What actually changed in GPT-5.6 pricing on 30 July 2026?
On 30 July 2026 OpenAI cut GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%, leaving flagship Sol unchanged. Luna input went from US$1.00 to US$0.20 per million tokens and output from US$6.00 to US$1.20. The gap between the cheapest and the most expensive tier is now 25 times, not 5.
The numbers come straight from OpenAI's own announcement thread dated 30 July 2026.
GPT-5.6 API pricing after the cut (USD per million tokens)
--- Sol: input US$5.00, cached input US$0.50, output US$30.00. Unchanged.
--- Terra: input US$2.00 (was US$2.50), cached US$0.20 (was US$0.25), output US$12.00 (was US$15.00).
--- Luna: input US$0.20 (was US$1.00), cached US$0.02 (was US$0.10), output US$1.20 (was US$6.00).
OpenAI also shipped a Fast mode for Sol in the API: up to 2.5 times the speed of standard processing at twice the standard price, with no change to the model's intelligence. In the API this is the same thing as setting the request's service tier to priority.
One more change matters if you never touch an API key. Auto-review in the ChatGPT app and the Codex CLI moved from GPT-5.4 to GPT-5.6 Luna, which OpenAI expects to cost roughly ten times less to run. The safety-check layer that approves risky actions is now a cheap model, not a mid-tier one.
Why did OpenAI cut the price, and does that tell you anything useful?
OpenAI attributed the cut to efficiency gains it found by pointing GPT-5.6 at its own serving stack: 20% lower serving costs from production GPU kernel improvements, and more than 15% better token-generation efficiency from improved speculative decoding. Those savings were passed to the API, Codex and ChatGPT.
That matters for one practical reason. This was an infrastructure win, not a capability change.
Luna did not get smarter on 30 July. It got cheaper. Any workflow you rejected three months ago because Luna produced weak output will still produce weak output, only for one fifth of the money.
The reverse is also true and more interesting. Any workflow you rejected because the volume was too expensive to justify is now worth re-testing tonight.
What is the difference between GPT-5.6 Sol, Terra and Luna?
GPT-5.6 is one model generation split into three capability tiers. Sol is the flagship built for long-horizon reasoning that persists across files, tests and follow-up fixes. Terra is the lower-cost everyday tier that OpenAI describes as competitive with the previous GPT-5.5. Luna is the fastest, cheapest tier for high-volume, low-reasoning work.
The naming is deliberately durable. The number, 5.6, marks the generation. Sol, Terra and Luna are tiers that can each advance on their own schedule, which is why you should build your workflow around the tier name rather than around a version string.
What each tier is actually for
--- Luna: summarising, labelling, tagging, extracting fields, classifying inbound messages, drafting a first scaffold, translating boilerplate. Work where a small quality difference does not change the outcome.
--- Terra: the everyday production tier. Scoped writing tasks, first-pass review, structured analysis of a document you will still read yourself.
--- Sol: the tasks you would otherwise block half a day for. Multi-step reasoning, work that has to stay coherent across many files or a long chain of decisions, anything where a wrong answer costs you a client conversation.
Most practitioners do the opposite of this by default. They pick one tier, usually the best one, and run everything through it because switching feels like admin.
How much does a real workload cost across the three tiers?
Take a realistic monthly job: 1,000 documents summarised, each about 4,000 input tokens with a 600-token summary out. That is 4 million input tokens and 600,000 output tokens. On Luna the run costs US$1.52. On Terra it costs US$15.20. On Sol it costs US$38.00.
The arithmetic is worth seeing because the ratios are cleaner than most people assume.
--- Luna: 4 x US$0.20 plus 0.6 x US$1.20 = US$1.52, roughly HK$12 at the pegged rate.
--- Terra: 4 x US$2.00 plus 0.6 x US$12.00 = US$15.20, exactly 10 times Luna.
--- Sol: 4 x US$5.00 plus 0.6 x US$30.00 = US$38.00, exactly 25 times Luna.
Before 30 July the same Luna run cost US$7.60. So the cheap lane got five times cheaper while the expensive lane stayed where it was.
Read that as a signal about where the industry is heading rather than a one-off discount. The same pattern is showing up in how vendors bill for AI features generally, which we covered in What Are AI Credits?
The practical takeaway: if a task is genuinely routine and you are running it on Sol, you are paying a 25x premium for reasoning you are not using.
How do you route work across tiers without writing code?
You do not need a router. You need a triage habit and one place where the model name is a variable, not a hard-coded assumption. In practice that means splitting each recurring task into a cheap first pass and an expensive second pass, then only paying for the second pass on the items that failed the first.
Here is the pattern that survives contact with real work.
Step 1. Split the task into extract and decide. Nearly every knowledge task has a mechanical half and a judgement half. Pulling the five key figures out of a supplier quote is extraction. Deciding whether to accept the quote is judgement. Send the first half to Luna, keep the second half on Terra or Sol.
Step 2. Put the cheap model first, not second. Have Luna produce a structured summary of all 40 inbound emails, then read only the summary on the ones flagged as needing a decision. The expensive model never sees the 33 emails that turned out to be routine.
Step 3. Make the model a field in your automation, not a constant. In Zapier, Make or n8n the model name sits in a dropdown on the AI step. Duplicate the step, change the tier, run the same 20 records through both, and compare the outputs side by side before you commit. If you have not built one of these flows yet, start with How to Automate Your Work With AI Without Writing a Single Line of Code.
Step 4. Escalate on a rule, not on a feeling. Ask the cheap model to end every output with a confidence line and a reason. Route only the low-confidence items to the expensive tier. This is the single change that makes tier routing feel automatic instead of fiddly.
Where does tier routing break down?
Tier routing fails in four predictable places: tasks that look mechanical but carry hidden judgement, long chains where a cheap early error compounds, prompts written for a strong model and inherited by a weak one, and any workflow where the cost of one bad output exceeds a month of API savings.
The hidden-judgement trap. Tagging support tickets by topic looks mechanical. Tagging them by urgency is not, because urgency depends on who the customer is and what they said last week. Cheap models handle the first and quietly guess at the second.
Compounding errors. If step one summarises a document badly, steps two through five build on a bad summary and every later output looks confident and wrong. Cheap tiers belong at the edges of a chain, not in the middle of one.
Inherited prompts. A prompt that works on Sol often assumes the model will infer structure you never specified. Move it to Luna and the missing structure shows up as sloppy output. The fix is not a better model, it is a more explicit prompt, which is the point of context engineering.
Asymmetric stakes. If a wrong figure in a client report costs you the account, the US$36 you saved this month is not a saving. Keep client-facing final passes on the tier you trust and spend the savings elsewhere.
One more caution on Fast mode. It buys latency at double the price and changes nothing about output quality. Turn it on for something a human is waiting for on screen. Leave it off for a batch job that runs at 3am.
Try it now: the 15-minute tier audit prompt
Run this once against your own list of recurring AI tasks. It takes about 15 minutes and usually finds two or three tasks sitting on a tier three times more expensive than they need to be. Paste it into any GPT-5.6 tier, Claude or Gemini, since it is a reasoning task about your workflow rather than a model-specific trick.
Copy this prompt:
You are helping me audit which AI model tier each of my recurring tasks should run on. I will give you a list of tasks. For each one, do the following.
1. Split the task into its mechanical component (extract, summarise, classify, reformat) and its judgement component (decide, prioritise, persuade, evaluate trade-offs). State both in one line each. If a task has no judgement component, say so explicitly.
2. Recommend one of three tiers for the mechanical component and one for the judgement component, using this guide: CHEAP for high-volume work where a small quality difference does not change the outcome; MID for scoped production work I will still review myself; TOP for multi-step reasoning or any output a client sees unedited.
3. Name the single failure mode most likely to appear if I run the whole task on CHEAP, in one sentence, and describe what that failure would look like in the output so I can spot it.
4. Write one escalation rule I can apply mechanically, in the form: escalate to the higher tier when [observable condition in the cheap output].
Finish with a table ranking my tasks by how much money I am likely wasting, highest first, and state what evidence you used to rank them. If a task is too vague to judge, ask me one clarifying question instead of guessing.
My recurring AI tasks are: [list 5 to 10 tasks, with rough monthly volume for each]
Then act on exactly one recommendation today. Duplicate that automation step, switch the tier, and run 20 real records through both versions. Compare the outputs yourself instead of trusting the benchmark table.
The takeaway: cheap is now a design decision, not a compromise
The 30 July price cut did not make any model better. It changed where the sensible default sits. When the cheapest tier costs one twenty-fifth of the flagship, running everything through the flagship stops being caution and starts being waste.
The practitioners who get the most out of this will not be the ones who switch everything to Luna. They will be the ones who can say, for each task they run weekly, exactly which half is mechanical and which half needs judgement.
That split is a skill, and it is the same skill whether the model in front of you is from OpenAI, Anthropic or Google. Model names change every few weeks. Knowing what your own work actually requires does not.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Turn a Tier Audit Into a Working System
Knowing which tier a task belongs on is step one. Building it into something that runs every day without you watching is step two. We'll walk you through every step, from task triage to model selection to a workflow that runs on its own.
UD has supported Hong Kong businesses through 28 years of technology change.
