v1.14.0 — Take a conversation in two directions
This release is about the middle of the work rather than the start of it.
Fork a conversation and keep your files
Duplicate & Edit copies a conversation up to the message you picked and lets you re-ask that message differently — so you can try a second approach without losing the first.
The fork brings the right files with it: the ones the conversation actually used up to that point, rather than everything ever attached. The message you're editing comes back with its own attachments already on it.
Files can also no longer get stuck as "Unsent" — attached to a conversation but linked to no message, with no way to clear them. Three separate paths produced that state; all three are closed, and the rule for when a file is still in use is now one rule instead of three that disagreed.
A lot more room on Free — and before you sign up
- Free: $2.50 → $15 a day, $9 → $35 a month
- Guest (browsing without an account): $1.50 → $7 a day
These are sized off measurement, not intuition. A full multi-agent conversation costs about $1.36 to run, so the old guest allowance couldn't finish a single one — you'd hit the ceiling mid-answer, which is the worst possible first impression. Free now covers roughly ten conversations a day and twenty-five a month: enough to work through a real multi-step task and still have room to iterate.
Newer models, and billing that matches the published price
Claude Opus 5, Gemini 3.6 Flash, Qwen3.7 Plus and Kimi K3 join the lineup, and every superseded lane is retired. The fast model behind routing and classification — the one that runs on nearly every turn — moved up two generations and is measurably quicker than what it replaced.
We also found and fixed four billing errors by checking every rate against what each vendor actually publishes. The worst: one GPT model was carried at five times its real price, draining budgets five times too fast and ranking it as pricier than a weaker model. Rates are now asserted against the published number itself, not just against each other.
Three more model families can see images
Kimi, Qwen and MiniMax now read images you attach, on the same acceptance rules as every other provider — and size ceilings are enforced before a call is sent, so an oversized file is refused immediately instead of failing deep in a request.
Citations you can check, in every discipline
Agents that cite the literature can verify a reference against the actual record — whether it exists, whether it's a preprint, and whether it has been retracted. This used to cover biology and medicine; it now covers physics, chemistry, mathematics, computer science, economics, the social sciences and the humanities.
More files that just open
3D models, CAD drawings and BIM models render in the viewer — rotate and orbit them in place. Biological sequences and molecular structures render too.
And every one of the fifteen PDF tools was driven end to end and reviewed. Removing a repeated page used to delete the wrong pages and report success. Rendering is now bounded, so one oversized request can't consume gigabytes and stall everything queued behind it.
Quieter, but worth saying
A round of privacy and security hardening. Long conversations keep more context — it's trimmed only when the model handling the turn genuinely lacks room. And long replies that arrived wrapped in a code block no longer shatter mid-render.
The full release note lives in the in-app What's New drawer.
Keep the feedback coming in Feature Requests — tell us what to build next.