On July 30, OpenAI cut the price of GPT-5.6 Luna by 80 percent. Not 10. Not 20. Eighty.
Luna went from $1 per million input tokens to $0.20 — twenty-one days after the model launched on July 9. Output dropped from $6 to $1.20. Terra, the mid-tier in the GPT-5.6 family, got a quieter 20% cut. Sol, the flagship, didn't move.
These cuts apply automatically. No account changes needed. If you're already using OpenAI's API or Codex CLI, your next invoice is already cheaper.
Why It Happened
This isn't generosity. It's a competitive response.
Chinese open-weight models have been eating into OpenAI's market share for months. Kimi K3 — Moonshot AI's 2.8-trillion-parameter open-weight model — runs at $3 per million input tokens and outperforms most US closed models on coding benchmarks. Xiaomi's MiMo-2.5-Pro costs less still. OpenRouter data shows that 61 percent of developer token consumption on the platform now comes from Chinese open-weight models.
Three weeks into GPT-5.6's life, OpenAI had to move.
The South China Morning Post's July 30 headline read "OpenAI blinks in face-off with Chinese rivals." That's the right framing. This is what a market correction looks like when open-source pressure actually lands.
What This Changes for Your Org
The price drop matters most not for chatbots — those were already affordable — but for automated, high-volume workflows where tokens accumulate in the background without anyone watching the screen.
Here's what "high-volume" actually looks like in practice: processing 500 documents daily, summarizing each one with roughly 2,000 tokens per output. That's about 1 million tokens per run. At the old Luna price of $1/M, that was $1/day, $30/month. At $0.20/M, it's $0.20/day, $6/month. That's not marginal — for an org paying $30/month on a workflow, that's the difference between keeping it on and cutting it.
The bigger shift is for workflows that weren't running because the token math felt uncomfortable:
Nightly document processing pipelines. Intake forms, grant reports, board submissions, vendor invoices — automatically categorized, summarized, and flagged for review. Processing 10,000 pages per night now runs under $30/month. That was a $150 line item six weeks ago.
Meeting transcription plus structured output. Feeding transcripts through a model to pull out action items, decisions, and who-owns-what. If your team runs 30 meetings per week, the transcript processing costs less than a coffee.
Email triage at scale. Inbound request classification, donor communication routing, intake screening. At $0.20/M tokens, most small org volumes are effectively rounding errors.
OpenAI confirmed that automatic code review in Codex CLI — one of the more token-heavy automated tasks — is now roughly ten times cheaper than before the cut.
Two Things to Watch
First: pricing velocity is now alarming. Luna launched July 9 and dropped 80% in 21 days. Before 2026, a frontier model's working life before deprecation averaged about 18 months. It's now closer to six. The lesson isn't "lock in to Luna." It's "build your automations so you can swap the underlying model without rebuilding everything." Use model aliases in your code, not hardcoded version strings. When the next cut — or deprecation — hits, you'll absorb it in a config change rather than a rebuild.
Second: this competitive pressure isn't stopping. The Chinese open-weight models that forced this cut are still releasing, and their prices are still dropping. OpenRouter now lets you route to the cheapest model for a given task class in real time. If you're processing non-sensitive content, that flexibility is worth building toward.
What to Do This Week
Pull up whatever you're paying for ChatGPT Work, Copilot, or API access. Find the line items consuming the most tokens. Then ask one question: at $0.20 per million input tokens, what did I consider too expensive to automate six weeks ago?
That audit usually takes an afternoon. The workflows you build from it might run for years.
If you want to map your highest-value automation candidates and work out what's worth building first, that's usually the starting point of a sprint with us. Get in touch.