What does Hydrate save your team?
Based on 12 benchmark runs using real Anthropic API spend. APJ Ltd run, April 2026. All figures are indicative - actual savings depend on your usage pattern and model mix.
Cost-per-session figures from real APJ Ltd benchmarks: baseline $0.158/session, Sonnet+Hydrate $0.206/session, Haiku+Hydrate $0.160/session, Firefighter (Haiku+Hydrate) $0.075/session. Cache hit rates measured: baseline 67%, Sonnet+Hydrate 96.1%, Haiku+Hydrate 96.8-97%. All figures are indicative - actual savings depend on your usage pattern, model mix, and session characteristics.
Lead architect uses Sonnet once to pin decisions. Developers use Haiku with Hydrate injecting context. Saves ~55% vs baseline at enterprise scale.
Haiku+Hydrate beats Sonnet on both cost AND quality across 3 sprints. $1.93 vs $2.54 - 24% cheaper with a higher quality score (8/10 vs 7/10).
Equal quality at 3.15× lower cost. Hydrate injects existing project facts - Haiku no longer needs to infer what it can be told directly.
Both models fix the bug correctly - the value is entirely cost reduction. At 1,000 sessions/month per dev, a 100-dev team saves $6,300/month on model costs alone vs running Sonnet.
| Configuration | Annual cost | Annual saving vs baseline | Cache benefit | Quality |
|---|---|---|---|---|
| Baseline - Sonnet, no Hydrate | ~$220,000 | - | 67% cold start | 8/10 |
| Quality - Sonnet + Hydrate | ~$160,000 | ~$60K saved (27%) | 96.1% throughout | ≥8/10 |
| Savings * - Haiku + Hydrate | ~$100,000 | ~$120K saved (55%) | 96.8% throughout | 7/10 |
| Firefighter - Haiku + Hydrate | ~$66,000 | ~$154K saved (70%) | 97% throughout | Equal to Sonnet |
| Company size | New hires / year | Onboarding sessions each | Baseline cost (Sonnet) | With Hydrate (Haiku) | Annual saving |
|---|---|---|---|---|---|
| Small - 20 devs | ~5 | 50 | $127.50 | $40.50 | $87 |
| Medium - 100 devs | ~25 | 50 | $637.50 | $202.50 | $435 |
| Large - 500 devs | ~125 | 50 | $3,187.50 | $1,012.50 | $2,175 |
| Enterprise - 2,000 devs | ~500 | 50 | $12,750 | $4,050 | $8,700 |
A new developer on Sonnet needs Sonnet's reasoning power because they don't know the codebase. They're asking Claude to infer what was decided, understand undocumented conventions, and avoid contradicting choices made weeks ago. Hydrate gives Haiku that knowledge explicitly - injected as structured facts before the session starts. The model no longer needs to infer what it can be told directly. That's why quality matches at 3.15× lower cost. Context poverty is the tax you pay when memory is absent.
Model direct API spend on your own Anthropic key. Three mixes
covering the common BYOK shapes. All rates pull from
src/data/anthropic-rates.json. Change the rates
there to refresh every number on this page.
Assumes 30% output token ratio; Hydrate cache resolves 96.8% of input tokens at the cache-read rate, 3.2% at the fresh input rate (compact-survival/v1.0 measured ratio).