QuiverAI ships Arrow 2 with editable SVG workspace and developer API
The San Francisco startup paired its vector models with canvas editing, Open Responses support and token-based API pricing.
By Ryan Merket · Published
Primary source: QuiverAI on X
Why it matters
QuiverAI is packaging specialist SVG research as a full design platform, betting that editable visual code will be more valuable to production teams than another finished image.
Joan Rodriguez (@joanrod_ai), the researcher who turned SVG generation into QuiverAI, released Arrow 2 and a more capable Arrow 2 Telos model on September 16th across the QuiverAI app and API. The release moves QuiverAI beyond generating vector files and into the harder business of owning the workspace where designers edit them and the infrastructure developers use to put them inside other products.
https://x.com/QuiverAI/status/2100295136261349802
QuiverAI first detailed the models in a September 7th product post, then made the broader app and API stack available on September 16th. Arrow 2 is positioned as the faster, lower-cost option for routine generation and iteration. Arrow 2 Telos is intended for complex briefs, longer contexts and designs that require additional refinement.
Both models generate SVG, a code-based graphics format whose individual paths, shapes and groups can be edited after an image is created. That distinction is QuiverAI's central bet. Most generative image systems return a finished raster image. QuiverAI returns the underlying visual structure, giving designers and software tools access to the elements behind the render.
From model demo to design workspace
The redesigned app begins with a conversation. Users describe an asset, receive a streamed drawing and can continue refining it through follow-up prompts. The same SVG can be opened on a canvas for direct changes, including selecting or duplicating child elements, navigating an undo history and adjusting the view. Saved canvas edits flow back into the conversation and creation history, according to QuiverAI's September 16th changelog.
QuiverAI also added animation to the canvas. Arrow 2 can animate shapes and groups already contained in an SVG, targeting assets such as logo reveals, loading states, product illustrations and interface icons. Arrow 2 replaced Arrow 1, Arrow 1.1 and Arrow 1.1 Max inside the app, although the older models remain accessible through the API.
QuiverAI says Arrow 2 produces cleaner geometry with fewer control points and improves spacing, padding and alignment. Those are company claims, and QuiverAI has not published a comparative benchmark for the new models. The practical test will be whether Arrow 2 reduces the manual cleanup that frequently follows generated design work. An SVG is only meaningfully editable when its internal paths and groups are coherent, rather than a dense reconstruction that happens to resemble the requested image.
An API built to fit existing agent software
The developer release matters because QuiverAI has adopted an Open Responses contract for its POST /v1/responses endpoint. Developers can point a compatible Responses client at QuiverAI, declare functions or custom tools and handle the resulting tool calls inside their applications. QuiverAI's endpoint is stateless, so clients must replay conversation history instead of relying on a stored response identifier.
The developer documentation lists a 131,072-token context window for Arrow 2 and a 1.05 million-token window for Arrow 2 Telos. Both have maximum outputs of 65,536 tokens and accept text, images and SVG inputs. SVG editing and animation depend on the capabilities enabled for each customer's API catalog.
QuiverAI also changed the economics of its API. Arrow 2 costs $4 per million input tokens and $20 per million output tokens. Arrow 2 Telos costs $6 and $30, respectively. Cached input is priced at $0.40 per million tokens for Arrow 2 and $0.60 for Telos. The older Arrow 1-series models retain fixed per-asset pricing, while API balances remain separate from app subscriptions. QuiverAI's consumer plans start at $8 per month.
The token model exposes QuiverAI to the same cost-versus-quality calculation faced by general-purpose AI labs. Telos charges 50% more in every token category and offers roughly eight times Arrow 2's context window. That gives developers a clear routing decision: use Arrow 2 for repeated asset production and reserve Telos for briefs whose complexity can justify longer inference and higher cost.
Research becomes a product stack
Rodriguez completed a PhD at Mila and ETS Montreal, where he worked on multimodal generative models, and previously conducted research at ServiceNow. Before QuiverAI, he led StarVector, an open model that treated image-to-SVG conversion as code generation. StarVector combined a vision encoder with a code model to produce SVG markup rather than flattened pixels.
That work supplied the technical thesis behind QuiverAI: a visual model should understand both how an image renders and how its underlying code is organized. Rodriguez later worked on reinforcement learning from rendering feedback, which compares generated SVGs with their rendered images to reward visual fidelity and structural efficiency.
QuiverAI disclosed an $8.3 million seed round on February 25th led by Andreessen Horowitz. The new release shows where that capital is going. QuiverAI is building a vertically integrated design product around Rodriguez's research, spanning a proprietary model family, an interactive editor, developer endpoints, billing controls and compatibility with agent software.
That stack gives QuiverAI several ways into a design workflow. Designers can work in its app, software vendors can embed the API, and coding agents can call Arrow when they need a structured graphic. Arrow 2's quality claims still need independent testing. The product strategy is already visible: QuiverAI wants the editable SVG to become a native output of AI software, then collect revenue wherever that output is created.