Matrix overall
100.0
AVG 61.1 across matrix cells
ActiveModel portfolio
DeepSeek-V3
Tap a star to rate
DeepSeek · Released 2024-12-26 · v2024-12-26 · Closed · Context 128K tokens
Cadle Score
132.3
#192 in index
Benchmark average
61.1
Not enough benchmark results yet
Pricing
$1.14
$4.56 / 1M out · 17.5 pts/$
MODEL MATRIX · TRUTH CENTER CELLS
Overall
100.0
Intelligence
100.0
Coding
49.2
Reasoning
51.5
Language
83.3
Knowledge
82.9
Agentic
47.4
MMLU
82.9
GPQA
51.5
Math
64.8
LiveCodeBench
49.5
HumanEval
82.6
Reason
93.7
ARC-AGI
—
Coverage AVG
61.1
Pillar Coverage:7 / 7 Core Pillars (100%)
Cadle Cap ranks models with within-benchmark percentiles so ARC-AGI ~40 and MMLU ~90 never compete raw. Coverage AVG is an importance-weighted mean of filled expert cells only — blanks are skipped, not zeroed.
ANALYSIS · TRUTH CENTER
Hard benches (GPQA · ARC-AGI · LCB · Agentic) sit left — a mid-40 on ARC-AGI can still be frontier. Broader benches on the right are easier and often higher.
CADLE INSIGHTS
Only measured signals
Benchmark avg
61.4
22 tracked scores
Strongest signal
epoch-eci
Score 132.3
Blended $/1M
$5.70
In $1.14 · Out $4.56
Value index
17.5
Quality ÷ blended $/1M (higher is better)
Context
128K
Provider context window (not max output)
Coding (matrix)
49.2
Model Matrix coding cell
Agentic
47.4
Tool-use / agentic matrix cell
SPEED · COST · QUALITY
Measured matrix and verified gateway observations only.
- Cadle Score
- 132.3
- Overall (Matrix)
- 100.0
- Cadle Intelligence
- 100.0
- Matrix Coding
- 49.2
- Input $/1M
- $1.14
- Output $/1M
- $4.56
- Value pts/$
- 17.54
DESCRIPTION + DETAILS
PROVIDER
DeepSeek
RELEASE
2024-12-26
CONTEXT
128K tokens
STATUS
ACTIVE
COMMUNITY REVIEWS
Community Reviews & Field Notes
Verified EngineersReal operator evaluations, prompt robustness observations, and deployment notes for DeepSeek-V3.
4.5(2)
Leave a Field Observation
Rating:5 - Frontier benchmark
Role:
Tag topic:
Reviews are publicly visible to the engineering community.
Inference OperatorProduction Deployer Verified
Nov 1, 2026
Tested DeepSeek-V3 on high-concurrency tool loops. Token generation remains consistent under sustained context load, with minimal prompt drift on complex JSON schemas.
Autonomous Agent LeadEval Maintainer Verified
Oct 28, 2026
Evaluated DeepSeek-V3 on multi-turn code synthesis. Code reasoning and diff application succeed on ~88% of test suites. Best paired with strict system prompt guardrails.
MODEL COMPARISON
| Metric | |
|---|---|
| Provider | DeepSeek |
| Release | 2024-12-26 |
| Input / 1M | $1.14 |
| Output / 1M | $4.56 |
| Context | 128K tokens |
| Cadle Cap | 132 |
| Cap board | Cadle Language Cap 132 |
| gpqa-diamond | 59.1 |
| swe-bench-verified | 49.2 |
| math-level-5 | 90.2 |
| training-compute-flops | 24.5 |
| livecodebench | 49.5 |
| math-500 | 90.2 |
| humaneval | 82.6 |
| winogrande | 70.4 |
| mmlu | 82.9 |
| hellaswag | 85.2 |
| aider-polyglot | 48.4 |
| frontiermath-2025-02-28-private | 3.0 |
DATA PROVENANCE
Scores may use different scales — check the Benchmarks tab for source links.
- Supabase models
- Cadle public observations
SOURCE TRACKING
Evidence linked to this model — listing, hub, benchmarks, or live gateway prices.
API
DeepSeek listing
MEASURED CELLS
Same Truth Center numbers as the datasheet above. Accuracy, reliability, safety, and cost scores are not invented.
Overall100
AA IQ100
Coding49.2
Agentic47.4
MMLU82.9
GPQA51.5
Math64.8
LiveCodeBench49.5
HumanEval82.6
TECHNICAL FIELD GUIDESAll 10 Guides →