As accurate as GPT‑6.1 Sol at finding objects, 982x faster

Name the things you want found, such as cars, people or trees, and get back a box around each one. SIE runs Meta's open SAM 3 detector for you behind one API call.

Get started Run in your cloud

SIE SAM 3

16 of 17 vehicles boxed

Asked for car, bus, truck

GPT‑6 Luna boxes 46% of the objects, Claude Sonnet 5.5 39%

GPT‑6 Luna, high reasoning

6 of 15 vehicles boxed

Asked for car, bus, truck

A drone photo of a motorway interchange with slip roads: a part of it with 15 vehicles, and the 14 boxes GPT-6 Luna at high reasoning returned there. — 14 boxes in this view; 6 cover a vehicle by half or more, the other 8 sit partly off a vehicle. Whole photo: 16 of 41 vehicles, in 63.5 s.

Claude Sonnet 5.5, default settings

0 of 11 vehicles boxed

Asked for car, bus, truck

A drone photo of a road junction with queues of cars: a part of it with 11 vehicles, and the 24 boxes Claude Sonnet 5.5 at default settings returned there. — 24 boxes in this view, none touching a vehicle. Whole photo: 0 of 42 vehicles, in 6.8 s.

Across 300 crowded photos, GPT‑6 Luna at high reasoning boxed 46% of the objects, Claude Sonnet 5.5 39% and Claude Haiku 5.5 34%. SIE SAM 3 boxed 73%.

NBC 5 Dallas-Fort Worth Dallas AI cameras missed a license plate they were meant to blur; the city covered it by hand
The city "had to add that black box to redact the plate number because it was still visible in the original image."
23 September 2026
arXiv:2607.06420 In crowded scenes, GPT-5.5 counted exactly right 1 time in 5 and Claude Opus 4.8 1 time in 100
HoloCount: A Holistic Visual Counting Benchmark for MLLMs

Deng, Qiao and Wan

“These results confirm that dense scenes cause systematic under-counting”
7 July 2026

More accurate than every model tested; only GPT‑6.1 Sol comes close

Run this eval

Objects boxed correctly

F1 F1 F1 F1 F1 ↑

SIE SAM 3

SIE SAM 3 0.62

GPT‑6.1 Sol high

GPT‑6.1 Sol high 0.60

Claude Sonnet 5.5

Claude Sonnet 5.5 0.39

GPT‑6 Luna high

GPT‑6 Luna high 0.36

Claude Haiku 5.5

0.31 Claude Haiku 5.5

GPT‑6.1 Sol high

GPT‑6.1 Sol high 0.92

SIE SAM 3

SIE SAM 3 0.87

Claude Sonnet 5.5

Claude Sonnet 5.5 0.81

GPT‑6 Luna high

GPT‑6 Luna high 0.76

Claude Haiku 5.5

Claude Haiku 5.5 0.75

GPT‑6.1 Sol high

GPT‑6.1 Sol high 0.81

SIE SAM 3

SIE SAM 3 0.77

GPT‑6 Luna high

0.42 GPT‑6 Luna high

Claude Haiku 5.5

0.25 Claude Haiku 5.5

Claude Sonnet 5.5

0.15 Claude Sonnet 5.5

GPT‑6.1 Sol high

GPT‑6.1 Sol high 0.50

SIE SAM 3

SIE SAM 3 0.49

Claude Sonnet 5.5

Claude Sonnet 5.5 0.32

GPT‑6 Luna high

0.18 GPT‑6 Luna high

Claude Haiku 5.5

0.16 Claude Haiku 5.5

SIE SAM 3

SIE SAM 3 0.35

Claude Sonnet 5.5

Claude Sonnet 5.5 0.27

GPT‑6.1 Sol high

0.19 GPT‑6.1 Sol high

Claude Haiku 5.5

0.08 Claude Haiku 5.5

GPT‑6 Luna high

0.08 GPT‑6 Luna high

Time and cost

seconds per photo USD per 1,000 photos ↓

SIE SAM 3

0.083 s SIE SAM 3

Claude Haiku 5.5

4.1 s Claude Haiku 5.5

Claude Sonnet 5.5

5.6 s Claude Sonnet 5.5

GPT‑6 Luna high

35.3 s GPT‑6 Luna high

GPT‑6.1 Sol high

GPT‑6.1 Sol high 80.7 s

SIE SAM 3

$0.25 SIE SAM 3

Claude Haiku 5.5

$0.76 Claude Haiku 5.5

GPT‑6 Luna high

$2.95 GPT‑6 Luna high

Claude Sonnet 5.5

$13.27 Claude Sonnet 5.5

GPT‑6.1 Sol high

GPT‑6.1 Sol high $59.98

Crowded-photo evaluation

300 crowded photos, 17,222 objects, four current OpenAI and Anthropic models

Photos: Traffic, parking lots, people and trees from drones and aircraft, four public benchmarks, 21+ objects each
Scoring: A box counts when its name is right and it covers the object by half. F1 is 1 when every object is boxed and no box is false. All is the mean of the four sets
Settings: SAM 3 at its shipped threshold. GPT‑6.1 Sol and GPT‑6 Luna at high reasoning, Claude at default settings, each photo at full resolution. GPT‑6 Luna at default reasoning scored the same as at high
Time: Median per photo. SIE timed on a GPU server (RTX PRO 6000) with the client alongside; the others over their public APIs from one client in San Francisco
Cost: Per 1,000 photos: SAM 3 $0.25 at SIE’s Serverless rate, GPT‑6.1 Sol $59.98. The others are at list price on the tokens each used

Send a photo and the names to count, get a box for each one

from pathlib import Path
from sie_sdk import SIEClient
client = SIEClient(api_key="API keysk-sie-…")
image = {"data": Path(imagejunction.jpgbrowse).read_bytes(), "format": "jpeg"}
result = client.extract(
    "modelfacebook/sam3",
    {"images": [image]},
    labels=[
        "labelcar",
        "labelbus",
        "labeltruck",
    ],
)
for obj in result["objects"]:
    print(obj["label"], obj["score"], obj["bbox"])

import { readFile } from 'node:fs/promises';
import { SIEClient } from '@superlinked/sie-sdk';

const client = new SIEClient(undefined, { apiKey: 'sk-sie-…' });

const image = await readFile("junction.jpg");
const result = await client.extract(
  'facebook/sam3',
  { images: [image] },
  { labels: ["car","bus","truck"] },
);
console.log(result.objects);

images_bytes=$(base64 < 'junction.jpg' | tr -d '\n')
curl https://api.superlinked.com/v1/extract/facebook%2Fsam3 \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"images\":[{\"data\":\"$images_bytes\",\"format\":\"jpeg\"}]}],\"params\":{\"labels\":[\"car\",\"bus\",\"truck\"]}}"

Build the "Detect" capability into my app using the Superlinked Inference Engine (SIE).

Context
- SIE is an OpenAI-style inference API. Python SDK: `from sie_sdk import SIEClient`; TypeScript: `@superlinked/sie-sdk`.
- Base URL: https://api.superlinked.com (or my regional endpoint). Auth: Bearer key from env `SIE_API_KEY` (never hard-code it).
- Model: facebook/sam3 (SIE primitive: /extract). Keep the model id configurable.

Task
- Input: an uploaded image plus the object labels to find.
- Behaviour: return the named objects with confidence scores and bounding boxes
- Call the selected SIE primitive once per request and map the response into your domain type.

Deliverables
- A typed client wrapper, an application-level function for this task, error handling for timeouts/empty input, and unit tests with a stubbed client.
- Wire it into my existing stack (ask me which framework if unclear) and add a short usage example.

Run

Output

51 objects returned for car, bus, truck

A drone photo of a road junction with queues of cars, with the 51 boxes SAM 3 returned for car, bus, truck.

label	score	bbox [x, y, w, h]
car	0.885	[3267, 363, 49, 46]
car	0.877	[3354, 459, 45, 50]
car	0.875	[3544, 1203, 49, 42]
car	0.872	[2370, 446, 51, 56]
car	0.860	[2722, 1377, 56, 31]
car	0.857	[282, 787, 60, 29]

and 45 more, highest score first. Score is SAM 3's own match score, not a probability.

Start finding objects in your photos

Serverless API

We host the models; you call one API.

No idle GPUs, pay for what you use
SOC 2 Type 2, US or EU data residency
$50 of credit to start, no card required

Get started

Self-hosted (Kubernetes)

SIE in your clusters, backed by our engineers.

Data stays in your own cloud account
Same models and API as Serverless
Terraform to your cloud in minutes

Talk to us Deploy guide

Run locally

The open-source server on your own machine.

docker run --gpus all -p 8080:8080 -e HF_TOKEN ghcr.io/superlinked/sie-server:latest-cuda12-transformers5

Quickstart

Switch to SIE in 5 minutes

from openai import OpenAI

client = OpenAI()
resp = client.chat.completions.create(
    model="gpt-6-luna",
    messages=messages,
    response_format=schema,
)
objects = json.loads(resp.choices[0].message.content)["objects"]

from google import genai

client = genai.Client()
resp = client.models.generate_content(
    model="gemini-3.8-flash",
    contents=[image, prompt],
    config=config,
)
objects = json.loads(resp.text)

import boto3

client = boto3.client("rekognition")
resp = client.detect_labels(Image={"Bytes": image_bytes})
# Rekognition's own label list; boxes only for the labels it knows
objects = [
    (label["Name"], box["BoundingBox"])
    for label in resp["Labels"]
    for box in label.get("Instances", [])
]

SIE

from sie_sdk import SIEClient

client = SIEClient(api_key="YOUR_SIE_API_KEY")
result = client.extract(
    "facebook/sam3",
    {"images": [image_bytes]},
    labels=["car", "bus", "truck"],
)
objects = result["objects"]

from sie_sdk import SIEClient

client = SIEClient(api_key="YOUR_SIE_API_KEY")
result = client.extract(
    "facebook/sam3",
    {"images": [image_bytes]},
    labels=["car", "bus", "truck"],
)
objects = result["objects"]

from sie_sdk import SIEClient

client = SIEClient(api_key="YOUR_SIE_API_KEY")
result = client.extract(
    "facebook/sam3",
    {"images": [image_bytes]},
    labels=["car", "bus", "truck"],
)
objects = result["objects"]