What is a context window?
A context window is how much text, code or instructions an AI can 'see' at once to reason about your task. Bigger means more can go in, but not always better answers.
Issue 009 · 6 min read
This week big reductions in both price and friction make it tempting to try AI everywhere. That's most useful right up to the point you move a draft or version into someone else's hands without looking at what changed. Unifying side panels and sped-up image edits mean it's simpler to whip up drafts in your flow, not just in a sandbox. Your real unlock isn’t more drafts, but a pause to check what's left your desk before you pass it on.

Rather listen?
Skip if you like
A context window is how much text, code or instructions an AI can 'see' at once to reason about your task. Bigger means more can go in, but not always better answers.
It's the bit at the side of your Gmail, Docs or Slides that lets you run AI tools inside your usual screen. You can draft, edit or pull info without switching apps.
A draft checkpoint is the pause where you check what the AI really changed before anything leaves your hands. It’s your chance to reject what doesn’t fit or you can’t stand by.
Start here
Newly announced
You could now run AI on every message, spreadsheet or file, but the lowest price doesn't cover who reads your data or whether the output holds up. The real lift isn’t just scale, but knowing when to pause for human sense.
See the detail and the sources →Newly announced
The faster, more flexible image drafts mean one button now spits out more versions before you have time to spot a classic AI blunder. You need to catch what changed when trying 'just one more version'.
See the detail and the sources →Newly announced
Any Workspace file can now sprout an AI suggestion without switching apps, so it's easy to let a draft get sent right from a slide or note. Reviewing what flows from one click keeps you from shipping half-baked work.
See the detail and the sources →
The one that mattersWhat's genuinely new
AI drafts can now be generated or tweaked from nearly any screen in your flow, at minimal cost and with barely any lag. Combined with deep price drops, this makes high-volume, multi-app and cross-medium drafting thinkable.
What the makers claim
Much of the week's announcement noise is about price (DeepSeek), feature blanket coverage (Gemini), or generation speed (ChatGPT). These show off scale and convenience rather than specific improvements in output reliability.
Where it helps
You’re editing a proposal, slide or reply in Workspace or any chat platform with AI built in. You tweak an image or doc and hit send from wherever you are, skipping the app switch and the second look.
Where it doesn't
When work is locked down, such as final sign-off rounds needing formal review, or where AI isn’t yet built into your main tool at all.
What remains yours
DeepSeek's V4-generation API models are deepseek-flash, now served by the DeepSeek-V4.1-Flash model, and deepseek-v4-pro (DeepSeek-V4-Pro-0813). Both carry a one-million-token context window and up to 384K output tokens, support thinking (the default) and non-thinking modes, JSON output and tool calls; the original deepseek-v4-flash and deepseek-v4-flash-vision-exp model names are retired and now route to deepseek-flash at its price. deepseek-flash is priced at US$0.30 per million input tokens (US$0.006 on cache hits) and US$1.20 per million output at peak rates, with off-peak billing at half those rates; deepseek-v4-pro is roughly three to four times that: US$1.32 input and US$3.96 output at peak.
Verify and say what new claim, image or task made it from AI output to your outgoing draft that you did not write or edit yourself.
Try this
Set up a draft checkpoint before anything leaves your scope.

The AI creates or edits a doc, image, or email draft with your usual instructions or reference sketches, right from the app side panel or chat.
Now, before a single AI-altered draft leaves your hands, catch what automated changes it actually made.
Pick one routine doc, mail, slide or image you’d usually push out in a hurry this week.
Ask the AI to generate or edit as you normally would, without hopping apps or opening new tabs.
Pause for two minutes before sending. Compare the new version with the last human draft. Name out loud what changed and what you can’t personally vouch for, then choose what to keep, strike or rewrite.
Verify and say what new claim, image or task made it from AI output to your outgoing draft that you did not write or edit yourself.
The Human Lens · what stays yours
No AI tells you what responsibility you take by sending a draft along. You’ll catch data leakage, misstatements or tone mismatches in places a model never will. When the tool works from anywhere, your trust boundary is what moves, not just your workflow.
One question
See you next Monday.
Bring it back to your work
Bit starts from this issue, then asks you enough about your own work to make it worth acting on. If the honest answer is that none of this touches your job yet, it will tell you that too.