Claude Opus 5 vs GPT-5.6 Sol: which is cheaper depends on a threshold OpenAI does not publish

This page was rewritten on August 22 because its premise stopped being true. Until August 21 both models charged five dollars per million input tokens, and every comparison came down to the output column. OpenAI then cut Sol by 20% on input and 33% on output, and the shape of the question changed with it. What follows is the comparison at today's published terms.

Short context: Sol is now cheaper on both columns

Sol charges $4 per million input tokens against Opus 5's $5, and $20 output against $25. Cache reads are $0.40 against $0.50. Batch halves both sides at both vendors: $2 and $10 on Sol, $2.50 and $12.50 on Opus 5.

Every one of those columns favors Sol, so within short context there is no crossover point to find and no mix of input to output that changes the answer. Price a job of 50,000 input and 5,000 output tokens and it costs $0.30 on Sol against $0.375 on Opus 5 — a fifth less, consistently.

Long context: Opus 5 is cheaper, and it is not close

Anthropic includes the full million-token window at standard rates and states it plainly: a 900,000-token request is billed per token exactly like a 9,000-token one. Caching and batch discounts apply across the whole window at the same rates.

OpenAI runs a second column. Sol's long-context rates are $8 input, $0.80 cached and $30 output — double the input, half again on the output. Push the same large job through both and the gap is wide: at 400,000 input and 20,000 output tokens, Opus 5 costs $2.50 and Sol costs $3.80, which is 52% more.

Here too there is no crossover. Both of Sol's long-context coefficients are higher than Opus 5's, so once you are in that column Opus 5 wins on every mix.

Opus 5 sits between Sol's two columns, and that is the whole comparison

Take one job — 50,000 tokens in, 5,000 out — and price it three ways. On Sol's short-context column it costs $0.30. On Opus 5 it costs $0.375. On Sol's long-context column it costs $0.55. Opus 5 is 25% more expensive than one Sol and 32% cheaper than the other, for the identical request.

So the question is not which model is cheaper. It is which of Sol's two columns your traffic will be billed under, and that is the one number OpenAI's pricing page does not carry: it publishes both columns and no threshold. You can find the answer on your own invoice at the end of the month, which is a strange place to discover the price of something.

Anthropic has no such column, and that is the actual difference between these two price lists. Not the rates — those move — but the fact that one of them can be computed in advance and the other cannot.

One price has an expiry date, the other does not

The $4 and $20 are promotional. OpenAI announced them on August 21 as a reduction of over 20% for the next three months, and the page says the promotional pricing is available at least through November 21, 2026. The floor is stated; the ceiling is not.

Opus 5's $5 and $25 carry no date. That does not make them permanent, but it does mean the two figures you are comparing are not the same kind of number: one is a standard rate and the other is an offer with a stated end. If your decision commits you past November, run the sum twice — once at $4 and $20, and once at $5 and $30, which is what Sol cost until last week.

At the old rates, on 1,000 calls of 4,000 input and 800 output tokens, Sol cost $44 and Opus 5 $40. Today the same thousand calls are $32 on Sol and still $40 on Opus 5. Nothing about either model changed.

Multipliers on both sides

Anthropic sells fast mode for Opus 5 at $10 and $50 per million, double the standard rate, and it stacks on top of caching and inference geography. Pinning inference to the United States multiplies every token category by 1.1 on Claude 4.6 and later.

OpenAI charges 10% more for regional processing on models released on or after March 5, 2026. Both vendors will charge you extra for speed or for where the computer physically sits, and neither charges it until you ask.

One thing no price list can tell you

Claude 4.7 and later use a newer tokenizer, which Anthropic says produces about 30% more tokens for the same text than the previous one. Opus 5 is on that tokenizer.

That figure describes Anthropic against Anthropic. It says nothing about how either model counts against OpenAI's tokenizer, and anyone claiming otherwise is calculating where they should be measuring. If the choice is close on price, count tokens for both on a real sample of your own text. It is an afternoon's work and the only number that describes your bill.