DCP Token Savings Report
Dynamic Context Pruning — compound savings analysis across 100 sessions
66.9M
tokens saved via compound DCP effect
Without DCP you would have used 261.9M tokens instead of 195.0M.
34.3% saved · $200.68 at $3/M
33.1x compound multiplier iThe multiplier shows how much compound savings exceeds one-time compression. A 31.9x multiplier means each compressed token saves ~32 tokens total across all subsequent requests in that session.
73/100 sessions had DCP
Understanding This Report
One-Time Savings measures tokens removed during each compression event — counted once.
If DCP compresses 40K tokens into a 1.5K summary, that's 38.5K tokens saved in that single event.
Compound Savings recognizes those 38.5K tokens are never re-sent on subsequent
requests. In a session that continues for 149 more turns, the real savings is 38.5K × 149 = 5.7M tokens.
Example: Session A has 164 requests. DCP compresses 173K tokens into summaries
across 6 blocks, starting around turn 15. Those 173K tokens are never re-sent on the remaining ~149 requests.
One-time savings: 173K. Compound savings: 14.9M (86x multiplier).
The Multiplier (33.1x) means that on average, each compressed token saves ~33
tokens across all future requests in that session.
One-Time vs Compound Savings
One-Time = tokens removed during compression (counted once per event)
·
Compound = same tokens × every subsequent request that benefits
One-Time Compression iSum of (compressedTokens - summaryTokens) across ALL compression events in ALL sessions. This is the traditional metric that only counts tokens removed once.
2.0M
Tokens removed during compression events — counted once.
Compound Savings (Real) iOne-time savings multiplied by the number of subsequent requests that benefit from each compression. This reflects the actual reduction in API token consumption.
66.9M
Tokens saved across ALL subsequent requests in each session.
Key Metrics
Tokens UsediTotal tokens (input + output + reasoning) consumed across all 100 sessions. This is what was actually sent/received after DCP compression.
195.0M
Tokens SavediCompound savings: one-time compressed tokens multiplied by subsequent requests that benefit. This is the real reduction in API consumption.
66.9M
Savings RateiCompound tokens saved divided by tokens used. Without DCP, you would have used 261.9M tokens total.
34.3%
USD Saved ($3/M)iEstimated cost savings at a blended rate of $3.00 per million input tokens. See sensitivity table below for other rates.
$200.68
Total RequestsiNumber of model inference calls (step-finish events) across all sessions. Each request benefits from prior DCP compressions.
4111
DCP BlocksiTotal compression events across all sessions. Each block compresses a range of conversation messages into a summary.
60
MultiplieriCompound savings divided by one-time savings. A 33.1x multiplier means each compressed token saves ~33 tokens on average across all subsequent requests.
33.1x
Sessions w/ DCPi73 of 100 sessions have DCP state files. Of the remaining 27: 14 were too short for DCP, 13 had enough activity (≥10 requests + ≥10.0K peak context) that DCP should have triggered but didn’t. DCP activates via a hard limit at 100.0K tokens or a soft nudge every 5 turns after turn 15.
73/100
Savings Rate
• 66.9M tokens saved by DCP (34.3%)
• 195.0M tokens actually used (65.7%)
Without DCP, total would be 261.9M
Top Sessions by Compound Savings
Multiplier = compound ÷ one-time for that session
Sessions ranked by compound token savings
#1
14.9M
164 req / 6 blk / 85.8x
#2
10.6M
105 req / 4 blk / 75.0x
#3
9.4M
101 req / 1 blk / 74.0x
#4
4.2M
145 req / 3 blk / 43.3x
#5
4.1M
212 req / 4 blk / 38.9x
#6
2.9M
70 req / 3 blk / 27.9x
#7
2.8M
57 req / 3 blk / 31.3x
#8
2.5M
49 req / 3 blk / 27.2x
#9
1.7M
72 req / 1 blk / 46.0x
#10
1.7M
56 req / 1 blk / 38.0x
#11
1.4M
45 req / 2 blk / 23.5x
#12
1.3M
44 req / 2 blk / 16.8x
#13
1.2M
45 req / 1 blk / 16.0x
#14
1.1M
36 req / 1 blk / 20.0x
#15
1.1M
29 req / 1 blk / 11.0x
Per-Model Breakdown
⚠ missed = ≥10 requests + ≥10K peak, no DCP
no DCP needed = all sessions short
Models ranked by compound savings · bar meta shows DCP calls, session coverage, and peak context size
glm-5.2
25.6M
18 calls | 15/17 sess | max 70.8K ⚠ 1 missed
qwen3.7-plus
17.3M
18 calls | 18/22 sess | max 107.2K
glm-5.1
12.3M
17 calls | 24/35 sess | max 37.6K ⚠ 8 missed
minimax-m3
9.4M
1 calls | 3/5 sess | max 87.5K ⚠ 1 missed
kimi-k2.7-code
1.2M
4 calls | 2/2 sess | max 99.8K
gpt-5.4-mini
1.2M
2 calls | 2/4 sess | max 67.8K ⚠ 2 missed
gpt-5.5
0
0 calls | 6/9 sess | max 59.7K ⚠ 1 missed
deepseek-v4-flash
0
0 calls | 1/1 sess | max 35.3K
deepseek-v4-pro
0
0 calls | 0/3 sess | max 24.8K no DCP needed
north-mini-code-free
0
0 calls | 1/1 sess | max 172.5K
nex-n2-pro:free
0
0 calls | 1/1 sess | max 19.0K
USD Savings Sensitivity
| Rate | Estimated USD Saved (Compound) |
|---|
| $2/M | $133.79 |
| $3/M | $200.68 |
| $5/M | $334.46 |
Blended input-token pricing from fungies.io, awesomeagents.ai, pecollective.com, stochasticsandbox.com, tldl.io (2026).
Missed DCP Opportunities
13 sessions had enough activity for DCP (≥10 requests + ≥10.0K peak context) but DCP didn’t trigger. These sessions used 7.3M tokens collectively. DCP was installed (files date to March 2026) but its soft nudge system (every 5 turns after turn 15) didn’t prompt compression — likely because context stayed under the 100.0K hard limit.
| Session | Model | Requests | Peak Context | Tokens Used |
|---|
| ses_13ccfadedffe1qWO8Mwq... | glm-5.1 | 22 | 25.5K | 971.9K |
| ses_14dcb2ab7ffepT2FuK4f... | glm-5.1 | 36 | 15.3K | 919.8K |
| ses_14db006e1ffeclQALCSy... | glm-5.1 | 32 | 12.8K | 773.1K |
| ses_146ac49acffepgQO1CA9... | glm-5.1 | 21 | 15.9K | 723.2K |
| ses_13cd10d76ffe8vX2oexK... | glm-5.1 | 18 | 12.6K | 678.5K |
| ses_138b4a6e6ffems9VY22j... | glm-5.2 | 15 | 12.5K | 619.3K |
| ses_138a8cec6ffeadERV1g1... | gpt-5.4-mini | 21 | 12.1K | 600.1K |
| ses_14db10caeffeqkdOpW71... | minimax-m3 | 15 | 29.8K | 516.7K |
| ses_143fe55a8ffeCicrkfWM... | gpt-5.4-mini | 18 | 17.3K | 475.9K |
| ses_13cd0336affeEhNmlmXN... | glm-5.1 | 17 | 13.3K | 408.2K |
| ses_13cb3a911ffeexIyNSG5... | gpt-5.5 | 13 | 15.3K | 226.2K |
| ses_139093da9ffezJ7JidV2... | glm-5.1 | 11 | 13.3K | 194.7K |
| ses_13cd2771cffeZWtv15e5... | glm-5.1 | 10 | 13.4K | 171.4K |
Insights
- 13 session(s) missed DCP (≥10 requests and ≥10.0K peak context, but no compression triggered): 7.3M tokens consumed across these sessions. DCP was installed but its nudge system didn’t fire — likely because context stayed under the 100.0K hard limit and the soft nudge (every 5 turns after turn 15) didn’t prompt the model to compress.
- 5 model(s) never triggered DCP: gpt-5.5, deepseek-v4-flash, deepseek-v4-pro, north-mini-code-free, nex-n2-pro:free. These were used only on short sessions with small context windows.
- Best compound multiplier: Session with 164 requests achieved 85.8x — 173.3K one-time compression amplified to 14.9M compound savings.
- DCP activation: 73 of 100 sessions (73%) have DCP data. 41 sessions were too short to need it.
- Compound effect is the story: One-time compression removed 2.0M tokens. Real API savings: 66.9M (33.1x more).
- Average DCP blocks per active session: 0.8 blocks across 73 sessions.
DCP saved 66.9M tokens across 100 sessions — 34.3% of total usage, or $200.68 in API costs at $3/M.
Without DCP, total cost would be 261.9M tokens instead of 195.0M. Compound effect amplifies one-time compression by 33.1x.
- 73% session DCP activation (73/100 sessions); 14 too short, 13 missed (7.3M tokens at risk)
- 11 models tracked, 5 never triggered DCP
- Top session: 14.9M compound on 7.6M used (85.8x)