10decodersTwo businesses, one workload, a 10× difference in the bill.
Same traffic. Same quality of answer where it matters. The only difference is whether anything in the stack decides, for each request, which model to call. The rates below are public. So is the arithmetic.
8,000 calls / 32M input tokens / 4M output tokens. Rates current at August 2026. Work through it yourself below.
Trusted by leading enterprises and healthcare teams
Most small businesses are paying flagship prices for filing-clerk work
Nobody decides to do this. It happens because the first pilot got wired to whichever model was in the news that month, it worked, and nothing since has given anyone a reason to look at it again.
Then the prices moved. On 30 July 2026 OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%, and Anthropic made Claude Sonnet 5's introductory $2/$10 rate permanent instead of letting it rise to $3/$15 in September. Inside OpenAI's own flagship family, the gap between the cheapest and most expensive tier is now 25 times on input. It was 2.5 times before that afternoon.
A stack built before August is therefore already mispriced, and it will be mispriced again by Christmas. That is the condition, not a blip to wait out.
Why this hits smaller companies harder
A bank has a team whose job is to notice a rate change and re-tier the traffic. A thirty-person company has one person who built the integration between other work, and they are not watching vendor pricing pages. The saving is available to both. Only one of them is set up to take it.
The two ladders, as they stand
Both vendors sell by the million tokens, split between what you send and what the model writes back. Output is the expensive half and it dominates most bills.
OpenAI
The GPT-5.6 family became generally available on 9 July 2026 with a 1.05M token context window across all three tiers. Three weeks later, two of the three prices came down.
| Model | Model ID | Input | Output | Cached input | Where it fits |
|---|---|---|---|---|---|
| GPT-5.5 Pro | gpt-5.5-pro | $30.00 | $180.00 | — | Premium reasoning |
| GPT-5.6 Sol | gpt-5.6-sol | $5.00 | $30.00 | $0.50 | Current flagship |
| GPT-5.4 | gpt-5.4 | $2.50 | $15.00 | — | Previous standard |
| GPT-5.6 Terra | gpt-5.6-terra | $2.00 | $12.00 | $0.20 | Balanced −20% on 30 Jul |
| GPT-5.4 Mini | gpt-5.4-mini | $0.75 | $4.50 | — | Legacy efficient |
| GPT-5.6 Luna | gpt-5.6-luna | $0.20 | $1.20 | $0.02 | High volume −80% on 30 Jul |
Anthropic
Claude Fable 5 arrived on 9 June 2026 as the most capable model Anthropic sells openly. Below it, Opus 5 handles complex agent work and Sonnet 5 holds the contested middle.
| Model | Input | Output | Cache read | Context | Where it fits |
|---|---|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | $1.00 | 1M | Long running agents |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 | 1M | Complex agentic work |
| Claude Sonnet 5 | $2.00 | $10.00 | $0.20 | 1M | Everyday production |
| Claude Sonnet 4.6 | $3.00 | $15.00 | $0.30 | 1M | Previous generation |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | 200k | Fast and high volume |
What the headline rates leave out
Four modifiers change the real number, and none of them appear in a side by side comparison.
Repeated text bills at a tenth
Both vendors charge roughly 10% of the base input rate for a cached prefix. Your system instructions, policy text and tool definitions are the same on every call, so most of what you send should never be paid for at full price twice.
Overnight work costs half
Anything that can wait 24 hours goes through a batch endpoint at 50% off both input and output. Document processing queues, nightly summaries and backfills belong here rather than in the live path.
OpenAI's meter roughly doubles past 272k tokens
Above about 272,000 input tokens the whole request reprices: Sol goes to $10/$45, Terra to $4/$18. Claude 4.6 and later include the full 1M window at standard rates, which is the sharpest structural difference between the two.
Output costs six times input
Across every GPT-5.6 tier the ratio is exactly 6:1, and reasoning tokens bill as output whether or not you ever see them. A verbose assistant costs multiples of a terse one answering the same question.
CloudZero surveyed 260 finance leaders in 2026 and found 42% had approved AI spending without a reliable cost projection. These four modifiers are most of the reason projections miss.
Put your own numbers in
This runs the published August 2026 rates against whatever volume you type. It is arithmetic, not a forecast, and no part of it depends on believing anything we say.
One model for everything
What most pilots do by default.
$280.00/ monthRouted by task
Simple work to GPT-5.6 Luna, standard to Claude Sonnet 5, hard to Claude Opus 5.
$46.84/ monthRouted and cached
Same routing, with the repeating part of each prompt read from cache.
$27.62/ monthThe honest caveat
All of this assumes the routing decision is right. Send a hard request to the cheap tier and you get a bad answer, and the retry costs more than routing it properly the first time. Building a classifier that is right often enough is the actual engineering problem, and it is why we would rather measure it against your data than quote you a percentage.
There are three routes, and most buyers do not realise they are choosing
Chat seats
Around $20 to $30 per person per month from either vendor. You get a capable assistant, admin controls, and a promise that your data is not used for training.
It automates nothing. A person still has to be sitting there deciding what to ask.
Straight to the API
Cheapest tokens available, and entirely viable if you have an engineer who enjoys this. You get real automation.
What you also get is the maintenance. Routing logic written in July is already stale, and the person who wrote it has other work.
An orchestration layer
Tokens at the routed rate, plus a platform that holds the routing, caching, retrieval, permissions and audit trail in one place.
Model choice becomes configuration. When prices move again, somebody changes a setting rather than opening a ticket.
Seats and orchestration are not competing purchases
We hear "we already have ChatGPT" in most first calls, and the honest answer is that you probably need both. Seats put a good assistant in front of a person who then decides what to do with it. Orchestration handles the work that happens whether or not anyone is at their desk: the document queue overnight, the ticket that arrives at two in the morning, the check that has to run on every single transaction rather than the ones somebody remembers.
You cannot hold a seat licence accountable for a workflow. That is the line worth drawing.
Six costs that sit above the tokens
The token bill is the part everyone compares and the smallest part of what a working system costs. In the model above, $27.62 of monthly tokens supports a workflow whose surrounding costs are an order of magnitude larger.
| Cost | What it looks like when nobody owns it | What an orchestration layer does about it |
|---|---|---|
| Routing decay | Traffic still going to models that were the sensible choice in June and are now more expensive than their replacements. | Model selection is configuration. Re-tiering after a price move is a settings change rather than a sprint. |
| Retrieval quality | The assistant is confidently wrong because chunking and reranking were never tuned to your documents. People quietly stop using it. | Grounded retrieval with a citation on every answer, and groundedness measured in production rather than assumed. |
| Evaluation | Nobody can say whether last week's prompt change helped or hurt. Regressions get found by customers. | An evaluation harness from the first sprint, so quality is a number somebody looks at. |
| Access control | The assistant answers from a document the person asking should not be able to read, because permissions did not follow the content into the index. | Role based access carried from the source system through retrieval into the answer. |
| Audit trail | An auditor, insurer or customer asks how a decision was reached and there is no record. | The evidence trail is produced by the system as it runs rather than reconstructed before a review. |
| Adoption | The thing works and nobody uses it. Deloitte put trained teams at 2.3 times the task completion rate of untrained ones in 2025. | Assessment, roadmap and enablement built into the engagement rather than sold afterwards. |
Where we would push back on a price conversation
Your token bill will land somewhere between twenty and three hundred dollars a month. That is not the number worth negotiating. The number worth negotiating is what it costs you the first time the system gives a wrong answer to a customer, or answers from a file that person was never meant to see.
The layer that makes the published discounts reachable
CheiAI is our AI platform. It takes unstructured content and returns explainable answers, using retrieval, computer vision and models tuned to your domain. It does not compete with OpenAI or Anthropic. It consumes them.
Model agnostic by architecture
CheiAI orchestrates OpenAI, IBM Watson and custom models side by side through an integration gateway. You pick or switch models on cost, compliance or performance grounds. This is the property that makes the routing arbitrage reachable at all, and it turns the next price cut into a configuration change.
It runs inside the stack you already bought
If your team has standardised on Microsoft, CheiAI sits between Copilot Studio at the front and Azure AI Foundry underneath. You authenticate through Entra ID and point it at SharePoint, Azure Blob and your databases. Nothing gets ripped out.
Your data stays in your tenant
CheiAI operates inside your existing Azure environment, and your permissions and access policies stay in force. Where the requirement is stricter, it deploys into your own cloud account, your VPC, or a fully isolated environment with no external model calls at all.
Answers you can defend
Every response is grounded in content you have approved and comes back with its source cited, with tracing and evaluators for groundedness and safety running in production. For a company without a compliance function, that is the difference between an AI you can stand behind and one you cannot.
What we would do first
A short diagnostic before anything gets built. We audit what you currently spend across seats and API calls, profile your actual traffic into simple, standard and hard, and produce a routed cost model against your own volumes rather than the illustrative ones above. You end up with a costed decision, including the option of not proceeding. If we cannot show a model below your current run rate, you have lost a few days and learned something useful about your own data.
From there, the usual shape is two workflows into production on CheiAI, with retrieval grounded in your sources, routing across tiers, caching configured, the evaluation harness in place, access control and audit trail wired from the start, and your team trained on it.
When we are the wrong answer
We would rather say this on a web page than forty minutes into a call.
- You are under ten people with no repeating workflow. If the work is genuinely different every time, buy chat seats. Orchestration is overhead you do not need yet.
- Your knowledge lives in people's heads. Grounded AI needs something to ground in. If there is no usable document set, you have a documentation problem first, and that is a different project.
- You are running a few hundred calls a month. Below that volume the routing saving is smaller than the cost of the meeting about routing. Come back when it is worth it.
- You want to buy a product and switch it on. Then buy a product. Our engagements involve integrating with your systems, which is a commitment of your time as well as ours.
- You want a guaranteed percentage before we start. We do not quote improvement figures without a named client behind them. The diagnostic exists because the honest answer is that we should measure it first.
The five that come up every time
Is it not cheaper to go straight to OpenAI?
Direct gives you the cheapest tokens and the most expensive engineering. The rates are public and anyone can call them. The questions are who writes the routing logic, who rewrites it when prices move, and who is accountable when a wrong answer reaches a customer. Prices moved four times between June and August 2026.
Should we wait for AI prices to settle?
They are not settling, they are spreading out. The gap between the top and bottom of OpenAI's flagship family went from 2.5 times to 10 times in a single afternoon in July. Waiting does not remove the decision. It just means paying flagship rates for volume work while you think about it.
How do we know it will not hallucinate at a customer?
Answers are grounded in content you have approved and come back with a citation, and groundedness is evaluated in production rather than assumed. You will also see in the diagnostic where retrieval is weak against your own documents, which is the honest place to find that out.
Our data cannot leave our environment. Does that rule this out?
No. CheiAI runs inside your Azure tenant using your existing controls, or into a fully isolated environment with no external model calls where that is the requirement. We hold ISO 27001 for information security and ISO 9001 for quality management.
You work with IBM and Dedalus. We are twenty people.
The governance was built for regulated enterprise clients, which means you get it without paying to develop it. That is the one real advantage of not being our first customer in a regulated industry.
Where the numbers came from
Rates change monthly. Everything here was checked on 25 August 2026 and we date the page each time we revise it.
- Anthropic, pricing and model documentation, platform.claude.com. All Claude rates, cache multipliers, batch rates and context windows.
- CloudZero, "OpenAI API pricing in 2026: every model after the July price cuts", published 10 July 2026 and updated 20 August 2026.
- OpenAI, model release notes and pricing announcements, 9 July and 30 July 2026.
- Deloitte, "State of Generative AI in the Enterprise", Q4 2025. Task completion multiplier for trained teams.
