How often catalog IDs actually move
Catalog ID add/retire/price only. Does not observe silent same-id behaviour change. The only catch for silent same-id change is a cadence re-decide.
| Provider | Added | Retired | Diffs |
|---|---|---|---|
| mistralai | 5 | 0 | 1 |
| ~openai | 4 | 1 | 1 |
| deepseek | 2 | 0 | 30 |
| qwen | 1 | 1 | 10 |
| z-ai | 2 | 0 | 10 |
| nex-agi | 0 | 2 | 1 |
| inception | 1 | 1 | 1 |
| inference-net | 2 | 0 | 1 |
| sakana | 2 | 0 | 1 |
| nousresearch | 0 | 1 | 1 |
| inclusionai | 1 | 0 | 1 |
| tencent | 0 | 0 | 12 |
Live JSON: /runtimeai/catalog-deltas.json
· ?since=ISO-8601 filters events. This is public evidence for the top of funnel, not a paid gate.
About · Market Reality
Download PDFModel churn is out of control. Manual gates cannot keep up.
OpenRouter — a leading public model gateway — already lists 422 model routes — 7.8× end-2024 — with input prices spanning 15,000× and session costs ~17,000×. On the same non-linear path: ~549 by YE ’26 and ~1041 by YE ’27. The market sells speed and rate cards; enterprises need control. RuntimeAI is the continuous ship / still-trust gate.
OpenRouter is a leading public model gateway — we use its live catalog as an observable proxy for routable-market churn and pricing. That is not the full universe: private models, on-prem weights, direct lab-only APIs, and silent platform defaults sit outside this view.
OpenRouter catalog — actual + quadratic forecast through YE 2027
Solid blue = observed catalog (Jan ’25 – Aug ’26). Rust dashed = anchored quadratic forecast. Grey dashed = same curve fitted through history. Markers: Aug ’26 actual, YE ’26, Q1–Q3 ’27, YE ’27.
Method: least-squares quadratic on month-end catalog size (idx 0 = Jan 2025). Anchored to live 422 at Aug 2026. Extrapolation — not a commitment; silent defaults / delists not modeled.
Forecast milestones
| Horizon | Catalog | Adds / mo | vs today | Basis |
|---|---|---|---|---|
| Aug 2026 (actual) | 422 | ~32/mo | — | Actual |
| Dec 2026 · YE | 549 | ~40/mo | +127 (1.3×) | Quadratic forecast |
| Mar 2027 · Q1 | 657 | ~47/mo | +235 (1.6×) | Quadratic forecast |
| Jun 2027 · Q2 | 774 | ~55/mo | +352 (1.8×) | Quadratic forecast |
| Sep 2027 · Q3 | 903 | ~63/mo | +481 (2.1×) | Quadratic forecast |
| Dec 2027 · YE | 1,041 | ~72/mo | +619 (2.5×) | Quadratic forecast |
Implied monthly adds from a separate quadratic on monthly route additions (same window). Catalog column is the cumulative quadratic (primary).
Where the 422 routes come from
Top providers by route count on OpenRouter — 53 orgs aggregated on one public gateway. OpenAI alone is 95 routes.
Pricing is all over the place
Live OpenRouter rates (334 paid options + 22 free; excludes offline batch variants). Input $/M spans $0.01 → $150 (15,000×). Output spans $0.03 → $600 (20,000×). Same fixed agent session: ~17,000× cheapest→dearest.
Price & metric chaos (live OR)
| Metric | Spread |
|---|---|
| Input $/M | $0.01 → $150 · median $0.44 · 15,000× |
| Output $/M | $0.03 → $600 · median $1.78 · 20,000× |
| Session cost (fixed tokens) | Cheapest → dearest · ~17,000× |
| Context length | ~4K → 2M tokens · 488× |
| Out / in price ratio | ~0.8× → 12× (median 4×) — metering shape also varies |
| Free vs paid | 22 free · 334 paid (interactive; excludes batch) |
Even inside one lab
| Provider | Input $/M range | Spread |
|---|---|---|
| OpenAI | $0.03–$150/M in | 5,000× |
| Anthropic | $0.25–$30/M in | 120× |
| Qwen | $0.03–$2/M in | 67× |
| $0.05–$2/M in | 40× | |
| DeepSeek | $0.04–$1.12/M in | 28× |
- Cheapest: $0.01
inclusionai/ling-2.6-flash· $0.017ibm-granite/granite-4.0-h-micro· $0.019mistralai/mistral-nemo - Dearest: $150
openai/o1-pro· $30anthropic/claude-opus-4.7-fast· $30openai/gpt-5.5-pro
OpenRouter catalog facts
| Metric | Value |
|---|---|
| Live catalog (today) | 422 model routes |
| End of 2024 | 54 routes |
| Growth since end-2024 | 7.8× (+368) |
| Added in 2026 YTD | 202 new routes |
| H1 2025 → H1 2026 | 64 → 134 adds (2.1×) |
| Paid / offline batch / free | 395 / 61 / 22 |
| Distinct providers | 53 |
| Our freeze Jul → Aug | 464 → 552 model-costs ids (+88) |
What teams do today — not control
| Approach | Why it breaks |
|---|---|
| Inspect | Traces, LLM judges — fuel, not a ship decision |
| Quarterly bake-off | Stale before the next Flash bump (<30 days) |
| Price sheet | Picks among 422 routes; never proves your workflow still passes |
| Router alone | Chooses a slug; does not clear still-trust after the swap |
Churn accelerating — new OpenRouter routes per month
Full timeline Jan 2025 – Aug 2026*. Dashed curves = quadratic trends (white halo for contrast). End callouts ① / ② match the legend. Y-axis: routes added by OpenRouter created timestamp.
Source: OpenRouter /api/v1/models · +57 variant slugs and 31 revision slugs in 2026 YTD
Proof it’s out of control
| Pattern | Evidence |
|---|---|
| Catalog mass | 54 → 422 routes on OpenRouter |
| Gemini Flash line | 3.5 → 3.6 → 3.7 in ~12 weeks (May–Aug) |
| Anthropic Opus | 4.6 → 4.7 → 4.8 → Opus 5 in ~5.5 months |
| OpenRouter July 2026 | 42 new routes in 31 days |
| Silent defaults | MAI into Copilot — no new public eval target |
Control is the empty seat
10-lab tracker — net-new vs revision vs platform
Same timeline · revisions = 71% of 2026 YTD tracker events. Quadratic trends show revision volume climbing through 2026.
Control vs chaos (2026 YTD)
| Without control | With ship / still-trust gate | |
|---|---|---|
| Swap candidates | 422 live OpenRouter routes (public proxy) | Same paths · pinned rubric · before/after |
| Decision unit | $/M tokens · arena rank · judge score | Pass rate + USD on your scenario |
| Cadence | Weekly drops · silent defaults | Continuous clearance on every change |
| Failure mode | Green dashboard · wrong invoice | Hard-fail on failure modes you refuse to ship |
Sources: OpenRouter live API (GET /api/v1/models) · AI Release Tracker · RuntimeAI Industry signals · scorecard freeze diff (+88 catalog ids Jul→Aug). Updated 23 Aug 2026.