OpenAI and Anthropic’s Price Cuts: What Repeated Prompts Save

OpenAI and Anthropic have made repeated conversation text much cheaper to process, changing what developers pay for AI agents that work through long tasks. An October 25 analysis by data scientist Max Woolf traced price reductions as large as 80%. The savings on a particular job depend on which text the service processes again and how much new text its model generates. Woolf

The largest discounts span several months and different models. OpenAI’s official changelog dates its 80% GPT-5.6 Luna reduction to July 30. It released GPT-6 Sol and GPT-6 Luna on September 22, followed by GPT-6.1 Sol on September 29. A comparison using those names needs to distinguish a price change on an existing model from a move to another model. Changelog

GPT-5.6 Luna currently costs $0.20 per million input tokens and $1.20 per million output tokens, with reused input at $0.02. Tokens are the small units into which an AI service divides text. GPT-6 Luna’s separate published standard rates are $0.10 for input, $0.50 for output and $0.01 for reused input. Both remain listed; their similar names do not make them the same product. Luna Prices

Opus cuts reused input more than new text

Anthropic’s September 22 Opus 5.5 launch illustrates why one discount cannot describe every customer’s bill. The company prices fresh input at $4 per million tokens and output at $20, each 20% below Opus 5. Reusing previously processed input costs $0.20 per million, down from $0.50. That last reduction is 60%. Anthropic

Anthropic estimated that Opus 5.5 would cost 40% less on typical workloads at default settings, reflecting both cheaper tokens and fewer tokens used per task. It also said the model requires less computing capacity to serve. Those are the developer’s own findings, covering its evaluated work. A customer who gets longer answers or runs more attempts can end up with a different saving. Anthropic

A calculation using a fixed workload separates those effects. Assume ten million billable tokens across many short-context requests: 9.8 million reused input tokens, 100,000 fresh input tokens and 100,000 output tokens. This is an illustrative mix, not a measured customer session. It excludes initial cache writes, tool fees and other charges so the comparison covers the same three token categories. Opus

At Opus 5 rates, that mix costs $4.90 for reused text, $0.50 for fresh text and $2.50 for output, totaling $7.90. At Opus 5.5 rates, the components are $1.96, $0.40 and $2, totaling $4.36. The reduction is about 44.8%. Even with 98% of tokens coming from saved text, the 60% cache discount does not produce a 60% reduction in these model charges: expensive output still contributes to the total. Anthropic

A 75% cache discount changes with the workload

OpenAI’s current standard short-context rates for GPT-5.6 Sol are $4 per million fresh input tokens, $20 for output and $0.40 for reused input. GPT-6.1 Sol costs $2, $10 and $0.10 respectively. Prices

Apply the same ten-million-token mix and GPT-5.6 Sol’s token charges total $6.32. GPT-6.1 Sol’s total is $2.18, a reduction of about 65.5%. A workload containing only fresh input and output would save 50% at those rates. The precise 75% reduction applies to the reused-input category. These calculations hold token quantities constant; they make no claim that both models need the same number of attempts to finish a real task. Prices

That distinction is central to Woolf’s argument. His analysis emphasizes cache-heavy coding agents and suggests that competition from open-weight models and alternative hosting companies helps constrain prices. Those rivals allow developers to buy inference, the computing used to produce answers, from several suppliers. The explanation is plausible market analysis, rather than a statement from OpenAI or Anthropic identifying why they chose each price. Woolf

Reusing a prompt has conditions and an initial cost

A prompt cache saves processing already performed on the unchanged beginning of a request. An agent can reuse instructions, documents and earlier conversation turns when that beginning matches a saved entry. OpenAI’s guide says keeping a session open does not guarantee a match. Changing earlier content can prevent later text from qualifying for reuse, even when much of that text looks familiar to a person reading it. Caching

Creating the saved entry also has a price. For supported current OpenAI models, a default 30-minute cache write costs 1.25 times ordinary input, while a 12-hour write costs 2.5 times. These are alternative rates for those tokens, rather than extra charges added to ordinary input. Reading the resulting entry is cheaper. A frequently rebuilt prompt can therefore have a different bill from a stable prompt reused many times. Caching

Anthropic has a different timing choice: five-minute writes cost 1.25 times base input and one-hour writes cost twice base input. A cache hit refreshes the default five-minute lifetime without another write charge. Its usage report separates newly written cache content, content read from cache and ordinary input. Counting only the ordinary-input field would miss text processed through the other two categories. Caching

Model behavior changes the output side as well. OpenAI bills reasoning tokens, the model’s internal problem-solving steps, at output rates even though developers do not see that text. Its documented output limit covers reasoning and visible output together. A short final answer can consequently accompany substantial generated-token usage. The quantities recorded by the service provide a firmer billing basis than counting words on screen. Reasoning

Anthropic’s Opus 5.5 documentation identifies a related migration issue: its adaptive thinking is always on, with an effort setting controlling depth. Applications that previously disabled thinking on Opus 5 need to change that configuration. Comparing the two models at their defaults therefore involves both prices and behavior, rather than simply applying a new price to an unchanged program. Opus

Broader usage data shows customers expanding activity as prices fall. In its October 15 production report, Vercel said September spending through its AI Gateway was about 13 times December’s while average price per token was 47% lower. The median paid customer active in both months more than doubled spending. This describes Vercel’s routed traffic, rather than all AI customers or either lab’s total revenue. Vercel

Vercel estimates spending at published list prices, and its token totals include input, output, reasoning, cache reads and cache creation. Those choices matter when interpreting an average: a shift toward cheaper models or repeated inputs can lower the aggregate price even while individual customers buy more service. Its report documents the usage mix and increased spending; it cannot establish which price change caused a particular team to expand. Vercel

One dated condition remains relevant for buyers comparing today’s models. OpenAI lists GPT-5.6 Sol’s promotional pricing as available at least through November 21, 2026. The continuation or replacement of that offer would change a comparison using it as the baseline. Customers reviewing the newer models can already check their recorded input, cache-write, cache-read and output quantities against the published rates. Changelog