2026 was the year splats became an ordinary file type: glTF and OpenUSD gained standard splat formats, and 3ds Max, SketchUp, V-Ray, Babylon.js and three.js all added native support.

Vantage is Chaos's real-time ray tracer, used by archviz studios for fast animations and client reviews of V-Ray scenes. Version 3.0 in October 2025 added splat import. 3.3.0 lets imported splat captures respond to lights placed or adjusted inside Vantage, so one session can mix modelled buildings and scanned environments under the same lighting. It also adds DLSS Frame Generation. This brings Vantage in line with the relighting in V-Ray 7 update 4, so a studio can keep one lighting setup for drone-captured context from interactive review through to final frames.

three.js r186 includes its first built-in Gaussian splat renderer and loaders, written in TSL for WebGPU and WebGL and contributed by Ben Houston. The release covers glTF import via KHR_gaussian_splatting, view-dependent colour through spherical harmonics, per-mesh bounding boxes, frustum culling and raycasting, with the addon renamed from GaussianSplatMesh to GaussianSplat. Spark and GaussianSplats3D remain the options for LoD and streaming. For three.js and R3F projects, simple splat embeds no longer need a third-party renderer.

Opus 5.5 is the first Claude 5.5 model and follows Opus 5 from July. Anthropic priced it at $4 input and $20 output per million tokens, with cache reads at $0.20, and reports output over 30% faster than Opus 5. It scored 66.4% on Terminal-Bench 4.0, and in one test it improved web app load times in 39 of 40 attempts. It is available as claude-opus-5-5 on the Claude API, AWS, Google Cloud and Azure, with Sonnet 5.5 and Haiku 5.5 announced for the following weeks.

For daily agent work it lowers the cost of using Opus as the default, and the load-time result is relevant to heavy 3D pages.

Safari 27.0 shipped on 17 September 2026. The <model> element, which shipped on visionOS a year earlier, now works in Safari on iOS, iPadOS and macOS. It embeds a 3D model (for example USDZ) the way <video> embeds video, with <source> fallbacks, environmentmap lighting, stagemode="orbit" and a JavaScript API. dynamic-range-limit now applies to it. visionOS sites can also offer immersive environments. WebXR Layers gain texture-array projection layers, and WebGPU adds clip_distances in WGSL.

Apple users now get a native, no-engine way to show a single furnished unit or product model.

XGRIDS introduced the Lixel L3 as its flagship spatial scanner at INTERGEO 2026. It combines LiDAR, a 4K HDR camera with a 1/1.3-inch sensor, integrated RTK and a 4.3-inch display. XGRIDS claims 5 mm relative accuracy, about 2.5 cm real-time absolute accuracy, and coverage of roughly 100 square metres in 10 minutes. One scan produces a true-colour point cloud, a mesh and a 3DGS model, and the LCC2 format works in 3ds Max and SketchUp. Pricing and availability were not announced. For studios capturing existing interiors or sites, it pairs survey-grade geometry for BIM with a splat for the visuals.

SuperSplat Editor 3.0.0 (GitHub release 8 September, blog post 9 September) rewrites both the renderer and the editor's data model on WebGPU, and it now needs WebGPU. On a 4.4 million splat scene, idle memory falls from 1,557 MB to 105 MB, and rendering 1080p video falls from 1,673 MB to 142 MB. On a 2.4 million splat scene, an orient click goes from about 8.5 seconds to under 25 ms. New features include a sphere brush selection, colour grading applied to selections, stochastic alpha rendering and a new Appearance panel. It runs on Engine 2.22.0.

Cleaning large captures in the browser is now practical on ordinary laptops.

GPT-6 Astra is the first model of OpenAI's GPT-6 generation. GitHub's testing found it plans and validates as it goes, batches diagnosis with verification and confirms results before declaring a task done, which cut steps on long coding tasks. It is available to Copilot Pro+, Max, Business and Enterprise with usage-based billing. On 22 September OpenAI added GPT-6 Sol and Luna to Codex at lower token prices than their GPT-5.6 predecessors.

For solo developers the GPT-6 family is the main OpenAI option for long unattended agent runs, and the Sol and Luna tiers reduce the cost of routine calls.

Seven months after the release candidate, the KHR_gaussian_splatting README in the KhronosGroup/glTF repository was updated to "Complete, Ratified by the Khronos Group", later than the Q2 2026 target. Listed contributors include Cesium, Niantic Spatial, Esri, NVIDIA, Huawei, Autodesk and Khronos. The SPZ compression extension (KHR_gaussian_splatting_compression_spz_2) was still an open pull request at the time of writing. Ratification means the baseline is fixed, so exporters and viewers can target it without expecting breaking changes; compressed glTF splats remain pending.

The SuperSplat Publishing API lets capture and training tools upload splats directly. It authenticates with PlayCanvas access tokens used as bearer tokens, supports resumable multipart uploads for large files, publishes as unlisted by default, and takes metadata including software attribution. Three integrations launched with it: Teleport by Varjo, XGRIDS LCC Studio, and LichtFeld Studio through an MIT-licensed plugin written by PlayCanvas.

An in-house splat farm could now push finished scenes to hosted SuperSplat viewers automatically, with no manual multi-gigabyte export step.

DJI Terra is the reconstruction software bundled with DJI's enterprise drone ecosystem. The July 2026 update spreads visible-light and LiDAR Gaussian splat jobs across a cluster of networked computers, so reconstructions are no longer limited by one machine. Exports cover TerraGS, PLY and 3D Tiles. Video files can now be used as input directly, with automatic keyframe extraction, and the update also brings rolling shutter correction, reduced water surface noise in splats, thermal 2D reconstruction and an AI assistant. Studios producing aerial context splats for developments can now process larger sites from standard DJI flights in a single job, without splitting them into tiles by hand.

Volinga, a commercial Gaussian splat suite and Unreal Engine plugin used in virtual production, made version 0.10.0 available after SIGGRAPH 2026. It adds support for UE 5.8 and for the SPZ and SOG splat formats alongside PLY. It lifts the previous 4 GB file ceiling, improves VRAM use and adds DLSS. Animated PLY sequences convert into NVOL assets for real-time 4D splat playback. Workflow additions include sublevel streaming, Sequencer keyframing of radiance field actor properties and user-supplied masks during processing. Since Epic still has no native splat support, this is one of the main supported routes for studios putting SOG-compressed captures into Unreal-based interactive property experiences.

This was the largest MCP revision since launch and includes breaking changes. The initialize handshake and Mcp-Session-Id header are gone, so any server instance behind a load balancer can answer any request. Multi Round-Trip Requests replace server-initiated calls, Mcp-Method and Mcp-Name headers allow routing without parsing bodies, list results carry cache hints, and authorisation moves to Client ID Metadata Documents. Tasks, MCP Apps and Enterprise Managed Authorization became formal extensions. TypeScript, Python, Go and C# SDKs were updated, with Rust in beta.

For a developer shipping an MCP server for a 3D or pipeline tool, remote hosting is now an ordinary stateless HTTP deployment, for example on a serverless platform.

Autodesk 3ds Max 2027.2 adds native Gaussian splat support. Splats import from PLY, SPZ (Niantic) and LCC (XGRIDS), preview in ActiveShade, and sit on a new Points object type. That means the new Point modifiers and existing modifiers can edit them. Turn to Point converts meshes and splines to point data, and Point Instance scatters objects at point positions at render time. Arnold 7.5.2 through MAXtoA 5.9.3 renders and relights splats. Smart Bevel is reported up to 20 times faster. As the main modelling host for archviz, 3ds Max can now crop, clean and place captured context in the scene file, without an external splat editor.

Engine 2.21.0 adds support for the KHR_gaussian_splatting glTF extension and for SPZ splats through a pluggable external parser. It also adds a physics backend abstraction, an instanced wide-line renderer and XR manipulation and navigation script upgrades. Engine 2.22.0 on 4 September followed with error-driven splat LOD selection, a LOD mode with per-splat detail falloff for streamed splats, and splats that write scene depth.

PlayCanvas can now load the same standard splat files as Cesium and three.js, so a viewer is no longer tied to SOG or PLY input.

Chaos V-Ray 7 update 4 was shown at SIGGRAPH 2026. Its headline for captured context is splat asset relighting: the baked-in lighting of a scan can be adjusted to match the CG scene, so captured locations blend with modelled buildings without the usual workarounds. Relighting works in V-Ray for 3ds Max, Maya and Houdini and in Chaos Vantage, with Blender support announced as coming. The update also puts Vantage real-time rendering, at final-frame quality, directly in the DCC viewport. Chaos quotes GPU interactive previews up to twice as fast. It supports 3ds Max 2027 and Blender 5.1 and 5.2. For studios mixing drone splats with V-Ray renders, the lighting mismatch between capture and CG no longer needs manual fixes.

GPT-5.6 introduced a naming scheme where the number marks the generation and Sol, Terra and Luna mark capability tiers. OpenAI priced them at $5/$30, $2.50/$15 and $1/$6 per million tokens respectively. The same day OpenAI launched ChatGPT Work, an agent powered by Codex and GPT-5.6, and GitHub made all three tiers available in Copilot.

The tiered naming is the practical change: developers can pick a cheap tier for high-volume pipeline calls and the top tier for long agentic runs within one model generation, and the naming carried into GPT-6 in September.

r185 adds WebXR support to WebGPURenderer and a fallback for XR sessions when WebGPU cannot be used, alongside removals of deprecated code and many TSL fixes. r186 in September continued the work with MSAA for WebXR on WebGPU, native WebGPU XR camera selection and upfront detection of WebGPU XR support.

Until now, a three.js scene moving to WebGPU had to give up headset support. With r185 a single WebGPU-first codebase can also serve Quest, Android XR and visionOS browsers, which matters for any immersive walkthrough built on three.js or R3F.

Sonnet 5 is Anthropic's mid-tier model for agent workflows, described as its most agentic Sonnet and approaching Opus 4.8. It became the default on Free and Pro plans and is available as claude-sonnet-5. Pricing was launched as introductory at $2/$10 per million tokens against a planned $3/$15, and a 10 August edit to the post made the lower price permanent.

This is the model that matters most for cost. Pipeline scripts that call Claude in a loop, batch asset tagging or CI review jobs get near-Opus results at well under half the Opus price, which changes what is affordable to automate on a personal budget.

Trimble's Gaussian Splats extension brings captured environments into SketchUp models. It imports .ply, .splat, .spz, XGRIDS .lcc and a new .sklat format. An import wizard handles crop, orient, align, scale and position, and splats work with Scenes, Section Planes and LayOut. Display modes switch between rendering, modelling and point cloud views. It is Windows only, needs a Pro or higher subscription, and costs nothing extra while in Labs. It does not train splats, so capture still happens in apps such as Scaniverse or Polycam. For archviz, this puts site captures in front of architects at the design stage, before the model reaches the visualisation studio.

Unreal Engine 5.8 makes MegaLights production ready for scenes with many dynamic shadowed area lights. It adds Lumen Lite, a beta global illumination mode based on irradiance fields that Epic says is twice as fast as high-quality Lumen. Other additions are a mesh-based terrain system, the Procedural Vegetation Editor for Nanite foliage, AxF measured material import for Substrate, accumulation depth of field in the path tracer and a Pixel Streaming 2 backend upgrade. Epic said 5.8 is the last UE5 release, with UE6 Early Access expected around early 2027. There is still no first-party Gaussian splat support, so splats in Unreal continue to depend on third-party plugins.

Fable 5 is a public version of Anthropic's Mythos-class model, with safeguards that route flagged cybersecurity, biology and distillation requests to Opus. Anthropic describes it working autonomously for longer than earlier Claude models and staying focused across millions of tokens; one early tester reported a codebase-wide migration in a 50-million-line Ruby codebase. It costs $10/$50 per million tokens as claude-fable-5. Subscription plans included it only until 22 June, after which usage credits applied.

For a solo developer it became the model for the hardest refactors and migrations, but at double Opus pricing it suits occasional heavy jobs rather than daily use.

PlayCanvas Engine 2.19.0 introduced a WebGPU renderer that moves culling, projection and sorting to compute shaders. PlayCanvas reports up to 5.7x speed-up over WebGL 2 at 35M Gaussians on an M4 Max and about 2x on an iPhone 13 Pro Max, with automatic WebGL 2 fallback. SuperSplat now processes every upload into Streamed SOG, loading the lowest LODs first and refining to a per-device Gaussian budget; the showcase scene has 24M Gaussians. All pieces remain MIT licensed. For web archviz this is the default open stack for large splats on phones.

Chaos Corona 15 for 3ds Max and Cinema 4D builds on the native splat rendering added in Corona 14. Imported Gaussian splats now bounce global illumination onto nearby 3D geometry, so a drone-captured context lights a proposed building instead of sitting behind it as a backdrop. The release also puts Veras generative AI inside the frame buffer on all plans. It adds a Glints layer to the Physical Material and faster animation parsing in 3ds Max that reprocesses only changed elements. Chaos Cloud collaboration gains browser-based 3D streaming links for clients. For archviz studios placing a scheme into a captured site, splat GI is one of the year's most practical rendering changes.

Rijsdijk, Peters, Weinmann and Marroquim avoid per-frame depth sorting by sampling pixel-sized, opaque points from each Gaussian and splatting them into a framebuffer with 64-bit atomics, which spreads work evenly across millions of GPU threads. The authors report real-time rendering of hundreds of millions of Gaussians. Published in ACM Transactions on Graphics (Proceedings of SIGGRAPH 2026). Depth sorting is a major per-frame cost in current viewers, so a sort-free approach that holds up at city or campus scale is relevant to anyone rendering very large captured sites.

The HTML-in-Canvas API adds a layoutsubtree attribute on canvas, draw calls for child elements (drawElementImage for 2D, texElementImage2D for WebGL, copyElementImageToTexture for WebGPU), and a paint event for updates. The DOM content stays in the accessibility tree and keeps find-in-page, translation and DevTools working. Cross-origin iframes cannot be drawn. PlayCanvas (engine 2.18 and 2.20) and three.js (HTMLTexture in r184) added support during the year.

If it ships, labels, price panels and floor-plan UI in a 3D property viewer could be real HTML inside the scene without losing accessibility.

At Google I/O, Google announced that Gemini CLI would be retired in favour of Antigravity CLI, a Go rewrite that shares its architecture with the new Antigravity 2.0 desktop app and runs multiple agents asynchronously. Agent Skills, Hooks and Subagents carry over. Consumer access, covering the free tier and Google AI Pro and Ultra, stopped on 18 June 2026; Gemini Code Assist Standard and Enterprise licences were unaffected.

Anyone with scripts, CI jobs or MCP setups built on the gemini command had about a month to migrate. The same keynote launched Gemini 3.5 Flash and Managed Agents in the Gemini API.

Three items in Chrome's I/O round-up bear directly on 3D web work. WebMCP, a proposed standard for exposing JavaScript functions and forms as tools to browser agents, moved to an origin trial in Chrome 149. Chrome DevTools for agents gives coding agents, including Antigravity and over 20 others, access to DevTools for debugging and performance checks. HTML-in-Canvas, also in origin trial, puts real DOM elements inside a WebGL or WebGPU canvas.

Together they let an agent inspect a running viewer's performance, let a 3D page expose its controls as agent tools, and allow accessible, searchable HTML labels inside a 3D scene.

Postshot is a widely used desktop tool for training Gaussian splats from photos and video. Version 1.1 adds a built-in VR/XR preview, and photometric compensation that corrects auto-exposure, white point and vignetting across input images. Selection gains lasso and rectangle modes with add, subtract and intersect. It also fixes camera tracking on NVIDIA Blackwell GPUs, improves memory handling for larger scenes, and imports RealityScan 2.1 exports and Pix4D OPF data. The Unreal plugin now supports UE 5.7. For studios training splats of sites or CG-rendered camera sets, the exposure correction and larger-scene handling bear directly on quality and turnaround.

SPZ 4 replaces single-stream gzip with six parallel ZSTD streams, one per attribute, giving 3 to 5x faster encoding and 1.5 to 2x faster end-to-end loading, while files stay about 10x smaller than PLY and 2.5% smaller than before. The 10 million point limit is gone, a plaintext 32-byte header exposes counts without decompressing, and opt-in vendor extensions carry extra metadata. SH quantisation is configurable from 3 to 8 bits, with SH degree 4 supported. The library was tagged v3.0.0 on GitHub the same day. SPZ is the compression proposed for glTF, so these changes feed into the standard.

VkSplat is a 3DGS training pipeline written fully in Vulkan compute rather than CUDA and PyTorch. The authors report 3.3x faster training and 33% lower VRAM use than a CUDA plus PyTorch baseline at matching quality, and show it running across GPU vendors; they describe it as the first fully Vulkan training pipeline at state-of-the-art performance. Code is at harry7557558/vksplat. For studios with mixed hardware, or anyone wanting to train off NVIDIA cards, it is evidence that CUDA lock-in for splat training is loosening.

Cesium now streams Gaussian splats with hierarchical LOD. 3D Tiles acts as the spatial index and glTF carries the payload, using KHR_gaussian_splatting and KHR_gaussian_splatting_compression_spz. It runs in CesiumJS, Cesium for Unreal and Cesium ion, with Unity and other runtimes planned. Cesium quotes SPZ as up to 90% smaller than PLY. Through the year, CesiumJS releases fixed splat tileset crashes above roughly 16 million splats and sped up SPZ decoding, and 1.145 in September added spherical harmonics for SPZ splats.

For site and masterplan work, this is the route to city-scale splat context streamed on a real globe.

DeepSeek V4 came in two open-weight mixture-of-experts sizes: V4-Pro at 1.6T parameters with 49B active, and V4-Flash at 284B with 13B active. Both default to a 1M token context, offer thinking and non-thinking modes, and the API accepts both OpenAI ChatCompletions and Anthropic formats. Legacy deepseek-chat and deepseek-reasoner were retired after 24 July 2026.

The Anthropic-compatible endpoint means Claude Code and similar harnesses can be pointed at DeepSeek by changing a base URL, giving a low-cost option for bulk pipeline tasks. The weights are too large for a laptop but can run on rented GPUs.

Opus 4.7 kept Opus pricing at $5/$25 per million tokens. It accepts images up to 2,576 pixels on the long edge, about 3.75 megapixels and more than three times prior Claude models, and ships a new tokenizer that can use 1.0 to 1.35 times more tokens for the same text. Claude Code added /ultrareview for dedicated review sessions, raised default effort to xhigh and extended auto mode to Max users. Task budgets entered public beta on the API.

The higher image resolution matters for 3D work: full-size viewport screenshots, render comparisons and plan drawings can go to the model without heavy downscaling.

Spark 2.0, World Labs' open source splat renderer for three.js, adds a continuous LoD system that picks a cut through an LoD splat tree so only about 500K to 2.5M splats are drawn per frame regardless of scene size. The new .RAD format is compressed and organised in 64K splat chunks for random access streaming, so a coarse scene appears first and detail near the camera loads next. Demo scenes range from 6M to 106M splats. For web delivery of large sites and masterplans it is a direct alternative to PlayCanvas Streamed SOG.

Lyra 2.0 follows the generative reconstruction route: a camera-controlled video model produces a long walkthrough from a single image, then a feed-forward step lifts it to 3D Gaussians and surface meshes. The paper targets the two failure modes of long-horizon generation, spatial forgetting and temporal drift, so revisited areas stay consistent. A distilled version cuts generation from 35 to 4 steps per frame, a 13x speed-up, and scenes import into Isaac Sim. Weights are on Hugging Face. For visualisation work it is a marker of how close image-to-explorable-scene tools are to usable context and concept environments.

Niantic Spatial turned Scaniverse from a consumer scanning app into a business platform. A new web interface lets teams upload, manage and process captures from the mobile app and from ordinary 360° cameras. Scaniverse then generates meshes, Gaussian splats and location maps for the new VPS 2.0 positioning service. Exports cover FBX meshes and splats in PLY and SPZ, Niantic's open splat format. Paid plans add 360° camera support, and the platform is aimed at areas up to thousands of square metres, including construction sites. For archviz studios, this gives a low-cost capture route for site context and existing interiors that exports straight into splat-aware renderers.

Cursor 3 reorganised the editor around agents. The Agents Window runs several agents at once across repos and environments, agent tabs show chats side by side, /worktree isolates changes in a git worktree, and /best-of-n runs one task across several models in parallel. Design Mode lets you annotate elements in the integrated browser and send them to chat.

For front-end and 3D web work, Design Mode and the built-in browser shorten the loop between seeing a layout or overlay bug and describing it to the agent. Later 2026 releases added subagents, cloud agents that react to events, and Projects for long-running work.

Microsoft released Babylon.js 9.0 on 26 March 2026. On the splat side it loads PLY, .splat, SPZ and SOG/SOGS, adds triangle splatting for opaque mesh-like results, lets splats cast shadows, and can combine several splat assets in one scene. Engine-wide additions include clustered lighting for hundreds of lights, textured area lights, volumetric lighting, a Frame Graph for pipeline control, a node particle editor, animation retargeting, geospatial rendering and the start of OpenPBR support.

With SOG loading, Babylon can now open the same compressed files PlayCanvas tooling produces, which makes it a practical alternative engine for splat viewers.

Auto mode sits between approving every prompt and skipping permissions entirely. A server-side probe screens tool output for prompt injection, and a two-stage transcript classifier decides whether each action can run. Anthropic published its own figures: a 0.4% false positive rate on real traffic and a 17% false negative rate on dangerous actions. It launched as a research preview and later became the default.

This changed how long unattended runs work in practice. A developer can leave an agent building, testing and fixing a pipeline tool without approving each shell command, but the 17% miss rate means deny rules and backups still matter.

NVIDIA's reference Vulkan renderer, shown in the GTC keynote, reached 2026.1 under Apache 2.0. The release adds a multi-instance architecture where splat sets are instanced and reused with global index tables and unified sorting, a bindless scene buffer, point, spot and directional lights with ray traced hard and soft shadows, front-to-back rasterisation with deferred shading, and DLSS Ray Reconstruction for anti-aliasing and denoising. It also has built-in MSE, PSNR and FLIP comparison. For archviz it is a working example of splats that are lit and shadowed alongside other content instead of baked.

Many kernel changes proposed for 3DGS need scenes retrained from scratch. This paper swaps the exponential kernel for a polynomial approximation combined with a ReLU, which stays compatible with existing datasets trained for the original kernel while allowing more aggressive culling of Gaussians. The authors report a 4 to 15% performance improvement across different 3DGS implementations with negligible impact on image quality, and discuss benefits for NPU hardware. Because it works on assets you already have, it is a low-cost speed gain for renderers and viewers rather than a new training pipeline.

NanoGS is a free plugin by Tim Chen, lead technical artist at Moonshine Studio, announced on LinkedIn on 18 March 2026. It breaks a splat scene into LOD clusters in the way Nanite handles geometry. It picks detail by screen-space error, compacts splats and sorts them with a GPU radix sort, so only splats visible to the camera are drawn. Demo footage showed more than four times the frame rate on an RTX 2070. Version 1.0 shipped as compiled binaries, with source promised later. For studios putting large aerial or site captures into Unreal for interactive property experiences, it offers a free path for streaming splats beside Epic's engine features.

This update brought streamed LOD to SuperSplat. Scenes are cut into small SOG chunks that load on demand, so large captures open quickly on phones and fill in on desktops. Walk mode adds click-to-walk first-person navigation backed by a voxel collision system. Easy Upload accepts PLY, SOG, Streamed SOG or LCC files straight from the homepage. Underneath is PlayCanvas Engine 2.17.0, released on 6 March, which added a WebGPU GPU-driven sort and culling pipeline and a scene-wide splat budget.

Chunked SOG streaming with a global splat budget is now the standard PlayCanvas pattern for walkable interiors.

Feed-forward reconstruction models such as VGGT grow quadratically in cost with the number of images. VGG-T³ distils the key-value representation of scene geometry into a fixed-size MLP through test-time training, so compute and memory scale linearly with view count. The paper reports reconstructing 1,000 images in 54 seconds, an 11.6x speed-up over softmax attention baselines, and can localise unseen images against the result. Accepted to CVPR 2026 with code at nv-dvl/vgg-ttt. It matters as a candidate replacement for slow COLMAP runs on large architectural capture sets before splat training.

PackUV maps every Gaussian attribute into a sequence of structured UV atlases, so dynamic splat sequences become image-native data compatible with standard video codecs such as FFV1 without quality loss. Its fitting method, PackUV-GS, optimises Gaussians directly in UV space with flow-guided labelling of dynamic Gaussians, and scales to sequences of up to 30 minutes. The team also released PackUV-2B, a dataset of 100 sequences from more than 50 synchronised cameras. Accepted to CVPR 2026, it points to 4D splats delivered through existing video streaming infrastructure rather than bespoke formats.

Foundry's Nuke 17.0 rebuilt the compositor's 3D system on USD and added native Gaussian splats that can be imported, edited, rendered and exported inside a comp. Projections, lighting and rendering now work non-destructively through USD, and the release also brings the BigCat machine learning node and Field Nodes for 3D data. For archviz studios that finish CG building renders over drone plates, splat context can now be composited in the same tool, with no round trip through a separate viewer. Nuke 17.1 followed on 20 August 2026 with animated splat sequences, basic relighting from Direct, Point and Spot lights, and SPZ and USD export.

Chrome 146 adds compatibility mode. You pass featureLevel: "compatibility" to requestAdapter() and get a slightly restricted subset of WebGPU that can run on older graphics APIs such as OpenGL ES 3.1. Android comes first, and the team is looking at ChromeOS and Direct3D 11 on Windows. The release also adds a TRANSIENT_ATTACHMENT texture usage that keeps render pass attachments in tile memory. Across 2026, Chrome also widened WebGPU on Linux: Intel Gen12+ from Chrome 144, then NVIDIA on Wayland from Chrome 147.

For mobile splat viewers, compatibility mode reaches older Android GPUs that would otherwise be stuck on WebGL 2.

Copilot CLI is GitHub's terminal coding agent, in public preview since September 2025. At general availability it gained plan mode and an autopilot mode that runs tasks end to end, support for MCP servers, plugins, custom agents and Markdown skills, /diff and /review commands, automatic history compression, and repository memory across sessions. Models from Anthropic, OpenAI and Google can be switched with /model.

For a solo developer already paying for Copilot, this puts a Claude Code style agent in the terminal at no extra cost, and skills and MCP configuration written for other agents largely carry over.

Pixar's OpenUSD 26.03 makes Gaussian splats a first-class USD prim type through the new UsdVolParticleField schema family. ParticleField3DGaussianSplat is the first concrete schema, and the family is designed to extend to other radiance field methods. The release ships a reference Hydra renderer, hdParticleField, and an example script that converts Gaussian PLY files to USD. It also adds WebAssembly builds. For studios with USD pipelines between 3ds Max, Houdini, Nuke and real-time engines, captured context can now travel as ordinary USD data instead of a sidecar file. Vendors such as Nuke and V-Ray built their splat support around USD during the year.

Lumion 2026.0 is a workflow release for one of the most widely used architectural real-time renderers. Area Placement scatters up to 5,000 objects across large outdoor areas with controls for density and variation. Fill-in Clipping Plane caps cut geometry so sections and cutaways look solid. The AI Upscaler gains a 4x mode for renders up to 15,360 by 8,640 pixels, and LiveSync now syncs PBR materials live from SketchUp and Revit. The library adds 31 photogrammetry trees and 73 exterior objects, and a project can now hold up to 300 panoramas. The release has no Gaussian splat support, which leaves Lumion behind Chaos, D5 and SketchUp on captured context.

Gemini 3.1 Pro replaced Gemini 3 Pro as Google's top reasoning model. Google reported a verified 77.1% on ARC-AGI-2, more than double Gemini 3 Pro, and highlighted code generation for animated SVGs and interactive designs. The Google Cloud post also cited improved understanding of 3D transformations and a 3D animation pipeline use case.

It shipped in preview on the Gemini API, Gemini CLI, Antigravity and Android Studio at once. For 3D web developers it was the first Gemini model where Google explicitly pitched spatial and animation code as a strength, which makes it worth testing on transform maths and scene setup code.

SuperSplat Studio is a new hosted app that works on splats you have already published to SuperSplat. You can place annotations, each holding a title, a description and a saved camera viewpoint; clicking one moves the camera to that view. You can also apply engine post effects: bloom, sharpen, vignette, colour grading and chromatic fringing. You open a scene in Studio from the Manage page.

For property work this gives a no-code way to build a guided splat tour with room callouts. It also shows the hotspot pattern PlayCanvas expects people to rebuild in their own viewers.

GPT-5.3-Codex was OpenAI's February coding model for Codex, aimed at repository search, terminal commands and debugging over long runs. GitHub rolled it out to Copilot Pro, Pro+, Business and Enterprise on 9 February, reporting up to 25% faster performance than GPT-5.2-Codex on agentic coding tasks, and later made it a long-term support model for business plans.

It matters mainly because it arrived the same day as Opus 4.6 and became available across VS Code, Copilot CLI and the Copilot coding agent, so developers on a single Copilot subscription could compare the two vendors' coding models on the same tasks without switching tools.

Opus 4.6 raised the ceiling for long agentic coding sessions. It added a 1M token context window (beta, Claude Platform), 128K output tokens, adaptive thinking with four effort levels, and context compaction in beta for long runs. Claude Code gained agent teams as a research preview, so several agents can work on a task in parallel. Standard pricing stayed at $5/$25 per million tokens, with a premium above 200K tokens.

For 3D web and pipeline work, the larger context means an agent can hold a viewer codebase, shader files and build scripts at once, and effort levels let you trade cost against depth per task.

WebSplatter is an end-to-end GPU rendering pipeline for browsers. It works around WebGPU's lack of global atomics with a wait-free hierarchical radix sort that runs deterministically across hardware, and adds an opacity-aware culling stage that prunes splats before rasterisation to cut overdraw and peak memory. The authors report 1.2x to 4.5x speedups over state-of-the-art web viewers. It is a useful reference for anyone tuning WebGPU splat viewers on mixed client hardware, where sorting and overdraw decide whether a large interior or site model holds frame rate on a phone.

Qwen3-Coder-Next is an open-weight model trained for coding agents rather than general chat. It has 80B total parameters but activates only 3B per token, with a native context of 262,144 tokens, and the model card lists harness support for Claude Code, Qwen Code and Cline, plus tool-call parsing for vLLM.

For a solo developer this is the most practical local option of the year. Quantised builds run on a well-specced laptop through LM Studio or similar, so you can point an agent harness at a private model for offline work or for pipeline scripts you would rather not send to a cloud API. The Apache 2.0 licence permits commercial use.

The Khronos Group released KHR_gaussian_splatting as a release candidate for public feedback. It defines uncompressed, GPU-ready splat attributes (position, rotation, scale, opacity, spherical harmonics) as glTF point primitives, is designed to be algorithm-agnostic and extensible, and falls back to a point cloud in viewers that cannot splat. Compression extensions were proposed alongside it, including Niantic Spatial's SPZ and Qualcomm's L-GSC. For architectural work this is the first route to shipping splats and meshes in one standard container, with CesiumJS 1.139 and Cesium for Unreal 2.23.0 adding support in March.

PPISP replaces per-frame latent codes and affine colour fixes with a physically based ISP module: per-frame exposure and colour correction, per-camera vignetting and camera response. A controller trained on the input views predicts ISP parameters for new viewpoints, much like auto exposure and white balance. Code is at nv-tlabs/ppisp, gsplat integrated it in January as an alternative to the bilateral grid, and LichtFeld Studio v0.4.1 shipped it. For architectural captures with mixed exposures, drone plus handheld sets, or sky-heavy exteriors, this targets the colour blotches and floaters that inconsistent photometry causes.

Poimandres published the first public alpha of React Three Fiber v10, together with an alpha of Drei 11. R3F can now drive either WebGLRenderer or WebGPURenderer, so state.gl becomes state.renderer. There are new TSL hooks (useUniforms, useNodes, useLocalNodes, usePostProcessing), and a rewritten scheduler lets useFrame run outside <Canvas />. Later alphas (up to alpha.5 on 8 September) kept adding WebGPU compute helpers. The stable line is still v9 (9.8.1 in September).

All v10 features are marked experimental. If your R3F property sites use custom shaders, the migration guide is worth reading now, before you plan a WebGPU port.

PlayCanvas released SplatTransform v1.0.0, moving the open source splat converter from a CLI-only tool to a library with a programmatic API (readFile, writeFile, processDataTable) that works in both Node.js and browsers, with URL, memory and ZIP file systems and a new --summary action. Later majors built on it: v2.0.0 (27 April) added voxel grids, floor fill and GPU floater filtering, and v3.0.0 (9 July) rewrote the pipeline so memory is bounded by chunk size, not scene size. For web pipelines this is the standard route from PLY to SOG and Streamed SOG, now scriptable inside your own tooling.

D5 Render 3.0 is the next major version of the real-time GPU renderer that links to SketchUp, Rhino, Revit, 3ds Max, Blender and Cinema 4D. It adds an Ocean object for animated water, a Volumetric Clouds system with presets, procedural buildings in City Generator and AI features such as Scene Match lighting from a text prompt. D5 Lite, a lightweight SketchUp companion for early design, launched alongside it. D5 could already turn rendered scenes into splats for its XR Tour web viewer. 3.0 goes the other way too, with experimental import of .ply and .gs splats as scene elements, so studios can drop captured context in beside their modelled buildings.

Firefox 147 shipped on 13 January 2026. Its WebGPU change is short but wide: the API is now enabled on all macOS versions running on Apple Silicon, where Firefox 145 had limited it to macOS Tahoe. Firefox has had WebGPU on Windows since version 141. Linux is still Nightly only and Android is still behind a flag, according to the gpuweb implementation status page.

For splat viewers this means a compute-based WebGPU path now reaches Firefox users on most modern Macs. You still need a WebGL 2 fallback for Linux and Android Firefox.