How volume pricing actually works
Half-price frontier tokens sound like they should be impossible. Here is the mechanism And, more usefully, here is why the discount differs per model and moves over time.
The mechanism
Model providers sell capacity in volume tiers. A buyer that commits to large aggregate throughput gets a materially better rate than one developer with a single API key. A gateway pools the demand of many developers, qualifies for the better tier, and passes part of the difference on.
Nothing about the model changes. You call claude-opus-5, you get Claude Opus 5. The only thing that differs is who is being billed and at what rate.
What you are taking on
| Risk | What it means | How to limit it |
|---|---|---|
| Uptime | You depend on the gateway's availability, not just the vendor's | Keep your direct key configured as a fallback |
| Model routing | You are trusting that the model you asked for is the model you get | Verify responses on a known benchmark before production |
| Continuity | Rates and availability can change | Do not prepay more than you would be comfortable losing |
| Compliance | Your data passes through a third party | Check the data retention policy if you handle regulated data |
Why the discount is not the same on every model
A single headline percentage would be wrong, so we publish the per-model rate instead. Three different situations sit behind those numbers:
- Verified against the vendor's own page. The Claude family and most of the GPT-5.6 family run at 50% of documented list price. These are the rates we can show side by side with a source.
- Official rate already discounted. GPT-5.6 Sol comes out at 37.5% off input and 25% off output rather than 50%, because OpenAI's own $4/$20 rate is promotional, halving it would overstate the gap.
- Platform rate, no discount claimed. DeepSeek, GLM, Kimi, Qwen and MiniMax bill by peak and off-peak windows and by cache hit. One percentage would mislead, so for those we publish the gateway rate and claim nothing.
What actually changes in your code
Two values: the base URL and the key. Your request bodies, model IDs and SDK calls stay exactly as they are.
export OPENAI_BASE_URL=https://aicomp.ai/v1
export OPENAI_API_KEY=sk-...Anthropic-native clients append /v1/messages themselves, so their base has to be set to the host without the /v1 suffix — otherwise the request goes to /v1/v1/messages and 404s. That single detail is the most common reason a setup "doesn't work".
How to confirm the route actually took
Do not trust the settings screen. Ask the endpoint what your key can see:
curl -s https://aicomp.ai/v1/models \
-H "Authorization: Bearer $OPENAI_API_KEY"A list of models means both the base URL and the key are working. An authentication error means the key is wrong; a 404 means the base URL is. Then send one real request and compare it against the same prompt run directly (same model ID, same answer). Rates verified 2026-11-02.
What it costs on a real workload
Per-million prices are hard to reason about, so here is one concrete month: 50M input tokens and 10M output tokens, roughly a small team running coding agents through the working day. Only models with a verifiable official rate are listed.
| Model | Direct at list | Through the gateway | Monthly difference |
|---|---|---|---|
| Claude Fable 5 | $1,000 | $500 | -50% |
| GPT-6 Astra | $1,000 | $500 | -50% |
| GPT-5.6 Sol | $400 | $275 | -31% |
| Claude Opus 5 | $500 | $250 | -50% |
| Claude Opus 4.8 | $500 | $250 | -50% |
| Claude Sonnet 4.6 | $300 | $150 | -50% |
| GPT-5.6 Terra | $220 | $110 | -50% |
| Claude Sonnet 5 | $200 | $100 | -50% |
| Claude Haiku 4.5 | $100 | $50 | -50% |
| GPT-5.6 Luna | $22 | $11 | -50% |
How to switch safely
- Start with non-critical traffic and compare output quality on your own tasks.
- Keep the base URL in an environment variable so you can revert without a deploy.
- Monitor spend in both places for the first month.
- Do not move regulated or highly sensitive workloads through any third-party gateway.