Agnify
LogisticsDocks, forklifts, pallets and trailers
ManufacturingUptime, stations and material flow
Metal fabricationCutting, forming, welding and the waits between
Oil & gasFlares, pads, rigs and plants
Equipment monitoringMachine states, changeovers and part checks
ExamplesInteractive runs on real video
Examples VLMs Blog
Sign in Book a demo
Agnify

Industries

LogisticsDocks, forklifts, pallets and trailers
ManufacturingUptime, stations and material flow
Metal fabricationCutting, forming, welding and the waits between
Oil & gasFlares, pads, rigs and plants
Equipment monitoringMachine states, changeovers and part checks
Examples VLMs Blog Sign in
Book a demo

Articles

Benchmarks, experiments and product notes from the Agnify team.

HOTA against API cost per hour of video, log scale. Gemini 4 Hydrogen 0.601 at $18.47, Gemini 4 Carbon 0.602 at $29.20 from 1 December ($14.60 until then), GPT-6.1 Terra 0.617 at $8.19, GPT-6.1 Sol 0.616 at $19, GPT-6 Astra 0.622 at $105, Claude Opus 5.5 0.594 at $43, Gemini 4 Argon 0.593 at $32.10 from 1 December ($16.05 until then), Gemini 3.7 Flash 0.570 at $3.70.Article

EngineeringLatest

Gemini 4 Carbon and Hydrogen fix Argon's shell game, mostly

Gemini 4 Carbon and Hydrogen make about a third as many shell-game mistakes as Gemini 4 Argon and tie the GPT-6.1 models. Hydrogen gets there by thinking much longer, at five times the cost of Gemini 3.7 Flash.

Anton Gritsay
Anton Gritsay
Nov 13, 2026 · 9 min read

More posts

AllArticlesVideos
HOTA against API cost per hour of video, log scale. GPT-6.1 Terra 0.617 at $8.19, GPT-6.1 Sol 0.616 at $19, GPT-6 Astra 0.622 at $105, Claude Opus 5.5 0.594 at $43, Gemini 4 Argon 0.593 at $16, Gemini 3.7 Flash 0.570 at $3.70.Article

EngineeringOct 30, 2026 · 6 min read

GPT-6.1 Terra matches Sol at less than half the price

GPT-6.1 Terra scores the same as GPT-6.1 Sol on grounded video QA for 43% of the cost. Most of what it gives up comes from following one instruction to the letter.

Read post →
Four views of the same frame, a car flying off a vertical wall on thrusters. SAM3 boxes it as car 0.97; RF-DETR L, D-FINE X and D-FINE-seg X, with only the car class enabled, find no car.Article

EngineeringOct 19, 2026 · 6 min read

When a car climbs a wall, only SAM3 still sees a car

Three fast COCO detectors and SAM3 watch a car drive up a wall and fly. On the ground the detectors find it; on the wall they lose it on most frames. SAM3 keeps it on 97%.

Read post →
Per-family HOTA difference, Gemini 4 Argon minus GPT-6.1 Sol, with 95% intervals. Argon is ahead on containers and pouring (+0.034) and behind on shell games (−0.049); −0.023 over all 300 questions.Article

EngineeringOct 14, 2026 · 7 min read

Gemini 4 Argon loses its points in the shell game

Gemini 4 Argon ties Claude Opus 5.5 on grounded video QA at about a third of the cost. It trails the GPT-6 models mostly on shell games.

Read post →
Paired HOTA difference against Gemini 3.7 Flash with 95% intervals: GPT-6 Astra +0.053, Claude Opus 5.5 +0.025, Gemini 3.7 Flash rerun +0.001, Qwen3.8-27B −0.089Article

EngineeringSep 30, 2026 · 6 min read

GPT-6 Astra takes the lead on grounded video QA

A small change to how pipeline stages refer to objects put GPT-6 Astra clearly ahead of Gemini 3.7 Flash, at 28 times the cost.

Read post →
Gemini and self-hosted Qwen answer the same shell-game grounding question; Gemini selects the annotated cup while Qwen selects the identical cup beside itArticle

EngineeringSep 14, 2026 · 9 min read

Detailed comparison of Gemini and Qwen on complex visual grounding tasks

Gemini is marginally better at video reasoning, but both are on par in grounding. We show that pipeline design still matters the most.

Read post →
Qwen3.8-27B general multimodal intelligence scores from the model card, covering MathVision, BabyVision, CharXiv, OmniDocBench, RealWorldQA and ERQAArticle

EngineeringAug 17, 2026 · 6 min read

Qwen3.8-27B is now live at Agnify.ai

The benchmark tables, what the visual gains actually average out to, and the latency we paid for them on our own hardware.

Read post →
D-FINE-seg semantic segmentation predictions on Cityscapes street scenesArticle

EngineeringJul 17, 2026 · 7 min read

D-FINE-seg adds semantic segmentation: a DETR model benchmarked against YOLO26 and RF-DETR

D-FINE-seg now supports three tasks on the same DETR-family architecture — and on Cityscapes it benchmarks better than YOLO26 and RF-DETR.

Read post →
Agnify

A 24/7 AI operator that turns your camera feeds into operational insights

info@agnify.ai

Industries

  • Manufacturing
  • Metal fabrication
  • Logistics
  • Oil & gas
  • Equipment monitoring

Explore

  • Examples
  • VLM tracker
  • FAQ
  • Blog

Company

  • Book a demo
  • Free feasibility check
  • Sign in

Legal

  • Terms of Service
  • Privacy Policy
  • Acceptable Use
© 2026 Agnify LinkedInX