The most useful thing we shipped this week is four new models you can pick in chat today: Flint, MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017 and Glimpse. We also fixed an error that could appear if you deleted a chat while an answer was still streaming, and we stopped refreshed blog posts from losing their hero images. It was a quieter week than the last two, which were mostly billing changes, and we would rather say that plainly than pad the list.
What shipped
| Feature | Category | Who it's for |
|---|---|---|
| Flint, MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017, Glimpse in the model picker | New | Anyone who compares models, especially for coding, long documents and multimodal work |
| Safe handling of chats deleted during a streaming reply | Fixed | Anyone who cleans up chats quickly, and API users running long requests |
| Blog posts keep original hero image, date and category | Fixed | Blog readers and anyone sharing our article links |
| Daily model catalog refresh and daily blog publishing | Improved | Anyone who wants new releases and how-to guides soon after launch |
Four new models in the picker
We added Flint, MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017 and Glimpse. Each one shows up in the model picker as soon as it appears in the catalog, which we pull from OpenRouter and refresh every day. This week the catalog refreshed on each of the last seven days, so nothing waited on a manual update.
Why it matters: new models arrive faster than anyone can test them. Our goal is that you can switch to a new release in one click and run your own prompt against it, rather than signing up for another account first. Our early notes on each model are in the blog. Flint is aimed at agentic coding and tool work, Glimpse at image, text and video input, and the other two at long-context drafting and research.
To try one, open /chat, click the model name at the top of the conversation, and search for the model by name. If you want to compare two side by side, send the same prompt to both and judge the answers on your own task, not on a benchmark. Paid models count against your plan the same way other models do, and workspaces with pay-as-you-go turned on can keep going past plan limits. See the changelog for how that works.

The picture above shows the picker. The new models appear in the same list as the rest, so there is nothing new to learn.
Deleting a chat mid-reply no longer errors
When you send a message, we stream the answer back and save the final text to the chat once it finishes. If you deleted that chat while the answer was still streaming, the save had nowhere to go, and the database rejected it with a foreign-key error. That surfaced as a server error in our logs, and it could also show up for API callers who delete a conversation while a long request is running.
We now treat that specific case as expected. If the chat is gone by the time the final message is ready, the write is skipped quietly, because there is nothing left to attach it to. We kept the rule narrow on purpose: other database errors still surface, so a real problem with saving messages will not be hidden by this change. We added tests for both the deleted-chat case and the unrelated-error case.
You do not need to do anything. If you tidy up your sidebar from /chat while a reply is still running, nothing should break now. The only visible effect is the absence of an error. We found this one by scanning production logs for repeated failures, which is also how we plan to catch similar races.
Blog posts keep their original hero images
Some of our older articles are refreshed through a publishing tool. When a refreshed article reused the address of an existing post and the tool sent no image, the post lost its hero image and showed the default logo. It also took the publish day as its date and landed in the general category. About twenty posts were affected, including our guide on whether Claude can access the internet.
We changed the publisher so that a post with a hand-written original keeps that original's cover, date and category. Posts that only ever came from the tool behave as before, and fresh tags can still update their category. We restored the cover, date and category on the twenty affected posts, and added tests that check the original values win.
If you shared one of those links and the preview looked wrong, it should now show the right image once the preview cache refreshes. You can browse the blog at /blog. We cannot control how fast other sites refresh their cached previews, so an old thumbnail may linger for a while.
Fixed this week
- Deleting a chat during a streaming reply no longer causes a server error when the reply tries to save.
- About twenty refreshed blog posts got their hero images, dates and categories back.
- Refreshing a blog post can no longer reset its date to the day it was republished.
- The daily model catalog refresh ran on all seven days this week, so the picker stayed current.
What we are working on next
This week was mostly maintenance and model coverage. We are still tuning the pay-as-you-go and friend-invite flows from last week, and the next recap will cover whatever comes out of that. Some of the work on the blog automation is rough around the edges, and we are iterating on it as we find problems.
Thank you to everyone who sent us bug reports and model requests, since several of this week's changes started that way. The full history of what we have shipped is on /changelog. See you next Friday.
Artem
Frequently asked questions about this week's update
Sources
Written by
Artem Vysotsky
Ex-Staff Engineer at Meta. Building the technical foundation to make AI accessible to everyone.
Reviewed by
Sergey Vysotsky
Ex-Chief Editor / PM at Mosaic. Passionate about making AI accessible and affordable for everyone.