Anthropic vs OpenAI
Claude Sonnet 5 vs GPT-5.6 Terra pricing
Identical headline input price. The difference shows up in output rates, in how each provider charges for long context, and in which one can go stale on you.
| Per million tokens | Claude Sonnet 5 | GPT-5.6 Terra Standard, short context |
|---|---|---|
| Input | $2 | $2 |
| Cached input read | $0.20 | $0.20 |
| Output | $10 | $12 |
| Batch input | $1 | n/a |
| Batch output | $5 | n/a |
Reading the table
Both models list the same input rate, $2 per million tokens, which is why a comparison that stops at the headline number concludes there is nothing to choose between them. The separation is on output, where Claude Sonnet 5 is $10 and GPT-5.6 Terra is $12. On a typical production workload the response drives most of the bill, so a difference on the output line matters far more than the same difference on input.
The two-band clause
Claude Sonnet 5 charges one rate whatever the request size, so a very large request bills at the same per-token rate as a small one. GPT-5.6 Terra is priced in two bands: input moves from $2 to $4 per million tokens and output from $12 to $18 once a request crosses into the higher band. Two teams on the same nominal rate card can therefore pay very different effective rates depending on the shape of their traffic, and that clause is invisible on the flat tables most comparison articles copy.
Where the rate card stops being the answer
Claude Sonnet 5 uses a newer tokenizer that produces roughly 30% more tokens for the same text than Claude Sonnet 4.6. Moving up a version at the same headline rate is not a flat trade.
Cached input is charged at 10% of the standard input rate in both bands.
None of this is knowable from a rate card, and none of it is the reason most teams lose money on model spend. The bigger problem is attribution: a provider invoice arrives as one number for the whole organization, so nobody can say which customer, team, or agent caused the increase. That question does not get easier by picking the cheaper model.
Frequently asked questions
Is Claude Sonnet 5 cheaper than GPT-5.6 Terra?
On input they are identical at $2 per million tokens. On output, Claude Sonnet 5 is cheaper at $10 against $12. Because output typically drives the majority of a production bill, that is usually the number that decides it.
Does the headline rate hold for every request?
Not on both sides. Claude Sonnet 5 charges one rate whatever the request size. GPT-5.6 Terra is priced in bands, so input runs from $2 to $4 per million tokens and output from $12 to $18 depending on the request. Two teams reading the same rate card can end up paying very different effective rates, which is why a comparison built on headline numbers alone tends to flatter the banded model.
Does switching from GPT-5.6 Terra to Claude Sonnet 5 guarantee a lower bill?
No. Headline rates only decide the bill if token volume stays constant, and it rarely does. Models tokenize the same text differently, reasoning effort changes output volume without changing a prompt, and retries and agent fan-out multiply calls. The only way to know is to price your own traffic against both rate cards using your real input and output token mix, not a vendor's example.
How do I compare these on my own usage rather than on list price?
Take a representative month of calls, split the tokens into input and output, and apply each rate card to your actual mix. If you cannot split them, that is the first problem to fix, because a provider invoice is a single number for the whole organization. Spendline records every call in a ledger with the model, the resolved cost, and the customer or agent it belongs to, which makes this a query rather than a spreadsheet exercise.
Know what this costs per customer, not just per token
Choosing the cheaper rate card saves a percentage. Knowing which customer, agent, and feature spent the money is what turns model spend into a cost line you can manage, and it is the question neither provider’s invoice answers.
Run the 5 minute assessment