one habit: treating what your assistant can do as a moving target you re-check, not a fixed trait you learned once.

The content

The obvious way to think about your assistant is as a tool with a known character — ChatGPT is good at this, our company’s AI is bad at that — the way you know what a spreadsheet does. The overturn is that the tool you touch is a thin wrapper around a model you don’t, and the model gets swapped, upgraded, and routed by people who aren’t you. The tool is not the model. What you learned about its limits has a shelf life measured in weeks.

This is sharpest inside the corporate assistant many people actually use, because there the model choice was never yours to begin with. Glean, whose platform sits over your company’s knowledge, is model-agnostic by design: administrators “curate which model families are available”, and the whole pitch is flexibility “in an environment where new models are regularly introduced” (March 2026). Sana says the same — “you can choose and switch between leading models as you need”. In Microsoft 365 Copilot the default model selector reads simply “Auto”, and Microsoft keeps changing what sits behind it: GPT-5.5 arrived in May 2026, and Claude only became selectable on 16 June 2026. So the honest answer to “which model just wrote this for me?” is often: you don’t know, and it isn’t the one it was last month. This is the flip side of owning the harness layer, not just the model — the layer you control is not the layer doing the thinking.

Three things follow for real work. A task you tried once and filed under “the AI can’t do this” may quietly have become one it can — that judgement is stale the moment the model moves. A workaround you built for an old limitation can turn into dead weight, or worse, steer a newer model away from a better route. And when your tool is set to “Auto”, your trust should attach to the output in front of you, not to a reputation the label carries — because the label and the brain behind it have come apart.

Try it

Pick one task you decided, weeks or months ago, that your assistant “couldn’t do” — the summary it mangled, the reasoning it fumbled, the format it kept breaking. Don’t re-prompt cleverly. Just run the same request again, today, in the same tool:

[paste the exact task or prompt you gave up on before]

Read the result against the verdict you filed last time. If it’s now good enough, you were carrying a limit that no longer exists — and probably a few more like it. If it’s still weak, you’ve re-confirmed a real boundary rather than assumed it.

Where it breaks: on a corporate assistant set to “Auto”, you may not be able to tell which model answered — check the tool’s model selector or settings if it has one. If it doesn’t, that opacity is the lesson: calibrate to what the output actually shows you, and re-run this check whenever you hear the tool changed, not on a brand you trust.

Additional reading

• Introducing Claude Sonnet 5 — Anthropic (2026-06-30) — the silent default swap in plain sight: Sonnet 5 becomes “the default model for Free and Pro plans”, “close to that of Opus 4.8, but at lower prices”. Millions of users changed models by doing nothing.
• Release notes for Microsoft 365 Copilot — Microsoft Learn (living page) — the vendor owns the roster and keeps editing it: GPT-5.5 Instant and Thinking added May 2026, “users can now select Claude as a model option” from 16 June 2026. The default selector is “Auto”.
• LLM choice in Assistant — Glean (2026-03-19) — the enterprise-wrapper view stated outright: administrators “curate which model families are available”, built for “an environment where new models are regularly introduced”. The person choosing the model is your admin, not you.
• Sana — product page (living page, 2026) — a second model-agnostic assistant making the same promise — “choose and switch between leading models as you need” — confirming this is the enterprise pattern, not one vendor’s quirk.

Editor’s note

This module is very squarely pointed at the enterprise model wrapper tools that we often encounter at work. You don’t get to choose the models. You don’t get to choose the guardrails. You probably don’t get notified when something that used to be a serious blocker to your work gets unlocked. Keep experimenting, anyway. Models and platforms will change, and every time they do, maybe your whole way of working can change with them.