← Back to dashboard
BigCode·Code — weighted competency·Tue 2 Jun 2026

StarCoder 3 15B

StarCoder 3 15B scores 58/100 competency for Code at $0.35.

Low confidenceOpen weights
Competency
58/100
category rank #8
Cost
$0.35/ 1M tokens
blended
Speed
90/100
Fast
Accuracy
4.6% halluc.
High confidence

Score history.

Weighted Code competency · last 12 weeks · weekly snapshot.

58/100
— flat
58

What it's actually good at.

Score on each business-task subset that goes into the weighted competency.

Test Generation10% of category weight
66
Code Generation40% of category weight
54
Review & Bug Fixing30% of category weight
58
Full-Repo Understanding20% of category weight
62

Best for. Watch out for.

✓ Best for
  • Strong at Code Generation for BigCode.
  • Strong at Review & Bug Fixing for BigCode.
⚠ Watch out for
  • Accuracy 4.6% — verify outputs.
  • Validate fit for your specific workload before committing.

Pricing.

Published usage pricing. Figures refresh daily from our pricing sources.

$0.35

Usage pricing

$0.35/1M tok
Input
$0.35 / 1M tokens
Output
$1.05 / 1M tokens

Self-hosting cost.

Open weights (15B parameters) — what it actually takes to run this in-house, beyond the per-token price. Estimated from parameter count; 4-bit is the practical default.

Minimum hardware

RTX 4090 24GB

$1,800to buy
VRAM (4-bit)
~9 GB
VRAM (fp16)
~35 GB
Or rent in the cloud

≈ $0.50/hr

$365/mo if always on
Best for
steady, high-volume, data-in-house workloads

Where it sits.

Position against the rest of the category. Ranked by weighted competency.

RankModelScoreCostSpeedΔ vs StarCoder
01GPT-5 Code79$14.0074 / 100+21
02DeepSeek V3.1 Coder64$0.676 / 100+6
03Qwen 3 Coder48$1.4078 / 100-10