GPT-5.6 Price Cut: Luna $0.20/$1.20, Terra $2/$12 (Jul 2026)

On July 30, 2026, OpenAI cut API pricing across the GPT-5.6 family. The short version: GPT-5.6 Luna drops 80% to $0.20 input / $1.20 output per 1M tokens, GPT-5.6 Terra drops 20% to $2.00 / $12.00, and the flagship GPT-5.6 Sol stays at $5.00 / $30.00. Sol instead gains Fast mode — the former Priority processing, renamed — running up to 2.5x faster at 2x the price. ChatGPT subscription prices are unchanged.

The Luna cut is the one that changes architecture decisions. Before July 30, Luna and Terra were 2.5x apart on price; now they are 10x apart. Any design that routes "most work to the cheap model, hard cases to the big one" has a materially different cost profile than it did yesterday.

Cached input fell by the same percentage. Luna's cached input is now $0.02 per 1M — another 90% off its already-reduced standard input. For RAG-style workloads that resend a fixed system prompt or reference document on every call, input cost effectively rounds to zero.

This is a price change, not a model swap. Model IDs and capabilities are identical to before:

The 272K-token surcharge threshold survives the price cut: requests with more than 272K input tokens are billed at 2x input and 1.5x output. A million-token window is available, but routinely filling it puts you in a tier that costs double the headline rate. If you work with long documents, plan for chunking, summarization, and caching around that threshold rather than against it.

The same announcement renamed Priority processing to Fast mode and raised the speed at which Sol runs. Fast mode delivers up to 2.5x the throughput of Standard processing at 2x the Standard price ($10 input / $60 output per 1M). OpenAI is explicit that this is a speed change only — no change in intelligence.

- Applies to: GPT-5.6 Sol only (not Terra or Luna)
- How to request it: set service_tier to fast in the API call (priority still works for compatibility)
- Price: 2x Standard ($10 / $60 per 1M)
- Quality: identical to Standard — speed and price are the only variables

Existing code using service_tier: "priority" keeps working, but the meaning shifted from "priority in the queue" to "faster processing." Worth updating comments and internal docs so the next engineer isn't reasoning from the old name.

Percentages are less useful than your own numbers. Take a workload of 10M input and 2M output tokens per month — a typical internal chatbot or document-processing batch — with no cache hits:

The last row is the interesting one. In a routed setup, 86% of the post-cut bill is now the Sol portion. The optimization target has moved: squeezing the cheap tier further barely moves the total, while shaving a few percentage points off the fraction of traffic escalated to Sol moves it a lot.

Post-cut, Luna lands near the bottom of the closed-model price range. For reference, alongside the same-generation models covered elsewhere on this site (per 1M tokens, from each vendor's published rates):

Luna now undercuts Gemini 3.5 Flash-Lite ($0.30 / $2.50) on both input and output. For the Gemini side of the comparison see Gemini 3.6 Flash explained; for the self-hosted open-weight alternative see DeepSeek V4 hardware requirements. A 10x drop in the hosted price also moves the break-even point for self-hosting — if you sized a GPU budget against the old rate, that calculation deserves a rerun.

1. Lower your routing threshold. With a 10x gap instead of 2.5x, work you previously sent to Terra "just to be safe" is worth retesting on Luna with an escalate-on-low-confidence second pass. The expected value of that two-stage design improved substantially overnight.
2. Caching matters less in absolute terms. At $0.02 per 1M cached input, the absolute dollars recoverable from input-side optimization on Luna are tiny. Engineering time may return more when spent trimming output tokens — at $1.20 they cost 6x what input does.
3. Audit where Sol is actually required. Sol's price didn't move, so its share of your total bill rises automatically. Identifying steps that pass on Terra or Luna is now the highest-leverage cost reduction available.

- Model IDs are unchanged — no code change is needed for the new rates to apply to requests after the cut
- Update the unit-price constants in internal cost models and dashboards (stale values will diverge from actual billing)
- Find every service_tier: "priority" call site and note the Fast mode rename in your docs
- Before moving work to Luna, run a 1–2 week quality A/B on real prompts rather than switching blind
- Check your request logs for the input-token distribution — anything above 272K is billed at the surcharge rate

This article reflects OpenAI's announcement and reporting as of July 31, 2026. API pricing changes frequently — verify current rates on the official pricing page before you build or quote against them. For the pre-cut landscape and a Grok 4.5 comparison, see GPT-5.6 vs Grok 4.5 API pricing.