Why Anthropic Held Its List Price While a Rival Ran the Same Task for 3 Cents

A Hundredfold Gap, in One Line

One comparison published by ZDNet Korea on August 4 sets the temperature of this market. Running a single complex real-world task cost $0.03 on DeepSeek V4-Flash and $3.15 on Anthropic’s Claude Fable 5, a gap of more than 100 times (ZDNet Korea). That gap opened after eight weeks in which Alibaba’s Qwen 3.8-Max, Moonshot AI’s Kimi K3, Zhipu AI’s GLM-5.2 and DeepSeek’s V4-Flash all shipped.

What buyers actually pay fell with it. On the Silicon Data index cited by Jefferies, the blended cost per million tokens went from $2.04 on May 31 to $1.45 in late July and $1.16 to $1.18 between August 6 and 8, the lowest reading of the year (SCMP).

The two American labs answered differently. OpenAI touched the list price and cut its GPT-5.6 series by as much as 80% (SCMP). Anthropic left its list price alone.

The Line Anthropic Chose to Defend

Opus 5, released in July, is priced at $5 per million input tokens and $25 per million output tokens, the same numbers its predecessor Opus 4.8 carried (ZDNet Korea). The claim the company put forward instead was cost per task: comparable or better results than the top-end Fable 5 at roughly half the operating cost.

Spend fewer tokens to finish the same job and the customer’s invoice drops while the rate card does not move. That splits an effective discount from a list discount, and only one of the two is easy to take back. A published price cut has to be explained to every customer on the day you reverse it. Cost per task gets re-set quietly with the next model.

The arithmetic explains the ordering. Anthropic closed its Series H in May at a $965 billion post-money valuation on $47 billion of run-rate revenue as of mid-May (Anthropic). Take 10% off list and $4.7 billion comes off that run rate on contact. Volume can refill it, but that is an awkward bet to place in the weeks when Morgan Stanley, Goldman Sachs and JPMorgan are reportedly lining up investor meetings for a listing as early as October (ZDNet Korea).

The Discounter Raised Prices First

On August 6 DeepSeek told API users that overall pricing was going up, warning of a “significant increase” and telling developers to plan their usage accordingly (SCMP).

Zhipu AI moved the same way. GLM-5 launched in February at $1 per million input tokens and $3.20 output. GLM-5.2 lists at $1.40 input and $4.40 output (ZDNet Korea). The version number went up and so did the bill.

Demand that concentrates eats capacity first. DeepSeek’s share of tokens processed on OpenRouter reached 27% on August 10, passing Google’s 25% (ZDNet Korea). Winning that share is what put the capacity problem on its desk. A team that has moved its entire workload onto one ultra-cheap provider receives the price notice as an announcement rather than a negotiation.

What Goes on Your Own Price Sheet

Keep list price and cost-to-serve in separate columns. When a competitor names half your price, the second column is the one you can cut, and delivering the same outcome with less consumption is far easier to walk back than a rate card.

Reprice the option value of a second provider. The cost of maintaining one extra routing path against the cost of being unable to move in a day: the DeepSeek notice just published that quote for you.

Then read the change-of-price clause in your vendor contract. How many days of notice, and whether existing rates survive any grandfathering window. For anyone running on discount inference, those two lines now matter more than the per-token number above them.

When Anthropic files, the prospectus will carry one line of gross margin. What that line reads is the scorecard for meeting a price war without moving your price.