MO§ES™ × 5 Measured Benchmarks

MO§ES™ leads all 5 measured economic categories — 7-day window (2026-05-08 → 2026-05-14)

Cache hit rate

96.88%

#1 — > field's 96.2%

SRC: Token Dashboard 7d

Output : Input

31.7×

30d 42.5× · 90d 22.1×

#1 — 83× field leader (0.38)

SRC: 3.90M out ÷ 123K in

Tokens / task

767K

#1 — 3.6× more efficient

SRC: 1.12B ÷ 1,465 tasks

Time / task

1.84 min

#1 — 3.2× faster

SRC: ~45 hr ÷ 1,465 tasks

Cost / LOC

$0.0007

plan · API: $0.044 · ccusage: $0.018

#1 — < 1¢ per line

SRC: $23.33 ÷ 35,242 LOC

Testing methodology — not the same thing being measured: AA field: per-task isolated runs on SWE-Bench-Pro-Hard-AA, Terminal-Bench v2, SWE-Atlas-QnA. Each task = one bug/issue. MO§ES™: sustained operator work on App Hub (real product) over 7-day measurement window (2026-05-08 → 2026-05-14), 21 sessions, 7,327 turns. App Hub = 5 build days within the window (5/10–5/14), 35,242 LOC shipped. Cost/task ($0.017 sub · $1.05 API) computed similarly but not on cards. $/LOC: AA = cost_per_task ÷ 20 LOC industry convention. MO§ES = $23.33/wk sub ÷ 35,242 actual LOC.

1. Cache Hit Rate

Cache reuse % · higher better · MO§ES sustained multi-project

96.9%

96.2%

96.1%

94.9%

94.5%

93.7%

92.8%

91.7%

87.8%

86.1%

85.3%

83.7%

79.8%

MO§ES™CC+Opus 4.7+op

Cursor CLIOpus 4.7 (Medium)

Claude CodeOpus 4.7 (Medium)

Claude CodeKimi K2.6

CodexGPT-5.5 (Medium)

Claude CodeSonnet 4.6 (Medium)

Claude CodeOpus 4.6 (Medium)

CodexGPT-5.4 (Medium)

Cursor CLIComposer 2

Cursor CLIGPT-5.5 (Medium)

Gemini CLIGemini 3.1 Pro (High)

Cursor CLIGPT-5.4 (Medium)

Claude CodeGLM-5.1

Claude CodeDeepSeek V4 Pro (High)

SRC: cache_read 1.084B ÷ (1.084B + 34.83M cache_create + 123K input) = 96.88%

2. Output : Fresh Input (log scale)

Output tokens per fresh-input token · higher = denser signal

31.7×

0.38

0.25

0.24

0.17

0.16

0.15

0.14

0.07

0.06

0.05

0.04

MO§ES™CC+Opus 4.7+op

Cursor CLIOpus 4.7 (Medium)

Claude CodeKimi K2.6

Claude CodeSonnet 4.6 (Medium)

Claude CodeOpus 4.7 (Medium)

CodexGPT-5.5 (Medium)

Cursor CLIComposer 2

Claude CodeOpus 4.6 (Medium)

CodexGPT-5.4 (Medium)

Gemini CLIGemini 3.1 Pro (High)

Cursor CLIGPT-5.5 (Medium)

Cursor CLIGPT-5.4 (Medium)

Claude CodeGLM-5.1

Claude CodeDeepSeek V4 Pro (High)

SRC: 3,902,803 output ÷ 123,246 fresh input = 31.7× · 30d 42.5× · 90d/all-time 22.1×

3. Tokens per Task

Total tokens · lower better

767K

2.74M

2.93M

3.24M

3.33M

3.75M

4.27M

4.41M

4.92M

5.42M

6.20M

7.28M

8.88M

MO§ES™CC+Opus 4.7+op

Cursor CLIGPT-5.5 (Medium)

Cursor CLIOpus 4.7 (Medium)

Gemini CLIGemini 3.1 Pro (High)

Cursor CLIComposer 2

Claude CodeOpus 4.7 (Medium)

Cursor CLIGPT-5.4 (Medium)

Claude CodeOpus 4.6 (Medium)

Claude CodeSonnet 4.6 (Medium)

CodexGPT-5.4 (Medium)

CodexGPT-5.5 (Medium)

Claude CodeDeepSeek V4 Pro (High)

Claude CodeKimi K2.6

Claude CodeGLM-5.1

SRC: 1.123B 7d total ÷ 1,465 task-equivs (7,327 turns ÷ 5) = 767K

4. Time per Task

Wall time · lower better

1.8m

5.8m

6.2m

6.9m

7.0m

7.1m

7.6m

7.8m

8.7m

9.2m

18.0m

21.6m

41.5m

MO§ES™CC+Opus 4.7+op

Claude CodeOpus 4.7 (Medium)

Cursor CLIGPT-5.5 (Medium)

CodexGPT-5.4 (Medium)

Claude CodeOpus 4.6 (Medium)

CodexGPT-5.5 (Medium)

Cursor CLIGPT-5.4 (Medium)

Gemini CLIGemini 3.1 Pro (High)

Cursor CLIOpus 4.7 (Medium)

Cursor CLIComposer 2

Claude CodeSonnet 4.6 (Medium)

Claude CodeDeepSeek V4 Pro (High)

Claude CodeGLM-5.1

Claude CodeKimi K2.6

SRC: ~45 hr active ÷ 1,465 = 1.84 min

5. Cost per LOC — all field models + MO§ES™ (log scale)

USD per line shipped · lower better · MO§ES plan basis $0.0007 leads field

$0.0007

$0.0035

$0.018

$0.038

$0.044

$0.050

$0.051

$0.062

$0.063

$0.073

$0.076

$0.080

$0.10

$0.11

$0.20

$0.26

$3.30

MO§ES™Plan basis · subscription

Cursor CLIComposer 2

Claude CodeDeepSeek V4 Pro (High)

Claude CodeKimi K2.6

MO§ES™API equiv basis

CursorIndustry low est.

Claude CodeSonnet 4.6 (Medium)

Claude CodeOpus 4.7 (Medium)

Claude CodeOpus 4.6 (Medium)

Cursor CLIOpus 4.7 (Medium)

Cursor CLIGPT-5.4 (Medium)

Gemini CLIGemini 3.1 Pro (High)

Cursor CLIGPT-5.5 (Medium)

CodexGPT-5.4 (Medium)

CodexGPT-5.5 (Medium)

Claude CodeGLM-5.1

CursorIndustry high est.

DevinIndustry low est.

DevinIndustry high est.

AA models: cost_per_task ÷ 20 LOC · Cursor/Devin: industry estimates · MO§ES plan: $23.33/wk ÷ 35,242 = $0.000662 · MO§ES API: $1,564.47 ÷ 35,242 = $0.0444

RAW DATA — 7-DAY MEASUREMENT WINDOW (2026-05-08 → 2026-05-14)

Input

123,246

fresh tokens in

drives cache hit · out:in

Output

3.90 M

3,902,803

drives out:in

Cache create

34.83 M

34,826,779

drives cache hit

Cache read

1.084 B

1,084,399,183

drives cache hit

Total tokens

1.123 B

1,123,252,011

drives tokens/task · time/task

ancillary inputs › 21 sessions · 7,327 turns · 1,465 tasks · ~45 hr active · 35,242 LOC · $23.33/wk plan
$1,564.47 API equiv

MO§ES™ × Artificial Analysis Coding Agent Benchmarks

1. Cache Hit Rate

2. Output : Fresh Input (log scale)

3. Tokens per Task

4. Time per Task

5. Cost per LOC — all field models + MO§ES™ (log scale)