Ship against every model. Operate one runtime.

Point the SDK you already use at tiers. We classify each call, route it to the cheapest model that clears your quality bar, inspect the session, verify the output, and return a signed receipt. Your provider keys and contracts stay yours.

Talk to the foundersView the integration
call_9f27e4 · production governed
classifycode-review · grade 2
routeclaude-sonnet-4.6
policyclean · failover armed
verify0.96 · passed
settle$0.0041 · signed
<1ms
added latency
14
providers
1
endpoint
56
models in the catalog
32
routing cells
87%1
reference workload reduction
BYOK
keys and contracts stay yours

The model market changes every week. Your application should not have to.

tiers absorbs provider churn behind a single endpoint that speaks both OpenAI and Anthropic wire formats, the two protocols nearly every SDK and framework already uses. Behind it, routing spans 56 models across 14 providers. New models enter through shadow evaluation. Better candidates are promoted only after they clear the task-specific quality bar.

01

Adopt without rewrites

Add a model, replace one, or move a workload without rebuilding the application around another provider SDK.

02

Degrade without failing

Provider brownouts, quota pressure, and access changes trigger cross-vendor failover while the mission is still running.

03

Stop paying models to do software's job

Deterministic work resolves locally. Arithmetic, date logic, transforms, and known templates never reach a model.

One call path. Four operating functions.

Routing is only the first decision. Production systems also need runtime policy, independent verification, and evidence that survives the request.

Gate · route

Classifies each request into one of 32 task and complexity cells, then chooses the lowest-cost model that clears the quality bar.

Aiglos · inspect

Evaluates tool calls, HTTP requests, and subprocesses before execution, across the full session rather than one prompt at a time.

Forge · verify

Checks the output against task intent, schema, and rolling quality baselines before the application consumes it.

Ulmo · settle

Records route, cost, policy, and verification in a tamper-evident receipt for every call.

The engineering objection: another proxy in the critical path.
tiers is designed as reversible infrastructure. One base URL in. One base URL out. It adds less than one millisecond to the reference path, keeps your provider relationships intact, and can fail open or closed by policy.
client.base_url = "https://YOUR_TIERS_ENDPOINT"
# same SDK · same messages · governed response
route claude-haiku-4.5 · verify passed
receipt rcpt_9f27e4 · cost $0.0003

Your teams can adopt the best model the day it ships, while the application keeps one contract with the runtime. Stop maintaining provider logic. Start owning inference.

Watch a governed call.

The full lifecycle, from the seat where it happens: routed, scanned, verified, settled, receipt signed. Try the injection. Try turning governance off.

WITHOUT TIERS
$0.0000
WITH TIERS
$0.0000
SAVED
$0.0000
you@prod ~ tiers session
GOVERNED

Simulated session using verified reference-workload figures.1 When the proxy is live, the same component runs against your tiers endpoint and nothing changes except the transport.

The same call gives the rest of the company what it needs.

Finance gets cost attribution. Security gets policy evidence. AI leadership gets quality telemetry. Nobody asks engineering to instrument the same workload four different ways.

One endpoint between your application and the model market.

Start with a shadow deployment. See the routes, quality, and counterfactual before moving production traffic.

Start a governed pilotRead the docs

1. Cost reduction measured on the tiers reference workload against a signed, immutable 14-day shadow baseline: $6.10 per day governed versus $47.00 per day ungoverned. Platform average across measured cohorts: 83.08%; optimized cohorts: 87.9%. Results vary by workload.

Make every AI dollar explain itself.

tiers turns inference into a governed cost center. Budget before execution, attribute every call, verify the work delivered, and compare governed spend against a signed parallel baseline.

Review the economics
inference · month to date reconciled
ungoverned counterfactual$47.00 / day
governed spend$6.10 / day
verified savings$40.90 / day
tiers fee2% of governed
receipt coverage100%
87%1
reference reduction
~335x
savings to fee
$0
provider markup
Per call
cost and owner attribution
Per loop
budgets and reserves
Signed
governed counterfactual
2%
of governed spend

AI spend becomes a governed operating line, not an invoice archaeology project.

Every receipt connects the call to a team, deployment, task, model, provider, cost, policy result, and quality outcome. Finance can see where spend went and whether the work cleared the bar.

01

Control spend before it happens

Per-session, per-team, and per-deployment budgets trip before a runaway loop becomes an invoice.

02

Defend the savings number

The shadow baseline runs the same workload governed and ungoverned, then signs both cost streams call by call.

03

Keep the economics neutral

No model markup, no seat fee, no incentive to route toward a more expensive provider.

The counterfactual is the product.

A generic benchmark cannot tell you what your workload would have cost. tiers measures the identical workload in parallel, preserves the evidence, and makes the difference auditable.

baseline_window 14 days
ungoverned $47.00 / day
governed $6.10 / day
quality_bar maintained
signature sha256:30:82:01:0a...
same calls · same period · two signed cost streams
The finance objection: every vendor claims savings.
The 87% is not presented as a universal promise. It is the measured result on the reference workload. The durable claim is the instrument: tiers gives each customer a signed counterfactual for its own traffic.

The goal is not a smaller AI program. It is more intelligence per dollar, with controls that survive scale. Every dollar of waste removed can fund the next deployment.

One receipt, four stakeholders.

Engineering gets route and latency. Security gets policy evidence. AI leadership gets quality. Finance gets the complete unit economics without another reconciliation step.

Put a signed meter in front of inference spend.

Run tiers in shadow mode first. Measure your own counterfactual before changing production behavior.

Review the economics

1. Reference workload result from a signed, immutable 14-day shadow baseline: $6.10 per day governed versus $47.00 per day ungoverned. Platform average across measured cohorts: 83.08%; optimized cohorts: 87.9%. The savings-to-fee ratio divides measured savings by a fee equal to 2% of governed spend. Results vary by workload.

Control the session, not just the prompt.

tiers inspects tool calls, HTTP requests, and subprocesses before execution. It evaluates behavior across the session, enforces provider and residency policy per call, verifies the output independently, and signs the evidence.

Request the security architecture
session_7a21 · policy trace inspecting
tool · read_file(config.yaml)allowed
http · api.vendor.com/v2allowed
subprocess · git diffallowed
trajectory · credential stagingblocked
receiptsigned · retained
3
interception surfaces
111
threat families
<1ms
inspection latency
Pre-exec
policy before action
Session
trajectory-aware defense
Per call
residency enforcement
Signed
evidence by default

Gateways see requests. tiers sees what the agent is trying to do.

A multi-step attack can look harmless one call at a time. tiers keeps state across the mission, evaluates the trajectory, and intercepts the action surfaces where intent becomes execution.

01

Stop actions before execution

Tool calls, outbound requests, and subprocesses are evaluated in-process, not reconstructed from logs after the fact.

02

Verify outside the provider

Quality and task alignment are checked by tiers, independent of the model that produced the output.

03

Produce the evidence automatically

Route, policy, verification, residency, and cost arrive as one tamper-evident receipt.

Policy follows the call.

Control is applied at runtime, where model, provider, geography, data class, tool access, and mission state are all visible together.

Session-level detection

Intent and anomaly signals accumulate across the complete mission rather than resetting on every request.

Sovereign routing

Residency and provider policy are enforced call by call, not assumed from a contract signed months earlier.

Independent drift detection

Rolling quality baselines surface provider-side regressions before they become a customer-support pattern.

Fleet control

Pause a session, a deployment, a team, a provider, or the entire fleet from the same operating plane.

The security objection: we already have a gateway and DLP.
Those controls remain useful. tiers covers the blind spot between them: the runtime sequence of model decisions and agent actions. It does not replace the perimeter. It governs the inference path.

Security stops being the team that blocks agents and becomes the reason the company can deploy them. The control and the evidence are created by the same call path.

The receipt is also the finance record and the quality record.

The company does not need separate instrumentation for security, cost, and model governance. Each function reads the same signed event.

Make runtime policy part of every AI call.

Review the architecture, threat model, deployment modes, and evidence format with the team that built it.

Request the security architectureDiscuss a governed pilot

Threat-family and inspection-latency figures reflect tiers' internal production architecture and mapping. Regulatory and standards alignment should be evaluated against each deployment's facts, jurisdiction, and control environment.

Own the model strategy. Do not inherit it from a provider.

Define task classes and quality bars once. tiers continuously evaluates the market, promotes cheaper challengers when they clear the bar, and rolls back when quality drifts.

See the governance plane
routing matrix · production learning
extraction · grade 1gpt-5.4-mini · 0.97
code review · grade 2claude-sonnet-4.6 · 0.94
research · grade 4gpt-5.6-sol · 0.91
math · grade 1local · $0.0000
shadow challengergemini-3-pro · evaluating
32
routing cells
56
models
14
providers
Task first
not provider first
Continuous
quality evaluation
Neutral
no model or token sales
87%1
reference workload reduction

Standardizing on one provider removes choice. It does not remove complexity.

Model quality, access, price, latency, and regional availability continue to move underneath the contract. tiers turns that volatility into a managed surface instead of an application dependency.

01

Make quality an owned capability

Your task taxonomy, quality bars, and production outcomes become portable intelligence rather than provider-specific configuration.

02

Benchmark continuously

Every cell keeps a production occupant and a shadow challenger. Promotions follow measured outcomes, not launch-day benchmarks.

03

Stay neutral by construction

tiers operates no models, sells no tokens, and earns nothing from where a call lands except that it landed correctly.

The four functions of an intelligence exchange.

A plural model market needs more than access. It needs price discovery, quality grading, market surveillance, and settlement.

Price discovery

Find the lowest price that can complete this task at the required quality, not the cheapest model in the abstract.

Quality grading

Verify what was delivered against the requested outcome before the application consumes it.

Market surveillance

Inspect the session for manipulation, unsafe actions, and provider or supply-chain anomalies.

Settlement

Clear the call with a signed record of route, policy, quality, and cost.

The AI leadership objection: we chose one frontier provider to avoid this complexity.
That decision simplifies procurement. It does not tell you which model should handle each task, whether quality changed, or whether the premium is justified. tiers absorbs the plurality, 56 models across 14 providers behind one endpoint, so the company can use it safely.

Stop betting on the winning model. Build the system that wins regardless. Your strategy becomes the governance position, not the logo on the invoice.

The same operating plane works for the teams that have to approve it.

Engineering gets a stable interface. Security gets runtime control. Finance gets the counterfactual. The AI strategy no longer depends on one function carrying the whole argument.

Turn the model market into an owned capability.

Map your highest-volume workloads into the routing matrix and measure how the current default performs against governed alternatives.

Run a routing workshop

1. Cost reduction measured on the tiers reference workload against a signed, immutable 14-day shadow baseline. Platform average across measured cohorts: 83.08%; optimized cohorts: 87.9%. Results vary by workload.

Deploy more AI without losing control.

tiers gives the company one operating plane for cost, quality, security, and evidence across every model and every team. The result is not a smaller AI ambition. It is an AI program that can survive its own success.

Talk to the founders
company AI · governed live
calls governed100%
outputs verified99.6%
policy interventions18
reference cost reduction87%
evidence coveragesigned · per call
1
governance plane
14
providers
All
teams and workloads
Cost
measured and attributable
Quality
verified before use
Risk
controlled at runtime
Evidence
signed on every call

AI adoption is outrunning the operating system around it.

Teams can add a model in an afternoon. The company still needs a coherent answer for spend, quality, security, ownership, and auditability. tiers makes those controls part of the call itself.

01

Scale without budget surprise

Budgets, routing, deterministic execution, and per-call attribution keep usage growth from turning into unexplainable spend.

02

Move faster without lowering the bar

Teams can use more providers and more capable agents because quality and policy are continuously enforced.

03

Give the board one answer

Every AI call is routed, inspected, verified, and settled under one operating model.

Governance is an expansion strategy.

Lower unit cost makes more workflows viable. Runtime trust makes more teams willing to deploy them. Evidence makes more regulated use cases possible. Governed inference expands the surface area of AI the company can responsibly use.

For the board

A measurable answer for AI economics, exposure, and control coverage.

For operators

One runtime rather than a different stack for each team, model, and vendor.

For customers

Outputs that were verified and policies that were enforced before the system acted.

For regulators and auditors

A durable evidence object produced by normal operations rather than reconstructed later.

The executive objection: is this just cost cutting?
No. Cost is the fastest way to prove the runtime works, but the strategic outcome is larger: more viable workflows, more provider choice, and more AI deployed under control.

The answer to “what is our AI strategy?” becomes simple: every call is governed, so the company can deploy more intelligence than competitors can safely or economically operate.

Bring the staff into the same operating model.

Each function gets the information it needs from the same governed call, without creating five separate AI programs inside the company.

Make AI intensity survivable.

Start with the workloads already consuming the most tokens. tiers will measure the current state in shadow mode before changing production traffic.

Schedule an executive briefing

Illustrative executive-console metrics are presentation examples except for the reference workload cost result. The 87% reduction was measured against a signed, immutable 14-day shadow baseline; platform average across measured cohorts was 83.08%, with optimized cohorts at 87.9%. Results vary by workload.

Common questions.

What does tiers give engineering teams?
tiers gives engineering teams a single governed endpoint that speaks both OpenAI and Anthropic wire formats, covering 56 models across 14 providers. Each call is routed to the cheapest model that clears the task-specific quality bar, inspected for policy compliance, verified before the output reaches the application, and settled with a signed receipt. Teams add tiers by changing one line of code — the base_url of the existing SDK client — and retain their own provider keys and contracts.
How does tiers integrate with the SDKs our engineering team already uses?
tiers exposes an OpenAI-compatible endpoint and an Anthropic-compatible endpoint. Integration requires changing one value: set base_url in the existing SDK client to the tiers endpoint and replace the provider API key with a tiers key. Prompts, application logic, and provider relationships stay the same. The same approach works for direct HTTP calls.
Does adding tiers to the call path increase latency?
tiers adds less than one millisecond on the reference path. Route selection, policy inspection, and verification run in-process within the governed call. The endpoint can be configured to fail open or closed by policy, so the overhead is bounded by design.
How does tiers help finance teams track AI spend?
tiers produces a signed receipt for every governed call, connecting each call to a team, deployment, task, model, provider, list-price cost, and governed cost. Per-session, per-team, and per-deployment budgets can be set to trip before a runaway loop becomes an invoice. The result is cost attribution at the call level without additional instrumentation.
How does tiers prove the savings it claims?
tiers runs a shadow baseline that measures the identical workload governed and ungoverned in parallel, then signs both cost streams call by call. The counterfactual is specific to each customer's own traffic, not a generic benchmark. On the tiers reference workload, the 14-day signed baseline showed $6.10 per day governed versus $47.00 per day ungoverned — an 87% reduction at the maintained quality bar. Results vary by workload.
What does tiers give security teams for AI governance?
tiers inspects tool calls, HTTP requests, and subprocesses before execution, evaluating behavior across the full session rather than one prompt at a time. Intent and anomaly signals accumulate across the mission, so multi-step attacks that look harmless individually can be detected from their trajectory. Every blocked and allowed action is recorded in a signed receipt.
What audit evidence does tiers produce for compliance?
tiers produces a tamper-evident signed receipt for every governed call, containing the route, model, cost, policy decision, verification result, and a cryptographic signature. Receipts are hash-chained so any alteration to a historical record is detectable. They are exportable and form the audit trail for compliance reviews, FinOps attribution, and regulated deployment sign-off.
How does tiers help AI leaders manage a multi-model strategy?
tiers organises inference into 32 task-and-complexity cells. Each cell has a production occupant — the cheapest model that cleared the quality bar — and a shadow challenger being evaluated against real traffic. AI leaders define the task taxonomy and quality bars once; tiers continuously evaluates the market, promotes cheaper challengers when they clear the bar, and rolls back when quality drifts. The strategy becomes portable intelligence rather than provider-specific configuration.
How does tiers stay neutral across model providers?
tiers operates no models and sells no tokens. Its fee is 2% of governed spend, which means it earns nothing from routing a call to a more expensive model. The catalog covers 56 models across 14 providers and adds new models through shadow evaluation rather than commercial relationship.
What does tiers give executives overseeing an AI program?
tiers gives executives one operating plane for cost, quality, security, and evidence across every model, provider, and team. Each AI call is routed, inspected, verified, and settled under one governance model, producing a measurable answer for AI economics, exposure, and control coverage. The result is an AI program that can scale without creating proportional budget surprise, security exposure, or audit gaps.
How does tiers support board-level AI governance reporting?
Every governed call produces a signed receipt that is exportable and hash-chained. Spend is attributable per team, deployment, and task. Policy intervention counts, verification pass rates, and counterfactual savings are measured continuously. Executives can answer questions about AI cost, quality, and risk exposure from the same evidence the operational teams use — no separate reconstruction step.

1. Cost figures verified against a signed, immutable 14-day shadow baseline: $6.10 per day governed versus $47.00 per day ungoverned on the reference workload; platform average 83.08% across cohorts, optimized cohorts 87.9%. The terminal session is simulated with reference-workload figures. Results vary by workload.