OpenRouter Luna 50% Off: GPT-5.6 Tier Cheat Sheet 2026

On 2026-08-03, OpenRouter applied a 50% promotional discount to GPT-5.6 Luna (permaslug openai/gpt-5.6-luna-20260709): prompt drops from $0.20/M to $0.10/M, completion from $1.20/M to $0.60/M. The same window repriced Terra / Terra-Pro to $1.00/M prompt and $6.00/M completion — a 60% cut versus the OpenAI list. Sol / Sol-Pro stayed at the $5/$30 list with no promo. This guide is the cheat sheet for deciding when the OpenRouter channel is the right call, when direct OpenAI wins, and how to wire the failover so a promo reversion does not break your cost model.

The 2026-08 pricing matrix

All numbers below were verified against https://openrouter.ai/api/v1/models and the per-model https://openrouter.ai/openai/gpt-5.6-{tier} pages on 2026-08-03. The OpenRouter price column reflects the live promo (where applicable); the OpenAI Direct column is the upstream list.

The headline number is Luna. Going from $1.00/M to $0.10/M puts it in the same cost tier as the budget-class models (Gemini 2.5 Flash at $0.075/$0.30, DeepSeek V4 Flash at $0.14/$0.28) while keeping the GPT-5.6 generation, reasoning controls, and the OpenAI Responses API surface. For teams that standardized on OpenAI's API surface and want to push budget-heavy workloads to a cheaper tier without rewriting prompts, this is the migration that pays for itself in a week.

Who should switch — and who should not

Switch to OpenRouter Luna if you fall in one of these buckets

• High-volume classification / extraction / routing — at $0.10/M prompt and $0.60/M completion, the cost-per-1K-requests drops to roughly $0.001-$0.005 for typical 2K-input / 200-output workloads. Routing traffic from a more expensive model to Luna via OpenRouter is a one-line routing change in any gateway.
• RAG agents with long system prompts — Luna supports reasoning and prompt caching on OpenRouter (5-10 minute TTL). For a 30K-token system prompt invoked 100 times per hour, the cache hit drops effective prompt cost to $0.01/M per cached read.
• Long-tail fallback for tool loops — if your Sol or Opus 4.8 call fails or rate-limits, falling back to Luna via OpenRouter keeps the agent responsive at 1/30 the cost of Sol.
• Eval pipelines — running 10K evals per day against a frontier model eats budget fast. Routing evals to Luna during development and only Sol for the final regression sweep is a common pattern that this promo makes much cheaper.

Stay on direct OpenAI for these workloads

• Zero-retention / HIPAA / SOC-2 paths — OpenAI's enterprise data handling (zero retention, training opt-out, EU data residency) is only available on direct OpenAI. OpenRouter cannot pass those contracts through.
• Long-TTL prompt caching (24h+) — direct OpenAI offers extended cache TTL for enterprise customers; OpenRouter honors the standard 5-10 minute window only.
• Sol / Sol-Pro frontier work — there is no OpenRouter discount on Sol, and OpenRouter adds a small latency overhead (single-digit ms typically, but real for time-critical paths). For the flagship tier, direct OpenAI is the right path.
• Workloads that need Microsoft Azure OpenAI Service contracts — OpenRouter does not proxy the Azure OpenAI endpoint, only the public OpenAI endpoint.

The decision cheat sheet

How to wire the failover (so a reversion does not break you)

OpenRouter's discount is a promo — at some point it will revert to the upstream list. The right architecture treats Luna as one tier in a tiered router, not a hard dependency. Three implementations, in order of effort:

Option 1 — Portkey (5 minutes)

Portkey's AI Gateway exposes OpenRouter as a virtual provider. You define one config that maps gpt-5.6-luna to openai/gpt-5.6-luna on OpenRouter today and openai/gpt-5.6-luna on OpenAI direct (or DeepSeek V4 Flash) tomorrow. The same client code, the same pricing rule, zero code change when the promo reverts.

{
  "strategy": { "mode": "fallback" },
  "targets": [
    {
      "provider": "openrouter",
      "model": "openai/gpt-5.6-luna",
      "weight": 1.0
    },
    {
      "provider": "openai",
      "model": "gpt-5.6-luna",
      "weight": 0.5,
      "conditions": [{ "type": "price_above", "value": 0.20 }]
    }
  ]
}

Portkey evaluates the conditions before each request, so if OpenRouter's Luna reverts above $0.20/M prompt, traffic falls back to direct OpenAI (or whatever fallback you list). This is the lowest-friction path.

Option 2 — Cloudflare AI Gateway (10 minutes)

Cloudflare AI Gateway accepts OpenRouter and OpenAI as providers and lets you write fallbacks in the request payload. The same pattern, hosted on Workers with zero egress fees. Best when your other agents already run on Cloudflare.

Option 3 — LiteLLM + a budget guard (15 minutes)

If you already run LiteLLM as your proxy, attach a per-model budget guard that throws a 429 when Luna's effective price crosses the threshold you set. The agent automatically retries the next-cheapest tier in your routing table. More code, but the most explicit.

How Luna stacks up against the 2026 budget tier

The budget-tier price war in 2026 has been brutal. With Luna at $0.10/M prompt, here is where it sits against the leading alternatives at the same capability band (single-step reasoning, 128K-200K context):

For pure classification or short extraction, Gemini 2.5 Flash and DeepSeek V4 Flash are still cheaper than Luna. The Luna discount matters most for workloads that need GPT-class instruction following, OpenAI's Responses API surface, or a drop-in replacement for code that already targets GPT-5.5. For those workloads, this is the cheapest credible path as of 2026-08-03.

Caveats worth knowing

• The discount is OpenRouter-specific. The same Luna tier on direct OpenAI is $1/$6. If you migrate a workload to OpenRouter for Luna's sake, you are opting into OpenRouter's contract terms, rate-limit policy, and routing choices. OpenRouter has been a reliable proxy since 2023 but the SLA is materially thinner than direct OpenAI's enterprise tier.
• Discount window. OpenRouter does not publish explicit end dates for these promos. The pattern across 2025-2026 has been: model launches with a 4-8 week promo, then either extends or reverts. The right move is to wire the failover now, not when the promo expires.
• Reasoning controls work but with subtle differences. Both Luna and Luna-Pro show supports_reasoning: true on OpenRouter, but the reasoning effort enum and the reasoning token accounting differ slightly from direct OpenAI's docs. Verify against your specific reasoning-effort setting if you depend on exact budgeting of reasoning tokens.
• Cross-region routing. OpenRouter routes Luna to whatever OpenAI region is cheapest at request time. For EU data residency requirements this is a problem — if you need the request to stay in EU, direct OpenAI with the EU endpoint is the only path.

Verdict

The 2026-08 OpenRouter discount on Luna is not a pricing gimmick. It moves GPT-5.6 Luna into the budget tier while keeping the GPT-5.6 generation's reasoning controls and OpenAI API surface. For teams that already run on the OpenAI ecosystem and want to push budget-heavy workloads to a cheaper tier without rewriting prompts or contracts, this is the migration that makes sense today. Three concrete actions for the next 24 hours:

1. Spin up a Portkey or Cloudflare AI Gateway route that maps your "cheap tier" to openai/gpt-5.6-luna on OpenRouter.
2. Run a 24-hour A/B on a non-critical workload: 50% Luna via OpenRouter vs 50% your current cheap tier. Compare quality, latency, and cost.
3. If Luna wins on cost without losing meaningful quality, make it the default and keep the previous tier as the failover for the day the promo reverts.

Routing GPT-5.6 Luna alongside Claude, Gemini, and DeepSeek on one key? If you also want a single dashboard for cost-per-search and cost-per-generation across providers and tiers, FreeModel exposes OpenRouter, direct OpenAI, Anthropic, and Gemini on one API key with usage breakdowns that surface when a tier reversion or rate-limit spike pushes your per-request budget past its guardrail — useful when the Luna promo ends and you need to decide within the hour whether to fall back to DeepSeek V4 Flash or stay on direct OpenAI.