Arrow 2 puts editable SVG behind a stateless Responses endpoint

A stateless endpoint sets the cost of an editing session. QuiverAI's POST /v1/responses keeps no conversation state, so a client refining a drawing replays the history on every call ^[13]. An SVG is text, and the replayed history carries it, so input tokens grow with the size of the file and the number of turns. Cached input is priced at 10 percent of the uncached rate for both models, $0.40 per million against $4 for Arrow 2 and $0.60 against $6 for Telos ^[18][17][1]. The reported documentation does not say when a request qualifies for that rate.

Output is where a single call gets expensive. Both models cap one response at 65,536 tokens ^[15]. At Arrow 2's output price that ceiling costs $1.31 a call; at Telos's it costs $1.97 ^[2]. QuiverAI's consumer plans start at $8 a month, and API balances are separate from app subscriptions ^[20][19], so the two are not one budget. For scale, $8 of Arrow 2 output tokens is about six full-length responses ^[4].

Runtimewire reports that Telos charges 50 percent more in every token category and offers roughly eight times the context ^[21], and the documented windows put that ratio at 8.0 ^[3]. The routing choice is narrower than that price sheet implies. The output ceiling is identical for both ^[15]. The extra window holds a longer brief and more input SVG ^[6].

QuiverAI says Arrow 2 produces cleaner geometry with fewer control points and better spacing, padding and alignment, and it has not published a comparative benchmark for the new models ^[10]. Runtimewire frames the real test: an SVG is only meaningfully editable when its internal paths and groups are coherent, and not a dense reconstruction that happens to resemble the requested image ^[11]. For that to hold on your assets, the groups the model emits have to match the parts you would want to select, and the node count per path has to be low enough to drag by hand. You can count both on ten of your own briefs.

The workspace is the harder thing to clone. The app opens as a conversation, streams the drawing back, and takes follow-up prompts ^[6]. That same file can be opened on a canvas where child elements are selected and duplicated, with an undo history, and saved canvas edits flow back into the conversation and creation history, according to QuiverAI's September 16th changelog ^[7]. Arrow 2 also animates shapes and groups already present in a file, aimed at logo reveals, loading states, product illustrations and interface icons ^[8]. On the API side, SVG editing and animation depend on the capabilities enabled for each customer's catalog ^[16].

A developer can point an existing Responses client at QuiverAI, declare functions or custom tools, and handle the resulting tool calls inside their own application ^[12]. The same client can be pointed somewhere else. The Arrow 1-series models keep fixed per-asset pricing on the API ^[19].

Rodriguez completed a PhD at Mila and ETS Montreal on multimodal generative models, did research at ServiceNow, and led StarVector, an open model, before QuiverAI ^[22].