The workflow connects Tripo's August mesh model to OpenAI's computer-use agent, while leaving deformation checks and physics setup to the creator.
By [RuntimeWire Staff](/author/runtimewire-staff)
· Published
Primary source: [Tripo AI on X](https://x.com/tripoai/status/2098785528527442276)
Why it matters #
AI 3D is moving toward assets that survive editing, rigging and game engines. If P2.0's topology holds up outside Tripo's demos, Song could make Tripo a dependable input layer for general-purpose agents. The required human checks show the remaining distance.
On September 13th, Tripo AI, the generative 3D platform founded by Simon Song, said P2.0 meshes can provide structured geometry for OpenAI's Astra computer-use model to handle downstream work in Blender. The company is making a larger bet: AI-generated 3D becomes valuable when the output can survive the rest of a production pipeline.
The distinction matters because image-to-3D systems have become good at producing an object that looks convincing from a favorable angle. A character that must be edited, rigged and animated faces a harder test. Its polygons need useful structure, its joints need room to bend, and its separate parts need to remain manageable inside tools such as Blender.
Song founded VAST, which operates Tripo AI, in 2023 after working on AI-generated animation and gaming at SenseTime and helping start large-model developer MiniMax. In an interview, Song described himself as a longtime gamer and anime fan whose interest in content creation pushed him toward 3D. Tripo's research materials describe a broader company vision of making interactive worlds accessible to more creators.
The asset has to survive Blender
Tripo introduced P2.0 Preview in its August product update, describing it as a Smart Mesh model that generates native quad topology from multiple reference views. Tripo says generation typically takes 10 to 40 seconds and supports outputs of up to 50,000 triangles or 25,000 quads.
Quads alone do not make a character production-ready. Their arrangement determines whether an artist can edit the model cleanly and whether a shoulder, knee or face deforms without collapsing. Tripo says P2.0 also separates semantic parts such as clothing and armor, giving downstream tools more useful boundaries for rigging and texturing.
The Astra example puts that claim into a broader workflow. According to OpenAI's September 3rd launch post, GPT-6 Astra has computer-use capabilities for multistep work in websites and desktop applications. Tripo's demonstration uses Astra through a configured Blender MCP connection, effectively making the model an operator inside the 3D software rather than treating it as another asset generator.
Tripo's full character tutorial shows how much work sits behind the short demo. The creator first prepares separate references for a character's body, head and hair, generates those parts in Tripo, imports them into Blender and asks Astra to align them. The tutorial instructs the creator to inspect neck joins, proportions, surfaces, materials and the silhouette, including front, side and back views.
Rigging has similar conditions attached. Tripo's preferred route imports a body already rigged in Tripo and asks Astra to reuse its armature and weights. Tripo's rigging instructions say Astra can create and bind a biped rig when no usable rig exists. The creator is then directed to test head turns, arm raises, elbow bends and knee bends, refining local weights where deformation fails.
Physics remains a separate step. Tripo's example uses keyed hair movement and explicitly says physics-based motion requires its own setup. The workflow therefore supports a narrow claim: structured Tripo assets can give an agent better material for downstream Blender work. It does not establish automatic, one-prompt physics simulation.
Song is building for the unglamorous middle
In the interview, Song said Tripo aims to eliminate manual cleanup and retopology, work that production teams want to avoid after generating an asset. He has described AI as a tool for improving creative efficiency and expanding what creators can make.
P2.0 follows that thesis closely. Its value proposition rests on topology, segmentation and export behavior, details that receive less attention than visual fidelity but decide whether a generated asset saves time after the initial render. Tripo's August update also described work on its bridges for Blender, Maya, 3ds Max and Unreal, along with faster FBX exports and browser-based model utilities.
The approach reflects the backgrounds around Song. Yanpei Cao, VAST's co-founder and chief scientist, previously worked on 3D digitization and generation at Tencent ARC Lab and Tencent AI Lab. That pairing gives Tripo a founder focused on how creators encounter the product and a research leader steeped in the geometry problems underneath it.
Fresh capital raises the stakes
Tripo's financing history adds a substantial commercial backdrop to the P2.0 preview. On September 1st, Tripo said it had raised approximately 3 billion yuan across Series B and B+ rounds led by MPCi. The company release listed Perfect World, BlueFocus and 37 Interactive Entertainment as strategic investors, CICC and CMC Capital Partners as financial investors, and said existing investors including Fortune Capital, Primavera Capital and 4399 Network increased their investments.
Tripo said the financing would support model research, 3D data infrastructure, computing capacity and commercialization. Those spending categories map directly onto Song's strategy. Better geometry requires specialized training data and compute, while winning production teams requires integrations, dependable exports and support around existing creative software.
Earlier 2026 financing announcements are difficult to combine cleanly because Tripo disclosed several rounds and tranches under different labels. A March announcement reported a $50 million Series A led by Alibaba and Hengxu Capital. A June 1st release separately described nearly $200 million across Series A+ and A++ rounds. Those figures should be read as individual disclosures rather than a reliable cumulative total.
Every 3D generator wants the workflow
Tripo is competing in a market where rivals are also moving beyond prompt-to-object demos. Meshy's August changelog lists Smart Topology, semantic segmentation, text-to-motion tools and a 3D agent that can manage generation and post-processing tasks. Hyper3D's Rodin line similarly offers controls over geometry modes and mesh density.
That competition explains why Tripo is putting P2.0 beside Astra so quickly after both products appeared. General-purpose computer-use agents could weaken the advantage of owning every downstream feature inside a single 3D application. They also create an opening for the generator that supplies the cleanest, most editable geometry to whichever agent a creator chooses.
Song is betting Tripo can own that input layer. The workflow gives the thesis a concrete shape: Tripo makes the asset, Astra operates Blender, and the creator reviews the places where geometry and motion still break. The cleaner those handoffs become, the closer Tripo gets to turning generated meshes from impressive outputs into dependable building blocks.