What changed on July 30, 2026
OpenAI announced a price drop for two of the three GPT-5.6 models. All prices are USD per million tokens, Standard tier:
gpt-5.6-luna: input $1 to $0.20, output $6 to $1.20, cached input $0.10 to $0.02. An 80 percent cut.gpt-5.6-terra: input $2.50 to $2, output $15 to $12, cached input $0.25 to $0.20. A 20 percent cut.gpt-5.6-sol: unchanged at $5 input and $30 output.
Source: OpenAI’s API pricing page and the price drop announcement. Always confirm against the provider before billing on any number, including ours.
Luna is now the cheapest current OpenAI model on output
The interesting comparison is Luna against gpt-5.4-nano, which used to be the budget floor. Nano costs $0.20 input and $1.25 output. Luna now costs $0.20 input and $1.20 output. Same input price, slightly cheaper output, and Luna is the more capable model.
Output tokens usually dominate a generation heavy bill, so for chat, summarization, and drafting workloads the floor of the OpenAI lineup effectively moved up a capability tier at the same price. If you routed traffic to gpt-5.4-mini ($0.75 and $4.50) purely for cost, that route now deserves a second look: Luna undercuts mini on both sides. The mechanics of picking the cheapest capable model are covered in model routing for LLM apps.
The part that bites: your stale price tables
Here is the honest part. Weckr recomputes cost server side from tokens times a price table, and for six days after the cut our table still had the old Luna numbers. Every Luna call we priced in that window was recorded at five times its real cost. Our weekly price watcher caught the drift on its first live run against published rates, we verified against OpenAI’s pricing page, and shipped the fix the same day, to the server, both SDKs, and the docs.
A stale table after a cut like this does three bad things:
- Spending caps fire early. A user capped at $10 of Luna usage gets blocked at $2 of real spend.
- Margins look worse than they are. Users on Luna heavy features show inflated cost, so profitable accounts can get flagged as underwater.
- Routing decisions invert. The table says Luna is expensive, so you route away from the model that just became the best deal in the lineup.
If your surprise is in the other direction, a bill that overshot instead, start with why your OpenAI bill is so high.
How to never carry stale prices again
Two approaches, and they compose:
- Read prices live instead of hardcoding. We publish a free live JSON feed of LLM prices at useweckr.com/pricing.json, generated from the same table Weckr bills with.
- Automate the diff. Our weekly watcher compares every price table against published rates and opens a reviewed pull request when something moves. Nothing touches billing math without a human approving the diff.
If you use Claude Code, the Weckr Claude Skillsread that live feed, so asking your assistant “what does gpt-5.6-luna cost” returns the post-cut number, not whatever its training data remembers.
FAQ
How much did OpenAI cut GPT-5.6 prices?
On July 30, 2026 OpenAI cut GPT-5.6 Luna by 80 percent, from $1 input and $6 output per million tokens to $0.20 and $1.20, and GPT-5.6 Terra by 20 percent, from $2.50 and $15 to $2 and $12. GPT-5.6 Sol stayed at $5 and $30. Cached input dropped in proportion: Luna to $0.02 and Terra to $0.20 per million.
Is GPT-5.6 Luna now cheaper than gpt-5.4-nano?
On output tokens, yes. Luna now costs $1.20 per million output tokens against $1.25 for gpt-5.4-nano, while keeping the same $0.20 input price. Since output usually dominates a chat or generation bill, Luna is effectively the cheapest current OpenAI model for output heavy workloads, with noticeably more capability than nano.
What should I do about my model routing after the cut?
Re-run the math on any route that avoided Luna for cost reasons. A workload you pushed down to gpt-5.4-mini or nano to save money may now be cheaper on Luna at higher quality. Also check any cost tracking, budgeting, or cap logic that hardcodes prices: at the old numbers it now overstates Luna spend by five times.
Why do stale price tables matter so much?
Any tool that computes cost from tokens times a hardcoded price silently drifts when the provider changes rates. After this cut, a stale table shows Luna traffic at five times its real cost, which can trigger spending caps early, flag profitable users as unprofitable, and push you to route away from a model that is now a bargain. We caught our own table doing exactly this six days after the cut.
How do I keep model prices accurate automatically?
Either read prices from a live feed instead of hardcoding, or automate the diff. We publish a free live JSON feed at useweckr.com/pricing.json generated from the table Weckr bills with, and we run a weekly watcher that compares every price table against published provider rates and opens a reviewed pull request when something moves.
Keep reading
Track the real number, not the remembered one
Price cuts are only good news if your cost tracking notices them. Weckr recomputes cost per user and per feature from a table that is watched weekly and published live, so your margins move the day the market does. See it on the live demo, or start with the AI cost and margin guide.