PUBLIC CALIBRATIONCSL-Meter Public Observatory — PUBLIC DATA - entries labeled Simulated are simulated baselines; entries labeled Server-Verified are real probe records. Golden baselines are being calibrated. (konspace.cc)
CSL-Meter
CSL-Meter · Independent Model Consistency Observatory
Consistency, Substitution & Latency (konspace.cc)
INDEPENDENT TELEMETRY ANALYSIS OF AI PLANS · KONSPACE.CC

Independent analysis of AI subscription & API plans

An independent public observatory for AI subscription & pay-as-you-go API plans
Sixteen real upstream API probes per round fingerprint each plan's response distribution, measuring the JS divergence between the claimed model and the served one (the CSL index), 14:00–18:00 peak drift and live throughput. Only reproducible readings are published; credentials are never stored.
Site-wide 7d Median CSL 12.6
Best Baseline Fit (Low Drift) 8.4 (OpenAI)
Highest Peak Drift (High Drift) 45.7 (OpenCode)
Verified Real-API Samples 4 / 15 targets
BENCHMARK · 24H PEAK & 30D DRIFT
24h peak slices (14:00–18:00 UTC+8) reconciled against 30d version-drift series
↗
METHODOLOGY · 16-PROBE JSD FINGERPRINT · DEMO-SIMULATED
16 real requests per round (hard cap 30K tokens) · simulated baselines strictly separated from real probes · credentials burned in memory (0 stored)
↓
All Targets (15) China Personal Plans Global Coding Subscriptions Global OA + Grok Official Pay-As-You-Go
Data Scope:
Incl. Simulated (All) Verified Probes Only (4)
CSL Fidelity Index
7d composite CSL index · lower means closer to the official full-strength baseline (Lower is better, <15.0)
⛁ “Simulated” marks simulated baselines · red ≥15.0 flags distribution deviation
OpenAI · gpt-5.6-sol
8.4
xAI · grok-4.7
9.1
Anthropic · claude-opus-5.5
9.6
Kimi · kimi-k3.1-pro
10.8
Kiro · claude-sonnet-5
16.2
Cursor · claude-sonnet-5
18.4
OpenCode · kimi-k3.1-pro
45.7
Range: 8.4 – 45.7 ↗ CSL-Meter
Peak Drift
14:00–18:00 peak vs off-peak ΔCSL · lower is steadier (Lower is better)
⛁ Covers the afternoon congestion window · ≥+5.0 flags significant peak shift
OpenAI · gpt-5.6-sol
+1.2
xAI · grok-4.7
+1.6
Anthropic · claude-opus-5.5
+2.1
小米 MiMo · mimo-v2.6-pro
+2.2
OpenCode · kimi-k3.1-pro
+5.4
智谱清言 · glm-5.3
+6.8
Cursor · claude-sonnet-5
+7.2
24h Peak Window (UTC+8) ↗ CSL-Meter
Throughput
Median output speed (tokens/s) and time-to-first-token · higher is faster (Higher is better)
⛁ Also records TTFT (ms) and the 429 rate-limit rate
MiniMax · minimax-m3
96T
xAI · grok-4.7
88T
火山引擎方舟 · kimi-k3
84T
OpenAI · gpt-5.6-sol
76T
Cursor · claude-sonnet-5
71T
GitHub Copilot · gpt-5.6-sol
70T
Command Code · deepseek-v4.1-pro
68T
Unit: Tokens / sec (TPS) ↗ CSL-Meter
PARETO FRONTIER CSL Fidelity Index vs. Median Throughput (TPS) · Pareto quadrant analysis
15 of 15 plans Open 24h peaks & 30d trends →
Most attractive quadrant(high fidelity CSL ≤ 13.0 · high throughput TPS ≥ 65) ····Pareto line (Pareto-optimal frontier) Normal zone (CSL < 15.0) High-deviation alert zone (CSL ≥ 15.0) ⛁ Higher = closer to the official baseline · further right = faster throughput
MOST ATTRACTIVE QUADRANT · low drift, high throughput (CSL ≤ 13.0, TPS ≥ 65) CSL = 15.0 significant-deviation line (high-drift zone below) 9.0 12.0 15.0 18.0 45 TPS 60 TPS 75 TPS 90 TPS MiniMax (11.4)xAI (9.1)火山引擎方舟 (13.2)OpenAI (8.4)Cursor (18.4)GitHub Copilot (14.9)Command Code (12.6)阿里云百炼 (12.1)Kilo Code (13.8)Kiro (16.2)Kimi (10.8)Anthropic (9.6)小米 MiMo (11.9)智谱清言 (15.8)OpenCode (45.7) ↗ Artificial Analysis Style Telemetry · CSL-Meter (konspace.cc)
LIVE PROBE Run a Real Upstream API Probe (real outbound requests · zero credential retention · no synthetic results)
Gentle 4-way concurrency · 16 real requests (~1,100 tokens, hard cap 30,000 tokens) · fewer than 8 valid extractions = no report
04Human verification (Cloudflare Turnstile)
Verification boundary: the server uses this credential to make 16 real requests to the official whitelisted API (about 4–10 seconds); the credential lives only in memory and is burned after use; an invalid key or an upstream 401/403 fails loudly and produces no report. Waiting for a real test credential...
MATRIX All Plans & Providers — Objective Statistics (flat, factual, by the numbers)
Incl. Simulated Verified Probes Only (4)
All China Personal Plans Global Coding Subscriptions Global OA + Grok Official Pay-As-You-Go
Provider / Plan Data Status Claimed Model [Harness] Samples N ↓ CSL Index Net-JSD Split-JSD Nearest Baseline (Top-1) Peak Δ Billing/Cache Throughput Cite
Command Code
Command Code GOAT ($10/mo · 10x 个人档)
Simulateddeepseek-v4.1-pro [opencode]356
12.6±2.9
0.1480.129 (4%)deepseek-v4.1-pro (94.0%)
+3.8
1.00x (97%)
68T
540ms · 0.8%
Peaks · Badge
Cursor
Cursor Pro ($20/mo 个人档 · Auto/Sonnet)
Simulatedclaude-sonnet-5 [cursor]289
18.4±5.1
0.1780.162 (8%)claude-sonnet-5 (86.5%)
+7.2
1.08x (92%)
71T
490ms · 1.1%
Peaks · Badge
Kiro (AWS Spec IDE)
Kiro Pro ($20/mo · 1,000 Credits 个人自服务档)
Simulatedclaude-sonnet-5 [kiro]195
16.2±4.0
0.1680.151 (6%)claude-sonnet-5 (88.9%)
+4.9
1.04x (90%)
60T
560ms · 1.5%
Peaks · Badge
Kilo Code
Kilo Pass Starter ($19/mo · 个人 Gateway)
Simulatedclaude-sonnet-5 [raw]176
13.8±3.1
0.1530.135 (4%)claude-sonnet-5 (91.8%)
+3.3
1.01x (91%)
61T
530ms · 1.2%
Peaks · Badge
GitHub Copilot
GitHub Copilot Pro ($10/mo · 个人档)
Simulatedgpt-5.6-sol [cursor]152
14.9±3.5
0.1610.142 (5%)gpt-5.6-sol (90.4%)
+4.1
1.03x (90%)
70T
450ms · 1.0%
Peaks · Badge
OpenCode
OpenCode Go ($10/mo · 6x 个人档)
Verified(4)kimi-k3.1-pro [opencode]4
45.7±3.8
0.2660.256 (5%)claude-opus-5.5 (89.8%)
+5.4
1.02x (89%)
18T
4890ms · 0.0%
Peaks · Badge
Method note: readings present only the statistical distance (JSD) between sampled distributions and claimed baselines; no subjective verdicts. Simulated demo baselines are never mixed into real probe aggregates. Live SVG badge:CSL-Meter Badge
STREAM Stored Snapshots (Server-Verified strictly separated from Simulated · 模拟生成)
View sample snapshot #a49d2e →
Snapshot ID Sampled (UTC) Plan Claimed Model Verification CSL Index Net-JSD Split-JSD Nearest Base Distribution Throughput Action
#65ea9ec12026-11-06T05:33:00ZOpenCode · OpenCode Go ($10/mo · 6x 个人档)kimi-k3.1-proServer-Verified52.40.2320.372claude-opus-5.5 (89.8%)17.7 TPSView snapshot →
#f9c046902026-11-06T05:32:40ZOpenCode · OpenCode Go ($10/mo · 6x 个人档)kimi-k3.1-proServer-Verified42.10.2800.199claude-opus-5.5 (78.8%)17.4 TPSView snapshot →
#92b8993d2026-11-06T05:32:19ZOpenCode · OpenCode Go ($10/mo · 6x 个人档)kimi-k3.1-proServer-Verified44.30.3200.166claude-opus-5.5 (85.1%)15.4 TPSView snapshot →
#e8caa71d2026-11-06T05:31:54ZOpenCode · OpenCode Go ($10/mo · 6x 个人档)kimi-k3.1-proServer-Verified44.20.2320.286claude-opus-5.5 (85.2%)19.6 TPSView snapshot →
#a49d2e2026-11-05T16:55:47ZCommand Code · Command Code GOAT ($10/mo · 10x 个人档)deepseek-v4.1-proSimulated · 模拟生成12.60.1480.129deepseek-v4.1-pro (100.0%)68.5 TPSView snapshot →
#8f3a9c2026-11-05T16:55:47Z火山引擎方舟 · 方舟 Coding Plan Lite (¥40/月)kimi-k3Simulated · 模拟生成13.20.1480.134kimi-k3 (99.9%)84.5 TPSView snapshot →