The short version: on 30 July 2026, OpenAI repriced two of its three GPT-5.6 tiers — roughly 80% off the cheapest model and about 20% off the mid tier, with the frontier tier's price unchanged but faster. The company attributed it to inference and serving efficiency gains, and framed the intent plainly: make high-volume work economical at much greater scale. If you're running AI in your business, the actionable consequence isn't "switch models." It's that any automation you costed and rejected on unit economics deserves a second look, because the arithmetic behind the rejection has moved by a multiple. And — this is the part that saves you from an expensive detour — for most of the projects we see, the model bill was never the binding constraint anyway.
What exactly happened?
- The lowest-cost tier dropped to roughly $0.20 per million input tokens and $1.20 per million output, from about $1 and $6 — a cut of around 80%
- The mid tier dropped to about $2 and $12 per million, from roughly $2.50 and $15 — around 20%
- The frontier tier's price was left alone; the improvement there was speed, not cost
- The family had only reached general availability three weeks earlier, which tells you the cuts came from engineering efficiency rather than a scheduled pricing cycle
- OpenAI's stated goal was better performance per dollar across enterprise workloads — pay the right price at each stage rather than the top price everywhere
Two caveats before anyone rebuilds anything. Published list prices move, in both directions, and competitive pressure is why this one moved — so treat any specific figure as a snapshot and check the current pricing page before you commit a budget. And a cheaper model is not a better model; the cut applies to the tiers built for volume, not for judgement.
Which kinds of work does this actually unlock?
The shape to look for is high volume, low judgement, structured output. Those are the workflows where token cost is a real line item, because you're running the same small operation tens of thousands of times. Where a cut of this size changes an answer:
- Enriching or normalising a large catalogue — attributes, categorisation, and description drafting across every SKU rather than the top 200
- Classifying and routing every inbound support ticket, instead of sampling
- Reading and structuring documents at volume: supplier invoices, purchase orders, spec sheets, inbound RFQs
- Deduplicating and cleaning CRM or product records across the full history, not the recent slice
- Monitoring jobs that check something on a schedule — every listing, every day, rather than a weekly spot-check
- Anything where you previously ran a sample and extrapolated because processing everything was too expensive
The pattern under all of those
Every one is a case where you had to choose between coverage and cost, and chose partial coverage. That's the real unlock — not doing a new thing, but stopping the sampling on something you already do. Full coverage is usually where the value was hiding, because the exceptions you never processed are the ones costing you money.
So should I go rebuild everything on the cheap tier?
No, and this is where the money usually gets wasted. Run the honest breakdown of a failed AI project and inference is rarely the largest number. Integration work, the human review loop, evaluation and monitoring, and the ongoing maintenance of a thing that now sits in the middle of a real process — those dominate, and none of them fell 80% last week. Industry research keeps making the same point from the other direction: Gartner has projected that a large share of agentic AI projects will be scrapped by 2027, and the reasons cited are escalating costs, unclear business value and inadequate risk controls — not the price of tokens.
The predictable mistake
Reading a price cut as permission to move everything to the cheapest model. What follows is a quiet quality drop on the tasks that actually needed judgement, discovered a month later by a customer rather than by you — and a support and rework bill considerably larger than the savings. Cheap tokens don't make a weak model do careful work; they make a suitable model do bulk work affordably. Those are different sentences.
The version that works
Route by task, not by policy. Bulk classification, extraction and first-pass drafting go to the cheap tier. Anything customer-facing, irreversible, or requiring judgement stays on the capable one. Measure quality per task with a small evaluation set so you can prove the routing is safe rather than assume it. This is also what makes the next price change a config edit instead of a rewrite.
How do I re-run the maths without wasting a week?
Go and find the list of things you decided not to automate. Most teams have one, even if it lives in someone's head. For each item, compute the cost per unit of work — per ticket, per SKU, per document — rather than a monthly total, because per-unit is the number that decides whether coverage scales. Then compare it against what the manual version costs you in loaded hours, and be equally honest about the build and review cost, which hasn't changed.
Rank whatever survives by volume times unit saving, take the top one, and run it end to end on real data for a week before committing. One workflow, measured, beats five estimated. If the top candidate still doesn't clear the bar at the new prices, that's a genuinely useful answer too — it means the constraint was never cost, and you can stop revisiting it every time a vendor cuts prices.
Where we fit
The hard part of this isn't the arithmetic, it's the portfolio judgement: which workflows are worth automating now, which were correctly rejected and should stay rejected, and which need a different design rather than a cheaper model. That's the question a Fractional Head of AI & Digital is for — someone senior enough to say no to the exciting project and yes to the boring high-volume one, who owns the roadmap and the vendor decisions rather than handing you a deck. Retaining us that way means the next price cut, model release or protocol announcement gets evaluated against your actual numbers by someone accountable for the outcome, instead of triggering another round of internal debate.
Sources
- CNBC — OpenAI cuts prices for two of its GPT-5.6 AI models as companies grow sensitive to costs (30 July 2026)
- PYMNTS — OpenAI cuts prices on select models to make high-volume work economical
- Forbes — OpenAI cuts GPT-5.6 pricing up to 80%, as AI costs come under scrutiny
- Unite.AI — OpenAI cuts API prices on its two cheaper GPT-5.6 tiers (per-tier figures and GA timeline)
- DEVOPSdigest — Gartner: 40% of enterprise apps will feature task-specific AI agents by 2026
