What Shipped This Week: One Credit Ledger for Every Feature, Long Chats Stop Crashing

The biggest change this week doesn't add a button anywhere — it removes a bunch of separate daily limits and replaces them with one credit ledger that every AI feature draws from. On top of that, we fixed a crash that was killing long streamed answers mid-sentence, restored PDF uploads that had quietly started failing, reconnected the Chrome extension to the current backend, and corrected video pricing for Seedance 2.0 and AppSumo plans.

What shipped this week

Feature	Category	Who it's for
Unified credit ledger across every AI feature	Improved	All paid plans, especially agentic chat and developer API users
Long streamed answers no longer crash (React error #185)	Fixed	Anyone with long chat sessions
PDF/DOCX uploads work again	Fixed	Anyone attaching documents to chat
Chrome extension reconnected	Improved	Chrome extension users
Model catalog & comparison pages rebuilt on real data	New	Anyone choosing between models
Seedance 2.0 / AppSumo video pricing corrected	Fixed	Video generation users, AppSumo customers

One credit ledger for every AI feature

We already told you a couple weeks ago that daily per-model message caps were going away in favor of a shared monthly credit pool. This week we finished the job: the ledger now covers agentic chat, the developer API, images, video, text-to-speech, speech-to-text, web search, and scraping — not just plain chat, images, and video.

The reason this mattered is that people were hitting inconsistent walls. You could run out of image generations while still having chat messages left, or burn through your daily agent-tool budget separately from your regular conversation budget, with no way to move capacity between them. A single ledger means your usage is fungible: if you're doing more chat and less video this month, you're not penalized for it.

Pricing stays the same idea as before: 1 credit equals roughly $0.01 of underlying provider cost, tracked in millicredits internally so cheap messages cost their true fraction of a credit instead of rounding up. Every model is now available on every paid plan — expensive models just draw down more credits per use instead of being locked behind a tier.

To see where you stand, open your usage page from account settings. If you're an AppSumo lifetime deal holder, your entitlements are grandfathered using your original purchase timestamp, so this change doesn't reset or shrink what you already have.

Long streamed answers stop crashing mid-sentence

If you've had a long answer suddenly die with an inline "An error occurred" card and a Retry button, that wasn't a model problem — the model had actually finished, and the full answer was already saved on the server. The crash was happening in your browser.

The cause was subtle: the smooth-scroll animation that keeps your view pinned to the bottom of a growing answer was calling a state update on every animation frame. On sufficiently long answers, those updates chained past React's internal limit for nested updates in a single render cycle, and React threw an error that looked like a chat failure but wasn't one at all.

We've fixed the scroll behavior so it no longer trips that limit. Long answers — the kind with extended reasoning, big code blocks, or multi-thousand-word responses — now render all the way through without an error card interrupting you. If you saved a workaround like manually scrolling during long generations, you don't need it anymore.

PDF and DOCX uploads work again

A backend upgrade to our AI SDK quietly broke document uploads. Writingmate pre-processes PDFs, DOCX files, and similar formats into markdown before sending them to a model, but the new SDK version was treating that processed markdown as a raw file attachment labeled with the original file's MIME type — so a converted PDF would show up to the model still tagged application/pdf while actually carrying markdown text. Anthropic's API rejected the mismatch outright with a 400 error; other providers handled it inconsistently.

This is fixed now. Drop a PDF, DOCX, or other supported document into any chat and it gets converted and read correctly, on every model that supports file input. If you had switched to copy-pasting document text manually as a workaround this week, you can go back to just attaching the file.

Chrome extension reconnected, model pages rebuilt

The Chrome extension (v2.8.0) got a dedicated per-installation developer key handoff through a new /extension-auth connection flow, replacing whatever credential path it was using against an older version of the backend. Keys are scoped and revocable individually, and extension credentials are now kept out of the browser's service-worker cache, so a compromised or uninstalled extension can't leave a lingering token behind. If your extension stopped working recently, reinstalling or re-authenticating from the extension popup should reconnect it.

Separately, we rebuilt the model catalog and model comparison pages so every listing is backed by real, sourced data instead of static copy — pricing tiers, context windows, benchmark scores, and Arena rankings pull from the same catalog that powers the rest of the product, with explicit "not available" states instead of guessed numbers when a data point is missing. The homepage's frontier-model strip and every one-on-one model comparison page now reflect the current 400+ model catalog. You can browse it at writingmate.ai/models.

Fixed this week

• Seedance 2.0 video generation was billing 4 credits per output second instead of the correct 30.34 — pricing is now accurate, and AppSumo video seconds move to the new 5/15/40/80 monthly tiers starting August 15, with existing AppSumo entitlements grandfathered at their original purchase amounts
• Trial subscription upgrades were failing with a Stripe configuration error when confirming a plan change — upgrade confirmation now works correctly
• The OpenAI-compatible API endpoint was rejecting any request that included a system or developer message, breaking roughly 85 requests an hour across every model — system messages now pass through as sent
• Requests with an invalid ID in the URL were returning a raw server error instead of a clean not-found response
• Free-plan users who also belong to a paid workspace were stuck on the upgrade paywall with no way to dismiss it and get back into the app — eligible users now get a close button
• Pages across /models, /docs, and /blog were serving slower dynamic responses instead of cached static pages — static rendering is restored, so these pages load faster again

Thanks for reading this far — a lot of this week's work was under-the-hood correctness that you'll mostly notice as things simply not breaking anymore. If you want the full history of what's shipped, past updates live at writingmate.ai/changelog. See you next Friday.

— Artem