Re-run your routing table: break-even math after GPT-5.6 Luna's 80% cut

Blended cost decides the route after the price cut

If your router still uses last quarter's rates, the GPT-5.6 price move changes the route before it changes the spend. The trade-off is paying a higher per-token rate for quality versus moving traffic to a cheaper model that still passes the task. Arithmetic: blended cost = input share x input rate + output share x output rate. OpenAI reduced the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% on July 30, 2026.

Pull current input and output rates for every candidate model. Build the input/output price ladder: one row per model and two rate columns. By August 20, 2026, GPT-5.6 Luna had the cheapest mainstream LLM API input rate at $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Sol was priced at $5 per million input tokens and $30 per million output tokens. That spread is the entry fee for quality.

Check whether request length crosses a pricing tier. The effective rate can differ from the headline rate. If production traffic sits near a boundary, use the tier rate for the whole request. A 400,000-token request through GPT-5.6 Terra was priced at $4 per million input tokens and $18 per million output tokens across the entire request. Long-context work moves the break-even in the opposite direction.

Weight each task by its real token mix from production logs, not demo prompts. If a task emits more output than input, the output rate dominates. DeepSeek R1's output price was $2.19 per million tokens, the lowest in BenchLM's selected frontier history set. The o3 model's output price fell from $40 to $8 per million tokens, an 80% reduction, in July 2026. An output-rate drop can flip a generation-heavy task before the input rate changes.

Compute the blended cost per task from the input and output shares in your traffic sample. Do not collapse input and output into one average rate; use the task's blended cost. Keep the incumbent only if its quality premium is worth the extra blended cost. Record the quality gate for each task. A route stays valid only if the cheaper model passes the task's acceptance test. If it fails, the cost saving is negative because retries, manual review, or user churn add cost. Set the route threshold as a cost value, not a model name, so the router can re-evaluate without a code change when prices move. A partial rollout verifies the threshold before full traffic.

Self-hosting breaks even only at volume

Estimate monthly token volume by task, split into input and output. A volume number without a task mix is a guess. A single monthly total hides the mix that drives cloud cost. If one task dominates output tokens, that task controls the break-even.

Treat the current token price index as the market trend signal. The frontier LLM token price index was 16 on September 3, 2026, 84% below its March 2023 baseline of 100. A falling index lowers the cloud path faster than a static self-host amortization.

Price the self-hosted path from your GPU index, utilization, and maintenance. Include hardware, power, network, and engineer time needed to keep the route healthy. The output is a monthly cost per million tokens, not just a server purchase price. Separate fixed cost from variable cost: hardware and base power are fixed, while incremental power, cooling, and support time are variable. The fixed part sets the floor; the variable part sets the break-even volume.

Solve for the monthly token volume where self-host total cost equals cloud blended cost at current rates. Below break-even, stay on the API. Above it, model the self-host savings against quality and latency risk. Re-run the worksheet when a major rate changes, task mix shifts, or the index moves materially. Keep the two-part worksheet in the same repo as the router config: route threshold from price ladder and task mix, self-host break-even from GPU index and monthly token volume, versioned with the route.