# Metrum GenAI Smart Router Product Page Handoff

Prepared for Saikat on September 7, 2026.

## Package contents

- `detailed-product-page-brief.md` — editable source brief with positioning, page structure, draft copy, detailed savings model, claim guidance, documentation links, SEO, analytics, and launch checklist.
- `detailed-product-page-brief.pdf` — shareable rendering of the detailed brief.
- `mock-preview.html` — responsive, branded long-form product-page concept. Open this file in a browser; it uses only the included local logo asset and has no runtime dependencies.
- `mock-preview.pdf` — print/share rendering of the branded page concept.
- `mock-preview-first-page.png` — quick visual preview of the opening PDF page.
- `savings-model.csv` — editable scenario outputs at 1, 10, and 100 developers across low/measured/high token volumes.
- `frontier-rate-card.csv` — verified September 2026 frontier list prices with tier, cache rates, and source URLs.
- `assets/metrum-logo-white.png` — Metrum logo downloaded from the current public Metrum site for this mockup.

## Recommended review order

1. Open `mock-preview.html` in a browser to review the narrative and visual direction.
2. Read `detailed-product-page-brief.pdf` for exact copy, claims, assumptions, and implementation notes.
3. Adjust `savings-model.csv` or the assumptions in the brief if the desired planning scenario changes.
4. Validate the GitHub repository is publicly reachable before publishing the product page.
5. Run the link and claim checks in the launch acceptance checklist.

## Core message

Metrum GenAI Smart Router is an Apache-2.0 open-source, provider-neutral GenAI gateway. It gives applications, coding agents, and subagents one stable OpenAI- and Anthropic-compatible API while deployment-owned policy selects among validated hosted and private models.

The on-premises story is intentionally precise: Smart Router can route to configured private inference services so enterprises can use smaller models on older or lower-cost hardware and reserve premium hardware or frontier APIs for demanding work. Smart Router does not schedule GPUs or place model replicas.

The enterprise-readiness story includes customizable static, weighted,
failover, dynamic-score, TypeScript, external-service, and contract-backed
routing; per-owner/caller-key/project/environment/client attribution; projects
mapped to applications, business units, or cost centers; request-time spending
and savings reporting; RPM/TPM/concurrency and token budgets; caller and
provider traffic shaping; adaptive upstream backoff; capacity pooling;
request-shape compatibility gates; scoped administration; PII filtering; and
request-level operational diagnostics.

## Savings headline

The central illustrative scenario is:

- **100 million tokens per developer per week** — measured from a 10-person team consuming about 1B tokens/week, not estimated from hours at a keyboard;
- 52 weeks per year, because automation and monitoring agents run continuously;
- a segmented request-shape mix (cache reads, cache writes, uncached input, and 25% output share including reasoning tokens);
- 75% of GPT-6 Astra requests above the 272K long-context threshold;
- a 50/50 frontier baseline of **GPT-6 Astra and Claude Fable 5.1** — true frontier peers;
- list prices verified September 7, 2026 (see `frontier-rate-card.csv`).

The workload deliberately spans more than interactive coding: concurrent coding
agents, delegated subagents, CI/CD and test-triage automation, 24×7 monitoring
agents, embedded enterprise applications, and scheduled batch jobs. As a sanity
check, one agent streaming at ~70 tokens/second around the clock produces about
42.3M tokens/week, so 100M/developer/week is roughly **2.4 always-on
agent-equivalents per developer**.

### Three pricing behaviors drive the number

1. **The 272K trapdoor.** Past 272,000 input tokens, GPT-6 Astra reprices the
   *entire* request at 2× input/cache and 1.5× output. Agents accumulating
   repository context cross it routinely. Anthropic does not do this.
2. **Reasoning bills as output.** Thinking is on by default on both flagships and
   reasoning tokens bill at the output rate — 5× input. A 1% output assumption
   badly understates agent cost.
3. **Always-on automation runs cold.** CI/CD and monitoring jobs are spaced beyond
   the 5m/1h cache TTL, so each run re-sends context at full uncached rates.

That yields a blended frontier rate of **$22.62 per million tokens** and a
frontier-only annual cost of **$117,643 per developer** (5.2B tokens/year, about
$9,800/month). An 80–90% reduction corresponds to approximately
**$94,115–$105,879 in annual modeled savings per developer**:

| Scale | Annual tokens | Frontier baseline | Savings at 85% |
|---|---:|---:|---:|
| 1 developer | 5.2B | $117,643 | $99,997 |
| 10 developers | 52B | $1,176,432 | $999,967 |
| 100 developers | 520B | $11,764,319 | $9,999,672 |

Do **not** use Claude Opus 5 ($5/$25) as the frontier baseline — it sits a tier
below Astra and Fable 5.1 and understates the baseline by roughly a third. An
earlier draft of this package made that error and reported $40,560.

This is an illustrative model, not a guarantee. The assumptions must remain visible near the claim. Actual results should be calculated from request-time Smart Router usage and a source-dated comparison baseline.

Commercial conversion copy should use **Get enterprise support**, not a generic “Talk to Metrum.” Open source remains first-class; enterprise support, evaluation, and optional private-managed deployment are the paid offer.

## Primary links

- Product documentation: https://llm-api.apps.metrum.ai/docs/
- Open-source repository: https://github.com/sysadmin-metrum-ai/genai-smart-router
- Enterprise support: mailto:contact@metrum.ai?subject=GenAI%20Smart%20Router%20enterprise%20support

The hosted product-docs URL and recommended deep links returned HTTP 200 when
checked on September 7, 2026. The repository still identifies
`https://docs.metrum.ai` as canonical and its public-docs checker currently
forbids the hosted deployment hostname. Reconcile that policy before adding the
hosted URL to checked-in public content.