# One model was 41% of spend: what routing actually saves (markdown mirror)

Source: https://aiapi-pro.com/blog/model-routing-savings-2026

Method: every paid text call in 30 days (model, tokens, cost), recomputed at the cheapest text model's price (deepseek-v4-flash: $0.08/1M in, $0.17/1M out).

| | Value |
|---|---|
| Actual cost | $188.74 |
| All-cheapest cost | $45.53 |
| Upper bound | **75.9%** reduction |

## Where the money went

| Model | Calls | Actual | At cheapest | Multiple |
|---|---|---|---|---|
| kimi-k3 | 146 | **$76.65** | $2.10 | 36.5x |
| deepseek-v4.1-flash | 2,920 | $59.64 | $15.24 | 3.9x |
| deepseek-v4-flash | 5,377 | $24.47 | $24.47 | 1.0x |
| glm-5.2 | 249 | $6.18 | $0.39 | 15.9x |
| deepseek-v4-flash-vision | 286 | $5.88 | $1.50 | 3.9x |
| qwen3.8-max | 177 | $5.19 | $0.22 | 23.5x |
| glm-5.3 | 146 | $3.85 | $0.24 | 16.3x |
| kimi-k2.8-preview | 289 | $0.93 | $0.08 | 12.2x |
| qwen3.7-max | 112 | $0.75 | $0.04 | 19.1x |
| mimo-v2.6-pro | 130 | $0.67 | $0.12 | 5.4x |

146 calls (~1% of traffic) carried 41% of the bill.

## Do not chase the 75.9%

The all-cheapest counterfactual assumes quality is irrelevant. It is a ceiling: routing is the largest single lever, bigger than prompt trimming or output tuning.

## Routing map

| Task shape | Tier | Example | $/1M in |
|---|---|---|---|
| Classify, extract, tag | Cheapest | deepseek-v4-flash, glm-5.3-flash | $0.080 |
| Everyday coding | Cheap-mid | glm-4.7-flash (free), doubao-seed-2.0-lite | $0.075-$0 |
| Hard coding | Mid | deepseek-v4.1-flash, kimi-k2.7-code | $0.308-$0.67 |
| Long documents (256K+) | Cheap long-context | doubao-seed-2.0-lite, glm-5.3-flash (1M) | $0.075-$0.08 |
| Hard reasoning | Flagship | kimi-k3, glm-5.3, qwen3.8-max | $1.25-$2.90 |

## Routing without rewriting code

Same endpoint, per-request model id:

```bash
curl https://aiapi-pro.com/v1/chat/completions \
  -H "Authorization: Bearer $NOVAI_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[...]}'
```

Router MCP tool: `pick_model`. Hosted endpoint: https://aiapi-pro.com/api/mcp (no install, no key).

## What this does not say

- Cheap models are not as good on hard tasks.
- Not every request should be routed (classifier latency/failure cost).
- One gateway's traffic, weighted toward agent/retrieval workloads.

What it does say: on the same tokens the spread between cheapest and most expensive model here is 36x, and a single model choice moved 41% of the bill.

## FAQ

**How much can model routing save?**
In 30 days of real traffic, routing every text call to the cheapest model would have cut the bill from $188.74 to $45.53, a 75.9% reduction. That is an upper bound, not a recommendation: quality-sensitive calls would suffer.

**Which model dominated spend?**
Kimi K3: 146 calls (about 1% of traffic) cost $76.65, which is 41% of all text spend. The same tokens on the cheapest model would cost $2.10.

**Should I just use the cheapest model for everything?**
No. Route by task: cheap models for classification, extraction and short answers; mid-tier for coding and longer reasoning; flagship only when the task genuinely needs it. The 75.9% figure shows the ceiling of the lever, not a target.

**How do I route between models without rewriting my code?**
Every model is reachable through the same OpenAI-compatible endpoint, so routing is a per-request model id change. NovAI also exposes a router MCP tool (pick_model) and a cheapest-per-task page that rank models by comparable cost.