Retrospective weekly edition for 28 September 2026, written 5 October 2026. The archive date is an editorial slot, not an earlier publication claim. This edition discusses announcements already public by that week.
Anthropic shipped Claude Sonnet 5.5 on 28 September 2026. The company says it generates output more than 30% faster than Sonnet 5 and can cost up to 30% less per task because it uses fewer tokens. The standard token prices stayed at $2 per million input tokens and $10 per million output tokens. Its launch page also claims clearer writing and stronger collaboration. These are Anthropic's reported comparisons; TextVeil has not run a writing benchmark on the model. The Sonnet 5.5 announcement.
Good. Less waiting is good. A smaller bill is good. I would happily retire the spinning cursor to a farm upstate.
But speed has a magnificent ability to make the wrong thing arrive earlier. A paragraph can now apologize for its own existence at industrial velocity.
The launch chart is not your manuscript
A successful software task has an unusually satisfying property: sometimes you can run the damn thing and see whether it works. A piece of writing has other obligations. It has to mean the intended thing. It has to earn the reader's attention. It may need to sound annoyed, funny, specific, or reluctant. Smoothness alone is a very low ceiling.
Imagine a model writes: “We are excited to share an update regarding our ongoing efforts to improve the onboarding experience.” Everything is polished. Nobody has learned anything. You could put a faster engine under that sentence and it would still be a stationary object.
Here is an invented alternative: “We removed the three forms that made new customers email us their own screenshots.” Now something happened. The reader can picture the old pain and judge the change. If those forms did not exist, the sentence is a lie. That last condition matters more than whether the model took two seconds or twelve.
Build a writing test out of your actual headaches
Keep a small set of drafts that reliably cause trouble: a blunt customer email, a technical explanation, a personal scene, a paragraph with citations, and a piece where the joke matters. Save the unedited versions. Write down what each piece must retain before you ask a model to touch it.
Then compare the outputs without the model names in your face. Did the angry email become passive-aggressive fog? Did the explanation lose the exception? Did the scene acquire a childhood memory nobody supplied? Did the joke become an explanatory footnote wearing a clown nose?
Count the repairs you actually had to make. That is your cost of using the model. The token bill is one line on the receipt; your time dragging the draft back to earth is another.
This is an editorial test, not a detector test. A classifier's approval would not rescue a rewritten invoice with the wrong amount. Nor would a classifier's suspicion make an original, excellent paragraph worse.
Spend the speed on another look
Faster models make iteration easier. Use the saved time for a sharper brief and a final comparison, rather than requesting ten extra paragraphs because generation has become cheap enough to ignore.
In TextVeil, the point of Humanize is to break the machine cadence while keeping the story. Restyler gives the draft a direction beyond “professionally enthusiastic mammal.” Those jobs stay the same when a provider replaces the motor. A new model does not get permission to replace the product's personality with the launch page's personality.
We want the cheaper, quicker engine when it earns its seat. We also want the weird human sentence that lands harder than five polished ones. The procurement department can have the bar chart. The reader gets the paragraph.