Anthropic Froze Sonnet 5's Price. Your Actual Bill Did Not Get the Memo.

Anthropic announced today that Claude Sonnet 5's list price is not going up. The company said on X that the introductory rate of $2 per million input tokens and $10 per million output tokens, in place since the model's June launch, is now permanent. The standard rate of $3 and $15 that was scheduled to kick in on September 1, a straight 50 percent increase on both numbers, is gone. BigGo Finance reported the move lands against a backdrop of Anthropic's confidential IPO filing this June, at a reported $965 billion valuation, with a public listing expected this fall.

My position: this is the free market doing exactly what it is supposed to do, and it is worth crediting when it happens. Nobody regulated Anthropic into holding its price. OpenAI cut its GPT-5.6 Luna model by 80 percent to $0.20 per million input tokens on July 30, and open-weight models like DeepSeek V4 Flash are pricing flagship-class capability under $0.15 per million tokens direct from the vendor, cheaper still through marketplaces. Anthropic had a price hike planned and shelved it two weeks before it would have taken effect. That is not generosity. That is a company checking what the field is charging and blinking.

Here is the steelman for the other read, because it deserves one: maybe this has nothing to do with competition and everything to do with the IPO. A company about to go public wants clean growth numbers and happy enterprise customers locked into predictable rates, not a customer revolt over a price hike landing right before the roadshow. Under that reading, the freeze is optics timed to a specific event, not a durable market signal, and it could reverse the day after the listing closes. That is a fair read. I do not think it is the whole story, because the competitive pricing pressure from OpenAI and the open-weight labs was already public and already real before the IPO filing existed. But a freeze announced this close to a listing earns skepticism, not applause, and I would rather say that than pretend the timing is a coincidence.

Now the part that actually matters more than the announcement. The Decoder published an analysis arguing that Anthropic has a pattern of holding the sticker price flat while the real cost per task climbs anyway. Its numbers: the average Sonnet 5 task now runs about $2.29, up from roughly $1.20 on Sonnet 4.6, even though the per-token rate has not moved. The reason is that Sonnet 5 burns close to 40 percent more output tokens per task and runs about three times as many agent loops as its predecessor. Do the math and the per-token price is a red herring. A model can hold its list price forever and still cost you twice as much a month later, because it is doing more work, or busier work, to get to the same answer. The Decoder also flags that Sonnet 5's average task now costs more than Opus 4.8's, $2.29 versus $1.97, despite Opus listing at a much higher per-token rate of $5 and $25. The cheaper-looking model is the more expensive one in practice.

I do not think this is necessarily deceptive on Anthropic's part. A more capable model that reasons longer and takes more agent steps will burn more tokens by design, that is closer to a mechanical result of the architecture than a pricing trick. But it means the freeze announced today answers a question nobody who runs a real workload was actually asking. The question is never "what does Anthropic charge per million tokens." It is "what did my invoice say last month, and why."

So the practical takeaway, for anyone whose business runs meaningful volume through Claude, GPT, or anything else billed by the token: stop tracking vendor list prices as a proxy for your AI budget. Track dollars per completed task, measured against your own workloads, on a schedule, because token consumption per task moves independently of the sticker price and nobody is going to announce that on X for you. This is the same discipline behind model routing, sending the routine 90 percent of requests to a cheap or open model and reserving the expensive reasoning for the hard 10 percent, except now it applies inside a single vendor's own model line, not just across vendors. A frozen price and a rising bill can both be true at once, and only one of them shows up in a press release.

If you want a straight read on what your actual per-task cost looks like instead of what a pricing page says, tell me what you are running.