Model releases, product launches, research notes, and corrections. We publish the corrections too — the July 2026 benchmark re-audit is on this page because a result we had already claimed turned out to be measuring our own harness.
Work underway, listed separately from the feed above because none of it has shipped. If it moves to the feed, it is live.
Need logos, product shots, or fast answers? Reach out and we will respond with assets tailored to your coverage.
Download the media kit17 updates · newest first
Product
September 5, 2026
POST /api/mcp is a Model Context Protocol server. Add the URL and your API key to Claude Code, Cursor, or any MCP client and extract_cad, extract_cad_from_url, rowe_respond, list_models, and list_supported_formats appear as tools. Calls are forwarded to the REST routes, so billing and limits are identical.
Product
September 5, 2026
POST /api/v1/cad/extract now accepts .drawio and draw.io .xml files. Each node keeps its absolute bounding box and container, each connector keeps its source, target, and direction, and the extraction adds derived spatial relations and a graph summary. An agent gets the diagram, not a caption of it.
Research
September 2, 2026
The rendered-image benchmark we listed as in progress has run to completion: 30 held-out cases, hunarbatra/SpatialThinker-3B on CPU, 4 scoring conditions, 1 seed. On the fair condition the competitor scored 16 / 30 overall and 10 / 21 on the capabilities that are not decided by vocabulary, the same as on the prior JSON prompt. Rowe's 30 / 30 on the same cases is a self-check against a deterministic geometry calculator, not a comparison. The confidence interval, the per-capability audit table, and a validity note on the invalid 12-case July attempt are on the page.
Early Access
A spatial reasoning model for 3D data, physical constraints, and object relationships.
Product
August 26, 2026
status.fourechelon.com now reports live probes for the web app, the CAD converter, and the Rowe runtime. The latency figure is a single probe sample and is labelled as one — it is not an SLA, and we would rather say so than let it be read as a guarantee.
Product
August 25, 2026
POST /api/v1/cad/extract is live. Upload STEP, IGES, DXF, PDF, or SVG and get four-cad.extraction.v1 back — B-rep solids and faces with exact volumes, drawing entities, measured dimensions, and annotations. No OCR anywhere in the path. Verified end to end on a box-with-hole STEP part: volume exact to 11497.3 mm³, cylindrical face radius 4.
Early Access
A spatial reasoning model for 3D data, physical constraints, and object relationships.
Company
August 25, 2026
A single page listing what is proven live today — health endpoints, CAD extraction, cad_scene inference, partner feature-tree import — next to what is explicitly not claimed: vendor superiority scores, latency SLAs, and native SolidWorks parsing. If a claim is not on that page, we are not making it.
Research
July 31, 2026
We re-audited our spatial benchmark and found the comparison was measuring our harness, not the models. A 64-token generation cap was truncating competitor outputs to as little as 7 characters, one model was mis-loaded entirely, and three of ten capabilities were unwinnable by construction. We raised the cap to 1024 tokens, rescored everything, published a scoring-validity notice on the results page, and pinned the correction in tests. The honest positioning is auditable metric geometry with error bars — not 'beats X'.
Early Access
A spatial reasoning model for 3D data, physical constraints, and object relationships.
Product
July 31, 2026
Self-serve billing is live end to end: Stripe-hosted Checkout, a base subscription that enables keys, and metered input/output spatial tokens on successful requests only. Failed requests are recorded for debugging and never billed. Current rates are always on the pricing page.
Product
May 2, 2026
An early preview of Rowe 1.0, a spatial reasoning model for point clouds, CAD files, maps, images, and sensor context.
Research
April 15, 2026
Shifted our core work toward models that treat 3D geometry, physical constraints, and spatial relationships as first-class inputs and outputs rather than as text to be captioned.
Milestone
March 13, 2026
Y Combinator ranked Four Echelon in the top 10% of P26 applicants for our work on 3D spatial mapping and physical-world intelligence.
Community
March 7, 2026
Spent time with builders pushing forward AI, infrastructure, and applied intelligence at the YC x Google DeepMind hackathon.
Product
February 9, 2026
Released a product video highlighting realtime spatial intelligence for live 3D mapping and enterprise operations.
Company
January 15, 2026
Announced the Four Echelon brand, reflecting our mission to build spatial intelligence systems for physical-world understanding.
Milestone
December 11, 2025
Selected as a top 10% applicant in the Y Combinator W26 batch for innovation in spatial intelligence and 3D mapping.
Milestone
November 6, 2025
Joined the Y Combinator Vibecon cohort alongside founders building the future of applied AI and enterprise infrastructure.
Product
October 15, 2025
An updated mobile stack with on-device scanning, improved indoor-outdoor handoff, and richer event timelines for teams on the ground.