🔥 Trending on HN

AI Prices Are Falling Fast. The Real Battle Is Cost per Task

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
cached input

Previously read information that a model reuses.

token

A small unit used to count text for AI billing.

open-weight model

A model whose key learned values are publicly released.

What happened

A Max Woolf analysis tracks major AI price cuts from late June through late September 2026. The companies are OpenAI, an AI company, and Anthropic, the company behind Claude. The analysis compares three charges: input tokens, output tokens, and cached input.

Its timeline shows a sharp example. GPT-5.6 Luna fell from $1 to $0.20 per million input tokens. Output fell from $6 to $1.20. Cached input fell from $0.10 to $0.02. Each change equals an 80% cut. OpenAI's July 30 announcement confirms the lower Luna prices. Anthropic also made Claude Sonnet 5's $2 input and $10 output launch prices permanent. It later said Claude Opus 5.5 costs 40% less on typical workloads than Opus 5.

The Hacker News submission received 213 points and 240 comments. That number shows community attention. It does not prove the analysis is correct.

Why the prices are falling

Woolf's main explanation is competitive pressure. Many models can now handle similar jobs. Users can switch models with little effort. Open-weight models add more substitutes. Multiple inference providers can also host one popular model. These conditions make it difficult for one company to keep a large price advantage.

The pattern looks like a back-and-forth contest between OpenAI and Anthropic. That is an interpretation of the timeline, however. The companies have not jointly confirmed that competition caused every cut.

Why cached input matters

The headline prices do not tell the whole story. Coding agents often reuse a long conversation. The model may read the same project context many times. Cached input charges cover that reused context.

Woolf writes that cached reads can make up 96% to 99% of token spending in agentic coding. He also describes cache-hit rates from 90% to 100%. Those figures are not universal measurements for every service. They explain why a cache discount can matter more than a small output discount.

Lower prices can create another effect. A cheaper task may encourage more tasks. A developer may test more models, run longer agents, or automate more work. So the price per token can fall while total spending rises. The analysis points to Vercel data showing token costs at 43.3% of their April level, while aggregate spending increased. That chart is not proof of an industry-wide rule.

What is confirmed

Official pages confirm several price changes. OpenAI's GPT-6 price table lists lower input and output prices for GPT-6 Sol than GPT-5.6 Sol. Anthropic says Sonnet 5's introductory price became permanent in its launch update. Its Opus 5.5 announcement lists 20% lower input and output prices, plus a 60% cache-read cut.

What remains uncertain

A lower list price is not the same as a lower cost per completed job. Results depend on model choice, answer length, cache hits, speed, limits, and retries. Public announcements also do not show each company's margins. We cannot yet separate the effects of competition, better hardware, improved software, and better model design.

What to watch next

The useful comparison is cost per finished task. Watch whether prices stay low, whether quality keeps improving, and whether open-weight models catch up. Also watch total usage. If cheaper AI causes much more work to move through these systems, lower prices may still produce larger businesses.

💬 Why Are LLMs Getting So Much Cheaper?

On the article’s pricing question, Hacker News commenters pointed to efficiency gains and competition, while debating sustainable profits and possible pre-IPO strategy.[Source article](https://minimaxir.com/2026/10/cheap-llms/)

  • Some users reported that Kimi and GLM were cheaper than OpenAI and Anthropic for their work, suggesting that more substitutes are putting pressure on prices.
  • Another user reported that Kimi was more expensive in their setup, both by token use and by the total cost of completing work. Price comparisons depend on the model, workload, and harness, and are not controlled measurements.
  • A technical explanation is that caching, smaller models, better hardware, and improved inference methods reduce the computing needed for the same task.
  • A user-reported estimate claimed that throughput and efficiency could improve by roughly 10x per year, but the thread did not establish that as a verified measurement.
  • A strong counterargument is that serving models still consumes electricity and memory as usage rises, making a truly unlimited flat-rate plan harder than a mobile-data plan.
  • Covering inference costs is not the same as covering training, R&D, and future model-development costs.
  • Some commenters speculated that expensive enterprise plans and investor funding may support cheap consumer plans, or that pre-IPO competition may be encouraging OpenAI and Anthropic to retain users. These are hypotheses, not established facts.

growing digest at 405 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Why AI Became Much Cheaper

📰 Full story: AI Prices Are Falling Fast. The Real Battle Is Cost per Task

OpenAI and Anthropic cut prices because AI companies are competing harder.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
cached input

Old information kept for reuse.

API

A way for a program to ask a model for work.

💡 The gist

  • OpenAI, an AI company, and Anthropic, the company behind Claude, cut prices.
  • GPT-5.6 Luna's input and output prices fell by 80%.
  • Hacker News discussed it widely. Attention does not prove truth.

AI companies charge for several kinds of work. Input is what the model reads. Output is what the model writes. Cached input is old information the model reads again.

A Max Woolf analysis compares prices from June through September 2026. GPT-5.6 Luna's input price fell from $1 to $0.20 per million tokens. Its output price fell from $6 to $1.20. Both new prices equal 20% of the old prices.

These numbers describe API use. An API lets a program ask a model for an answer. They do not automatically describe every monthly subscription. Some products use quotas or bundles instead. Users should compare the final bill, not only the model's price table.

OpenAI's official announcement lists the lower Luna prices. Anthropic also kept Claude Sonnet 5's launch price. Later, Anthropic said Opus 5.5 costs less on typical work.

Why would companies lower prices? More AI models can do similar jobs now. Customers can try another model more easily. Open models add more choices. Companies must compete for users.

Cached input makes this more important. Coding tools often reuse the same long project information. A cache keeps information that the model already read. Reusing it can cost less than reading everything again. This matters when one tool works through many steps.

A lower price can also increase usage. Developers may test more models. They may run agents for longer. They may automate more tasks. Therefore, each task can become cheaper while total spending grows.

The Hacker News post received 213 points and 240 comments. Those numbers show interest from that community. They do not prove every claim in the analysis.

One question remains. Does a lower token price mean a lower price for finished work? Not always. The answer depends on retries, speed, limits, and model quality. Quality also matters when comparing two prices. The best number to watch is the cost of completing one useful task.

💬 Why are AI models getting cheaper?

Commenters gave two main reasons: running the technology may be cheaper, and more companies are competing. They also warned that low prices do not prove lasting profits.[Source article](https://minimaxir.com/2026/10/cheap-llms/)

  • Some users reported that Kimi and GLM cost less than OpenAI or Anthropic for their work. More competition can push prices down.
  • Another user reported that Kimi actually cost more for their work. The result depends on the model, the task, and the software coordinating the AI.
  • Smaller models, caching, better hardware, and better methods may reduce the computing needed. A user-reported estimate said efficiency could improve by about 10x a year, but this was not a verified measurement.
  • AI still needs electricity and memory whenever people use it, so a truly unlimited plan is difficult.
  • Paying the cost of running a model does not necessarily pay for training and building the next model.
  • Commenters also suggested that business customers, investors, or a pre-IPO push might help support cheap personal plans. Those ideas remain speculative.

growing digest at 405 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

AI Work Costs Less Money

📰 Full story: AI Prices Are Falling Fast. The Real Battle Is Cost per Task

OpenAI and Anthropic made AI cheaper to use.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Hacker News

A website where people discuss technology news.

OpenAI (an AI company) and Anthropic (the company behind Claude) lowered prices.

AI reads words. AI writes words. Both jobs can cost money.

Sometimes AI reads the same old words again. A cache keeps those words. Then AI does not read everything again. That can cost less.

Many AI companies can do similar jobs. They want people to choose their AI. So they compete by lowering prices.

A Hacker News post got 213 points and 240 comments. That shows interest, not proof.

Cheaper AI may make people use it more. One job can cost less. All jobs together may still cost more.

Max Woolf's article says this price fight may continue. It is an analysis, not a promise.

💬 Why does AI cost less now?

AI companies are charging less money. People in the discussion gave a few reasons.

  • Companies compete, so they lower their prices.
  • New machines, smaller AIs, and clever tricks can make the same job use less computing. A user-reported estimate said the improvement might be about 10x each year.
  • But AI still needs electricity and memory every time it works. Paying for one job does not pay for building the whole AI.
  • One user reported that a supposedly cheap AI cost more for their own work. Cheap plans might also be helped by investors or business customers.

growing digest at 405 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

Sources