On 30 July 2026, OpenAI made GPT-5.6 Luna 80% cheaper, three weeks after the model family launched on 9 July. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from $1 and $6, according to The New Stack. Terra fell 20%. The flagship, Sol, did not move. Vendors usually hold prices for months after a launch, so a cut inside a month is the story inside the story.
What changed on 30 July
- GPT-5.6 Luna: $0.20 input and $1.20 output per million tokens, down from $1 and $6 (80% cheaper).
- GPT-5.6 Terra: $2 input and $12 output, down from $2.50 and $15 (20% cheaper).
- GPT-5.6 Sol: unchanged at $5 input and $30 output.
Developers already on Luna needed no code changes. Their bills just dropped.
Why the price fell
OpenAI credits infrastructure. Its engineers rewrote production GPU kernels, cutting serving costs by about 20%, redesigned Sol’s speculative decoding to make generation over 15% more efficient, and reworked the agent runtime to reuse more cached prompt computation. The New Stack adds a second driver: cheaper open-weight models from Chinese labs such as Moonshot are forcing commercial providers to compete on price, not just performance.
The longer trend points the same way. Epoch AI found that the price of reaching a fixed performance level has fallen between 9x and 900x per year, depending on the benchmark, with about 40x per year for GPT-4-level scores on PhD-level science questions. That analysis dates from March 2025, and Epoch noted the fastest declines came in the most recent year, so it was unclear whether they would persist. Cheaper is the direction. The speed is not a promise.
Why agents change the math
A chatbot makes one call per question. An agent often makes dozens or even hundreds of model calls to finish one task, which is why The New Stack argues serving costs can matter more than small benchmark differences. So the number to watch is cost per task, not cost per token.
Here is an illustration with assumptions we picked: an agent run of 40 calls, each with 6,000 input tokens and 800 output tokens. That is 240,000 input and 32,000 output tokens per run. At the prices above:
- Sol: $2.16 per run, or $21,600 a day at 10,000 runs.
- Terra: about $0.86 per run, or $8,640 a day.
- Luna after the cut: about $0.09 per run, or $864 a day. Before the cut, the same run cost about $0.43.
The gap between Sol and Luna is 25x. It ignores prompt caching, retries and tool costs, and it assumes Luna finishes the work Sol would. If the cheaper tier needs three attempts, the savings shrink fast.
What open-weight data adds
Kilo analyzed 10,643 AI code reviews between 22 June and 23 July 2026. Open-weight models handled about 74.8% of its tokens at about 16% of user-facing cost, which works out to roughly 16x cheaper per token. Kilo is clear that this compares traffic, not identical tasks, and that its data shows what models flagged, not whether the flags were right. Read it as a signal that cheap tiers are production options, not as proof they match flagship quality.
Five moves for your own bill
- Price the task, not the token. Calls per task times tokens per call times price.
- Log calls per task. You can’t cut what you haven’t counted.
- Split steps by difficulty. The New Stack’s view is that most agent steps don’t need a flagship model, so send easy steps to a cheap tier and keep the expensive one for hard calls.
- Re-price every quarter. A cut landed three weeks after launch. Last quarter’s spreadsheet is stale.
- Test before you switch. A lower price says nothing about quality. Run the cheap tier on your own tasks and count failures and retries.
Frequently asked questions
How much did OpenAI cut GPT-5.6 API prices?
OpenAI cut GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% on 30 July 2026. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from $1 and $6. Terra costs $2 and $12, down from $2.50 and $15. The flagship Sol stayed at $5 and $30.
Why are LLM prices falling?
Serving efficiency and competition. OpenAI credits rewritten GPU kernels, faster speculative decoding and more prompt caching, while The New Stack points to cheaper open-weight models from Chinese labs such as Moonshot. Epoch AI measured the price of reaching a fixed performance level falling 9x to 900x per year, depending on the benchmark.
How do I estimate the cost of an AI agent task?
Multiply the calls per task by the tokens per call and the model’s price. In our example, 40 calls with 6,000 input and 800 output tokens each cost about $2.16 on GPT-5.6 Sol and about $0.09 on Luna at the new prices. Ignore caching and retries at first, then add them from real logs.
Should I switch to open-weight models to save money?
Only after testing them on your own tasks. Kilo’s data shows open-weight models handled about 75% of its tokens at about 16% of the cost, and two led on critical code-review findings. Kilo also says that is aggregate traffic rather than a controlled test, and it does not prove the findings were correct.
Leave a Reply