GPT-5.6 Pricing: OpenAI Kept the Same Price and Promised a Discount Anyway
OpenAI's new flagship GPT-5.6 Sol costs $5 input and $30 output per million tokens, the same as GPT-5.5. The promised saving sits in a number buyers cannot check.
The Bright Recap
OpenAI released GPT-5.6 on 9 July 2026 with its flagship model, Sol, priced at $5 input and $30 output per million tokens. That is the same rate its predecessor GPT-5.5 charged, so the saving OpenAI advertises depends on the model consuming less, not on the price being lower.
Anthropic prices Claude Fable 5 at twice that rate. Both companies now compete on how much their software eats to finish a job, which is the one figure a buyer cannot read off a price list before signing.
To know more about this topic, read our related articles:
- Why GPT-5.6 arrived late
- AI's hidden cost problem
- What AI actually saved the NHS
- Financial technology explained
Bright Answers
Is GPT-5.6 cheaper than the model it replaces?
The price is identical. GPT-5.6 Sol charges $5 per million input tokens and $30 per million output tokens, the same as GPT-5.5. A lower bill arrives only if the model uses fewer tokens on your work.
How can a company check the saving is real?
By measuring what one finished task costs, on its own work, before and after the switch. Vendor efficiency figures are produced on the vendor's own tests, and OpenAI states its cost estimates come from simulation.
Software has never been able to charge you for how hungry it is. GPT-5.6 pricing changes that, and it does so without changing a single number on the price list. OpenAI released its new flagship model, Sol, on 9 July 2026 at $5 per million words of input and $30 per million words of output, which is what the previous model charged. The company is still promising customers a lower bill.
The promise rests on appetite. Sam Altman told CNBC that the new model gets through coding work using 54% fewer of the units it bills for. A model that eats less costs less to run, even when the menu price holds. Every finance director now approving an AI budget is being sold a discount that exists inside the product rather than on the invoice.
Cantica is The Bright Minded's Fin-Tech Intelligence Layer
The unit nobody in the meeting can count
AI companies bill by the token, which is roughly a fragment of a word. Your company pays for the words it sends the model and the words the model sends back. Nobody in a procurement meeting can say in advance how many words a piece of work will take, which means the price per token tells you almost nothing about what the year will cost.
That gap is where the entire competition now lives. OpenAI's own release notes set the rates at $5 and $30 for Sol, $2.50 and $15 for the mid tier, and $1 and $6 for the cheapest, and the same page carries a footnote admitting that its cost and speed figures were produced by simulating the models offline, with real results varying substantially. The claim and the caveat are printed on the same document.
OpenAI vs Anthropic: the cost issue that could make Claude lose clients
Anthropic sells the opposite proposition, and its price list says so. Claude Fable 5 costs $10 per million input tokens and $50 per million output, double Claude Opus 4.8 and double OpenAI's flagship. The argument for paying it has always been that a better model finishes harder work in fewer attempts, so the total is lower even when the rate is higher.
OpenAI has now taken that argument and printed a number on it. It claims its flagship beats Fable 5 on a coding index while producing less than half the output and costing about a third less, a comparison it ran itself. Anthropic's answer sits in the same set of published tables, where the Claude models lead on the benchmark closest to real engineering work by roughly fifteen points. A buyer who reads only the rate card sees a rival charging half as much and saying it wins.
Why OpenAI can afford to hold the line
OpenAI does not need this price list to make money. The company's confidential filing for a stock market listing came alongside reporting that it loses more than a dollar for every dollar it earns, and most of what it does earn comes from subscriptions rather than from the developers buying tokens. A rate card funded by investors is a marketing asset, and one held flat while claiming a 54% efficiency gain is the cheapest discount a company can advertise, because it gives away nothing per unit sold.
Anthropic has the harder and more honest position. It charges a premium and asks customers to believe the work will take less of it, which is a claim the customer can test on their own workload within a month. But does the premium still buy what it used to?
The trap both companies have built
A vendor billing by the token, whose product is engineered to use fewer tokens, is competing against itself. Sell the efficiency and you sell less of what you charge for. The escape is to price the finished job instead of the fuel, and both companies are already edging towards it through seats and subscriptions.
Anyone who has bought software in financial technology recognises the shape of this. Metered pricing gets optimised down by the customer until it stops paying the supplier, while pricing tied to a person or an outcome survives, because it cannot be engineered away. AI is arriving at the same conclusion, slowly, and its customers will get there first.
There is one number worth having before the next renewal, and neither vendor publishes it: what one finished piece of work costs on your own workload, measured twice. Until a company can answer that, an AI discount is a claim about someone else's benchmark, and it has never appeared on anybody's invoice.
Editor's note
Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.