1Classify
Score every request before a single token is spent.
A hybrid classifier reads intent, context, tools, and risk. Rules first, embeddings second, an LLM only when the case is genuinely hard.
docscomplexity 2/5risk lowdev-tools
Every prompt paysthe right price.
Steeros routes every prompt to the cheapest model that can handle it. Up to 85% lower spend, same quality. One proxy, no code changes.
87.7%
cost cut on our held-out eval
139 prompts, router vs all-flagship
95%
flagship quality kept
RouteLLM study, LMSYS 2024
92.4%
lower than Opus-only routing
our traffic, 1,011 requests
$0.004
average light-tier request
vs $0.068 unsteered
Pick a flagship and the door flies open for everything. Teams burn 60-70% of their spend on prompts a small model handles perfectly.
Like taking a Ferrari to buy milk. Every single time.
Four moves between you and the right model.
2Route
The cheapest model that clears the bar wins.
Four modes, one policy: cheap, fast, balanced, quality. You set per-tier cost ceilings; the router spends inside them, every time.
light$0.003
balancedchosen$0.012
flagship$0.20
3Convert
Anthropic format in, native format out.
Token caps get clamped, thinking blocks stripped, mid-chat system roles demoted. Each transform is logged so nothing surprises you.
max_tokens clamped to 8192
thinking block stripped
system message demoted to user
4Account
Every cent accounted, to six decimals.
Provider spend, cache reads and writes, classifier calls, compaction. Measured against a fair baseline, session and all-time.
req #48211 deepseek-chat $0.000041
cache read 14.2k tok ×0.1 $0.00002
baseline: same tokens on a fixed model
Three rungs. One ladder.
The smallest model that clears the bar wins. You set the ceilings; the router spends inside them. One proxy in front of DeepSeek, Anthropic, OpenAI, and Gemini.
The errand tier.
Light tier
cost ceiling per request
$0.003
- git messages
- README edits
- typo fixes
- test stubs
51% of a typical mix lands here
The daily driver.
Balanced tier
cost ceiling per request
$0.012
- features
- refactors
- code review
- most debugging
32% of a typical mix lands here
When it actually earns it.
Flagship tier
cost ceiling per request
$0.20
- architecture
- hard bugs
- novel design
- long-horizon planning
17% of a typical mix lands here
The boring details, handled.
Cache-aware pricing
Every request is priced the way the provider actually bills it: cache reads, cache writes, full inputs. Our own traffic banked $22.79 in cache savings.
anthropic read
×0.1
anthropic write
×2.0
deepseek cache hit
~3% of input
deepseek full input
×1.0
Fair baselines
Savings are measured against the same token bundle on a fixed model, at that model’s own cache rates. No flattering math.
1,011 requests, actual: $22.41deepseek-only: $24.31 · opus-only: $294.68
Benchmarks
Held-out eval across chat, math, mcq, and code, scored blind against a judge and ground truth. Small samples on the non-chat categories.
chat 100 · math 80 · mcq 100 · code 5087.7% cheaper than all-flagship
Carbon
Spend and carbon fall together. Fewer flagship tokens means fewer GPU-hours.
66%
fewer flagship tokens, same output
Compaction
Long sessions get compacted; past answers get reused, so repeat prompts skip straight to the right tier. 399 calls on our traffic, $0.12 total.
summary: 3.4k tokens → 412route: remembered → light
Guardrails
Monthly caps that degrade to the light tier. Automatic escalation when a cheap model fails. PII scrubbing on the roadmap.
over_cap_policy: degrade_to_light
Live dashboard
Watch every route, token, and cent in real time, on your own machine at localhost:3000.
tokens, cost, routes
Do the math on your burn.
Slide your team size and see spend before Steeros, spend after, and what stays in your budget.
10
1 to 100 seats
60
10 to 300 prompts
without Steeros
$857/mo
12,600 requests
all at flagship prices
all at flagship prices
with Steeros
$299/mo
−65%
Kept in your budget$558every month
Same output, roughly 65% fewer flagship tokens, so energy falls with the bill.
Modeled on our pilot mix: 51% light at $0.004, 32% balanced at $0.020, 17% flagship at $0.090 per request. Your distribution will differ.
Free to pilot. Built for fleets.
Steeros Free25kreq/mo
- Working config template, routes.free.yaml
- One-line localhost proxy
- Cost ceilings you edit by hand
- Enough for a solo dev or a small team pilot
EnterpriseCustom pricing
- Every agent: Claude Code, Cursor, Copilot, Gemini CLI
- PII scrubbing and redaction
- Self-tuning classifier on your prompt mix
- Unlimited seats, real support