1300 166 138 Tell us what you need
News / OpenAI

GPT-5.6 Luna is now 80 per cent cheaper

By Monika Tantau • Published 4 months ago
GPT-5.6 Luna is now 80 per cent cheaper

Key takeaways

• GPT-5.6 Luna drops from $1 and $6 per million tokens to $0.20 and $1.20, an 80 per cent cut.
• Terra falls 20 per cent to $2 and $12 per million; the Sol flagship stays at $5 and $30.
• A new Fast mode runs Sol up to 2.5 times faster at twice the price, replacing Priority Processing.
• Subscription prices are unchanged, but the cuts also stretch usage allowances in Codex and ChatGPT Work.

Most Read

OpenAI cut the price of two GPT-5.6 tiers on 30 July 2026. Luna, the smallest and fastest model in the family, dropped 80 per cent. Terra, the mid tier, fell 20 per cent. The flagship, Sol, did not move at all.

That asymmetry is the story. OpenAI is not making its best model cheaper. It is making the cheap end cheap enough that work which never justified automation now does, and it lands at an awkward moment for Australian finance teams that spent the first half of the year trimming AI budgets.

What the new rate card says

Luna now costs $0.20 per million input tokens and $1.20 per million output, down from $1 and $6. Combined, that is $1.40 per million against $7 before. Terra moves from $2.50 and $15 to $2 and $12. Sol stays where it launched, at $5 input and $30 output. All figures are in US dollars.

There is one addition on the other side of the ledger. Fast mode, available in the API for Sol only, runs up to 2.5 times faster than standard processing and costs twice as much. It replaces the previous Priority Processing option, and per OpenAI’s announcement, subscription prices are unchanged, though the lower rates also stretch how far usage goes in Codex and ChatGPT Work.

Why the flagship did not move

OpenAI attributed the cuts to efficiency rather than competition. Better routing keeps hardware busy, optimised production software generates tokens more efficiently, and smarter context management stops agents redoing work they have already finished. The company also said Sol contributed by autonomously optimising its own production systems.

The competitive reading is harder to ignore. Chinese open weight models have undercut the US labs all year, and Microsoft has started promoting its own MAI models as cheaper alternatives, as reported by CNBC. Cutting the volume tiers while holding the flagship protects margin where reasoning quality still commands a premium.

How it compares with Claude

Against Anthropic’s ladder, Luna now sits well below anything Claude offers on price. Claude Opus 4.8 runs $5 input and $25 output, and Fable 5 is the premium option at $10 and $50. On Artificial Analysis’ Intelligence Index, Luna scores 51 against 60 for Fable 5, so the capability gap is real, and so is the twenty five fold difference in output price against Sol.

That is the trade Australian teams now have to price properly. Nine points of index score matters on a contract summary and is close to irrelevant on a job that sorts three thousand support tickets into six buckets. Buying reasoning you do not use has always been waste. It is simply more visible waste at these rates.

How it compares with Claude

Why Australian businesses should care

The practical effect is a change in what is worth automating. At $7 per million tokens combined, high volume and low judgement work such as classifying inbound email, extracting fields from invoices or drafting routine replies often sat just outside the business case. At $1.40 it usually does not.

That matters here specifically, because Australian firms spent the first half of 2026 reining in AI spending as costs outran budgets. The common response was to cut scope. The cheaper option now is to route the routine steps of a workflow to Luna and reserve Sol for the steps that genuinely need reasoning, which is the same tiering logic behind our document automation work.

One caution before anyone rebuilds a forecast. Per token price is not per task cost. A cheaper model that burns more tokens, retries more often or needs more review can still cost more per finished job. The number to model is cost per completed task on your own workload, not the rate card.

What to check this week

Anyone running GPT-5.6 in production should re read their tier assignments, because the arithmetic that justified them changed on 30 July. Workloads parked on Terra for cost reasons may now belong on Luna, and workloads rejected outright may now clear the bar. Our breakdown of the three GPT-5.6 tiers has the updated numbers.

It is also worth watching whether the rest of the field follows. Anthropic moved the other way this month, announcing a 50 per cent rise on Claude Sonnet 5 for 1 September, though it cancelled that rise on 11 August. If the volume tier turns into a price war and the frontier does not, Australian buyers will end up running mixed stacks by default rather than by design.

The tiering decision is now worth more than the vendor decision. Teams that route per task will spend a fraction of what teams that standardise on one model spend, and the gap widened considerably on 30 July.

Scroll to Top