Claude Sonnet 4.5 output token usage increased 52% vs 7-day average

criticalactiveAnthropic Claude

Current impact

$3,100

vs baseline

Forecasted impact / mo

$3,600

if trend continues

Change vs baseline

+52.0%

Output Tokens

Confidence

76%

likely cause attribution

Event Context

VendorAnthropic Claude
StartedApr 27, 2026
Detected15d ago
OwnerLegal Ops / Product Engineering
ModelClaude Sonnet 4.5
FeatureAI Contract Review
Endpoint/api/contracts/analyze

Usage Timeline — Output Tokens

Actual vs 7-day rolling baseline

Usage Metric Breakdown

Current vs baseline — all contributing metrics

Input Tokens

1.4M

Baseline: 800k

+75%

Output Tokens

710k

Baseline: 300k

+137%

Cache Read Tokens

496k

Baseline: 420k

+18%

Cache Write Tokens

220k

Baseline: 180k

+22%

API Requests

4k

Baseline: 2k

+81%

Root Cause Analysis

1 confirmed · 2 need confirmation

MeasuredFirst observedInferredAnnotated
Measured · Anthropic usage API
79%

Calls to /api/contracts/analyze: 1,400/day → 2,140/day (+53%)

Measured · Anthropic usage API
68%

Avg completion_tokens per call: 1,800 → 2,900 (+61%)

Interpreted as: Longer output suggests contracts are more complex or the response format changed. Check if the analysis prompt template was modified.

Was the contract analysis prompt or response format changed around Apr 20?

First observed · First observed in usage data
58%

Call volume to /api/contracts/analyze jumped 53% on Apr 20 with no prior ramp.

Did an internal rollout or feature flag change happen on Apr 20?

Likely cause

Increased usage of AI Contract Review workflow after internal rollout

76% confidence · 2 signals awaiting confirmation

Connect a deployment webhook to auto-correlate releases with future spend events.

Suggested Next Steps

Owner: Legal Ops / Product Engineering

1

Confirm internal rollout scope with Legal Ops team

2

Enable prompt caching on /api/contracts/analyze to cut input token costs

3

Review output length — shorter responses may be viable for initial triage

4

Set cache write/read ratio target of 80% to optimize costs

Related dimensions

endpointmodelcache_policyteam