GPT-5.6 Luna Drops 80%: The Price Cascade That Resets Your Routing Tier Decisions

OpenAI cut GPT-5.6 Luna pricing by 80% on July 30, 2026, and simultaneously renamed Priority Processing to Fast mode with a backward-compatible API change. For routing teams, these two moves together change the cost math that drives tier selection across the entire GPT-5.6 family — and they put Luna in direct price competition with Chinese providers for the first time.

The operational implication is not subtle: if your routing policy was built on the original Luna pricing from July 9, the thresholds that determined when to fall back from Terra to Luna have shifted by an order of magnitude. Operators who have not recalculated are routing to a more expensive tier than their workloads require.

The before/after pricing table every routing policy depends on

GPT-5.6 pricing on July 9 vs. July 30 (standard processing, 1M tokens):

These are the raw prices from https://developers.openai.com/api/docs/pricing. Long-context input and output carry a 2× multiplier for Sol and Terra above the standard short-context price.

The Sol:Luna cost ratio is now 25:1 on input (was 5:1). The Terra:Luna ratio is 10:1 on input (was 2.5:1). If your routing logic used the old ratio to decide when Luna was "good enough," those thresholds are stale by 4–5×.

Fast mode replaces Priority Processing: the service_tier rename every proxy must know

On the same date, OpenAI renamed Priority Processing to Fast mode. The API change is backward compatible but introduces a new canonical value:

// Before: only option for priority
{ "service_tier": "priority" }

// After: both work identically for gpt-5.6
{ "service_tier": "fast" }
{ "service_tier": "priority" }  // still works

For gpt-5.6-sol specifically, Fast mode delivers up to 2.5× faster latency than Standard processing at 2× the price. For earlier GPT models, the response field service_tier in the response object still returns "priority" whether you sent "fast" or "priority" — so any downstream parsing of the response's service_tier field must handle either value.

What breaks if you don't update:

• Routing proxies that validate service_tier values against an allowlist need to add "fast" to avoid rejecting new client requests that adopt the new name.
• Logging pipelines that aggregate by service_tier will see a split between historical "priority" events and new "fast" events — they map to the same tier.
• Fast mode has its own ramp-rate limit: if your traffic exceeds 1M tokens in a short window, some requests may be downgraded to Standard and billed at Standard rates. The response returns service_tier: "default" when this happens. Your retry and billing reconciliation logic should handle this case.

The cross-provider routing decision: Luna vs. DeepSeek V4 Flash vs. Qwen

The view that only a router has: Luna at $0.20/$1.20 now lands in the same cost band as the leading Chinese provider models.

Approximate published prices at time of writing (per 1M tokens, input/output):

Luna is now within a 1.25–1.5× cost premium of DeepSeek V4 Flash and within 1.5–2× of Qwen-Plus — before accounting for regional latency, compliance requirements, and per-provider reliability characteristics. For teams routing globally, this changes the calculation from "Luna is a premium domestic option" to "Luna is price-competitive for high-volume workloads that need OpenAI API semantics or residency in OpenAI infrastructure."

The original GPT-5.6 launch article on TheRouter established the routing policy for Sol, Terra, and Luna. That policy framework still applies; what has changed is the input data that drives the tier selection: Luna is no longer a fallback that costs 5× less than Sol — it costs 25× less.

How to update your routing policy today

Three things to check in your routing configuration:

1. Recalculate tier thresholds. Any rule that routes to Terra when cost-per-token exceeds a threshold, with Luna as fallback, needs to be revalidated. The Luna:Terra cost ratio is now 10:1 on input, not 2.5:1.

2. Add "fast" to your service_tier allowlist. If your gateway validates or logs service_tier values, add "fast" alongside "priority" and "default". Treat them as synonyms in aggregation and billing logic.

3. Monitor for Fast mode ramp-rate downgrade. Implement handling for service_tier: "default" in responses from Fast mode requests. This is not a failure — it's a billing event: the request was served at Standard speed and should be billed at Standard rates, not Fast mode rates. Log and reconcile accordingly.

OpenAI's engineering post describes Luna's performance on professional tasks (Agents' Last Exam) as outperforming Fable 5 at roughly 99% lower cost. The relevant operator question is not whether Luna is better than Fable 5 — it is whether Luna is now good enough for the tasks you currently route to Terra, and whether the 10× cost difference justifies the evaluation work to confirm it.

What TheRouter users should watch or try

If you route to the GPT-5.6 family via TheRouter, update your model tier selection logic using the new prices above. Luna at $0.20/$1.20 is a material change in the cost model for high-volume agent workloads — the kind where context is short, tasks are well-specified, and the model is implementing rather than planning. Review your routing rules against the docs to ensure your tier-selection conditions reflect the updated Terra:Luna cost ratio.