Anthropic Cut Task Costs Up to 45% and Says Margins Held

Anthropic says it cut what customers pay per task by up to 45% with its September 2026 models while keeping its margins, in the Form S-1 it filed with the SEC on 26 October 2026. “Because of our compute efficiency, we have been able to make these improvements while preserving our margins”, the filing says.

The S-1 says Claude Fable 5.1 reduced the cost of typical workloads by approximately 25% compared with Fable 5, with savings of up to 45% for complex coding and highly agentic tasks, and that Opus 5.5 reduced customers’ cost per task by approximately 40% compared with Opus 5.

What the Filing Says Made Tasks Cheaper

The filing names two sources of the savings: “Customers’ cost per task can decrease for two reasons: newer, more capable model versions can complete tasks more efficiently, and we may reduce the prices we charge for our models.”

It says each September cut was “a combination of a lower price and improved efficiency, the contributions of which depend on the task”. It does not split any of the three figures between the two.

Source: Anthropic, PBC Form S-1, filed with the SEC on 26 October 2026, Business section. Both models were released in September 2026, the filing says.

What Happened to Revenue After the Last Cut

The S-1 gives one earlier case. It says Opus 4.5, launched in November 2025, came with “greater capabilities and a lower price and cost per task than its predecessor Opus 4.1”. After that launch, the filing says, “our total quarter-over-quarter revenue growth accelerated from 48% for the three months ended December 31, 2025 to 132% for the three months ended March 31, 2026”.

The filing does not credit price alone. It says that growth “reflected the combined effect of several factors”, naming lower prices and improved capability that unlocked new use cases, the emergence of new products, and expanded customer adoption.

Source: Anthropic, PBC Form S-1, filed with the SEC on 26 October 2026, Quarterly Results of Operations (US$ millions, unaudited). Growth on the prior quarter is our arithmetic, except 48% and 132%, which the S-1 states.

Tokens Are the Meter

Most of the revenue is metered. The S-1 says pay-for-intelligence “currently represents the substantial majority of our total revenue”, and that “Under this model, customers pay for usage, which is metered in tokens and priced according to the model selected.”

That is why the price of a task matters to the top line: a cheaper task lowers what each piece of work earns, and the filing’s Opus 4.5 example is its own account of customers then using more.

Connecting the Coverage

Our compute piece reported compute and infrastructure expense falling from 251% of revenue in the first quarter of 2025 to 48% in the second quarter of 2026. Our S-1 piece covered the headline revenue. Our Claude Code piece and our revenue-per-employee piece covered how the filing says Anthropic runs on its own model. What is new here is the filing’s account of price: cheaper tasks, margins it says it preserved, and the revenue growth that followed the last cut.

The Structural Read

As we read it, the filing describes a pattern rather than a promise. After the last lower-priced launch, Opus 4.5, quarterly revenue growth rose from 48% to 132%, and the S-1 lists lower prices among several causes.

The same filing says it expects its revenue growth rate to decline over time, and it names agentic work, which is where its biggest September savings sit, as structurally more compute-intensive.

As we read it, the test is whether customers keep using more Claude for each dollar the price falls. The S-1 gives one quarter of evidence for that, and no per-token figures to check it against.

Three Implications

FOR BUYERS OF AI The S-1 says a cheaper task comes from both lower prices and more efficient models, and that the split depends on the task.

FOR INVESTORS The filing pairs price cuts with a statement that margins were preserved, but the passage gives no margin figure for those models.

FOR COMPETITORS The S-1 says open-weight models could exert pricing pressure, and that it may have to reduce prices or offer task-based pricing.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The guide quotes Zenskar: “The world moved from a per-query model to a per-workflow model — where one action can trigger dozens of model calls — so lower cost per token multiplied by exploding token volume increases total spend.”

As we read it, the S-1’s Opus 4.5 example is a seller-side view of the same mechanism: a lower cost per task, followed by customers using more. The guide was published in June 2026 and does not discuss this filing.

What Is Not Established

The S-1 gives no price per token and no count of tokens processed. The passage does not quantify the margins it says were preserved, and the filing does not say how much of any cut came from price.

Independent measurement can differ. Our 2 September piece reported Artificial Analysis’s first evaluation of Fable 5.1, which on that firm’s own workload put cost per task about 20% above Fable 5, driven by about 1.7 times the output tokens. The S-1’s figure refers to typical workloads.

The filing also lists the opposite risk. It says “our agentic offerings are structurally more compute-intensive than our non-agentic offerings”, and that without pricing that reflects this, “our compute efficiency could deteriorate as agentic usage grows”.

It says it “may be required to reduce our prices”, that consumption revenue may fall if customers “shift usage to lower-priced models”, and: “We expect our revenue growth rate to decline over time and to fluctuate from period to period.”

The Bottom Line

Anthropic’s S-1 says its September 2026 models cut customers’ cost per task by approximately 25% to 40%, and by up to 45% for complex coding and agentic work, while preserving its margins. It says that after the lower-priced Opus 4.5 launched in November 2025, quarterly revenue growth rose from 48% to 132%, which it attributes to several factors including lower prices.