We Just Cut Our Own AI Coding Bill 83% in 12 Hours — Here's the Data

TL;DR — Our pitch is "smart routing": each request gets classified by intent (planning, implementing, debugging, testing, documenting) and routed to the cheapest capable model. Opus only when truly needed. But last week's audit caught us: 80%+ of traffic was bypassing the router because client defaults (Claude Code / Cursor / Codex) hardcoded model: "claude-opus-4-7" and we honored it. 12 hours of fixes later, real data: per-request cost dropped from $0.19 to $0.122 (-36%), Opus share of spend from 80%+ to 45%, honest savings from 60% to 91%, ~$90K/month saved. Here's the process and the lessons.

1. The embarrassing thing we found

We sell smart routing. Each request should get classified by intent — planning, implementing, debugging, testing, documenting — and routed to the cheapest model that can handle it. In theory: Opus only for the heavy reasoning, V4-Pro for implementation, V4-Flash for tests and docs. Most everyday coding requests should NOT use Opus.

But last week's product audit, on 5,000 real requests:

Requests where the router actually ran:    15%
Requests that bypassed the router:         85%

Why? Users' clients (Claude Code / Cursor / Codex) ship with default configs that hardcode model: "claude-opus-4-7" — and we dutifully forwarded those to Opus. Smart routing got skipped entirely.

This was a product-level embarrassment: the core value we were selling didn't get a chance to fire on most client default configurations.

2. Our call: force auto on Smart-tier plans

The fix was a product positioning decision more than a technical one. Two options:

A. Educate users. Email everyone, ask them to flip model: "auto" in their config. B. Override the model field server-side. Smart-tier plans always go through smart routing, regardless of what the client sends.

We picked B. Reason: smart routing IS the product. If a customer paid for the Smart tier and their client default bypassed the routing, that's our responsibility, not the user's.

But B has a precondition: give users who genuinely want manual model control a way out. We already had the Direct plan (pay-as-you-go, you pick the model, we charge provider list + 15%). It became the explicit-mode escape hatch.

So the product positioning sharpened into two distinct lanes:

• Smart plans (Starter / Solo / Pro / Studio / Team): subscription, smart routing, whatever model you send is rewritten to auto
• Direct plan: pay-as-you-go, the only plan where your model: choice is honored

3. The data, 12 hours later

After the fix, 12 hours of real production traffic — 2,094 requests:

| Phase | Share | Routed to (top 3) | |---|---:|---| | debug | 41.7% | Opus 44% + Sonnet 22% + GPT-5.5 18% + V4-Pro 15% | | implement | 32.2% | V4-Flash 62% + V4-Pro 16% + GPT-5.4 12% | | test | 17.2% | V4-Flash 47% + Sonnet 34% + V4-Pro 19% | | plan | 5.1% | V4-Pro 77% + Sonnet 8% | | document | 2.5% | V4-Flash 94% | | small_edit | 0.9% | gpt-5-mini 100% |

Each phase landed on the right model mix. Opus is still the first pick for genuinely reasoning-heavy debug work, but everywhere else it stepped aside for cheaper models.

4. The cost numbers

12-hour actual spend:            $254.89
12-hour all-Opus baseline:     $2,819.63
Honest savings:                    91.0%

Annualized:

• Actual cost: ~$15K/month
• All-Opus baseline: ~$169K/month
• $154K/month saved

Versus our pre-fix rate:

• Pre-fix 12h extrapolated: ~$1,780
• Post-fix 12h: $254
• $1,500+ saved every 12 hours, ~$90K/month

5. Three counterintuitive findings

A. Planning doesn't need Opus

We'd assumed planning required Opus-tier reasoning depth. Reality: of 107 plan-phase requests in 12 hours, 77% routed to V4-Pro and 0% to Opus.

Reason: V4-Pro and Opus both score top-tier on plan-quality benchmarks, but V4-Pro costs ~10× less. The balanced strategy picks V4-Pro on quality/cost ratio. User feedback didn't show any quality regression.

Takeaway: "Opus is mandatory for planning" was an assumption, not a fact.

B. One cheap model carried 32% of traffic for $2

V4-Flash (DeepSeek's flash tier) handled 659 requests in 12 hours for a total cost of $2.13. Same volume on Opus would have been $300+.

It's especially good on long-context Chinese-language sessions, where 98% prompt-cache hit rates make per-token cost essentially free.

C. Most "routing problems" are actually "config problems"

We'd spent weeks improving the routing classifier — multilingual support, tool-call inference, multi-turn fallbacks. But what actually moved savings from 60% to 91% wasn't the algorithm. It was:

1. Forcing client requests through the routing pipeline (the silent rewrite)
2. A sensible default for ambiguous cases (default to "implement" → cheap models)

The classifier was already good enough. The problem was that most requests never got classified.

6. Three lessons

1. Product defaults beat algorithm precision. A clever router doesn't matter if client defaults bypass it. If you're selling routing, you have to actually route.
2. Build a real escape hatch for the "I want control" segment. Smart plans = forced auto, Direct plan = explicit selection. Two clear lanes, no fuzzy middle.
3. Public data is the moat. Most LLM resellers don't publish real audit data because their differentiation is just "cheaper API." Our differentiation is the routing logic itself, which transparency only strengthens.

7. What this means for you

If your team uses Cursor / Claude Code / Codex and your monthly bill is over $1,000, there's a good chance you're stuck in the same trap — every request defaulting to Opus or GPT-5.5 because that's what your client config says.

How to check: pull your last 30 days of usage. Look at the share of cost going to Opus (or GPT-5.5 / GPT-5.4). If it's over 60%, and your team isn't full of architects doing nothing but planning, the excess is likely overspend.

CodeRouter gives you 1M tokens free to try. Pro plan ($99/month) includes 30M tokens + 500K Opus quota — enough for most mid-sized teams for a full month. Try it for a week. If your bill doesn't drop to ≤30% of what you were paying, we'll refund you.

---

All data here comes from real production audits on our own infrastructure. We publish honest retrospectives because they're more persuasive than marketing copy.