Cheaper AI Tokens to Put Anthropic’s Economics to the Test 

AI is getting cheaper per token, and businesses are spending more on it anyway. Providers keep cutting the price of many tokens, but customers are feeding models longer documents, pulling more output, and running agents that can call a model several times to finish a single task.  

Anthropic’s current price list runs from $0.10 per million input tokens for Claude Haiku 5.5 (on prompts up to 100,000 tokens) to $10 for Claude Fable 5.1. Output runs from $0.50 to $50 per million tokens across those models.^¹ As unit prices fall, the question for the market is who keeps the savings?  

What You’re Actually Paying For  

Language models read and write in tokens, and that’s how they bill, not by the document, the question, or the finished task. Input tokens cover instructions, source material, and conversation history. Anthropic’s token-counting tool estimates input usage before a request goes out, so developers can manage cost, rate limits, routing, and prompt size.^²  

Output tokens are what the model writes back, and they usually cost more because the model generates them one at a time. Claude Sonnet 5.5, for example, lists at $2 per million input tokens and $10 per million output tokens.^¹ That makes the shape of the workload matter. Document analysis can mean a big input and a short answer. An agent can generate more output and call the model again and again.  

The Sticker Price Is Only the Start  

How a workload is built changes what it really costs. Repeated instructions, documents, and conversation history can be cached and reused instead of being billed at the full input price on every call. Anthropic lists cache hits for Claude Sonnet 5.5 at $0.10 per million tokens and says caching cuts both cost and processing time on repetitive prompts.^³  

Model routing, context length, retries, and the number of task steps all move the final bill, too. Anthropic also notes that newer tokenizers can turn the same text into more tokens, so a lower quoted price doesn’t always mean a proportionally lower workload cost.^¹ The number that counts is the cost of the finished job.  

Better Models, Smaller Price Tags  

In February 2026, Anthropic priced Claude Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens.^⁴ Claude Sonnet 5 arrived in June at $2 and $10, and Anthropic called it a substantial improvement over Sonnet 4.6, close to Opus 4.8 on some tasks.^⁵ The model, tokenizer, and capabilities all changed, so it isn’t a perfect apples-to-apples comparison. But the direction is clear: providers are competing on capability per dollar.  

Epoch AI tracked the lowest inference prices needed to hit fixed performance levels and found annual declines ranging from ninefold to 900-fold, depending on the benchmark. In one series, the price of reaching GPT-3.5-level performance on a general-knowledge benchmark dropped from $20 per million tokens in November 2022 to $0.07 in October 2024.^⁶ Work that once needed a premium model can now run on a cheaper one.  

The Bull Case: Lower Prices Unlock More Work  

Cheaper tokens can make AI pencil out for more customer-service interactions, longer document reviews, lower-priced software plans, and multi-step agents. Companies can send routine tasks to lower-cost models and save premium systems for the harder work. When applications that didn’t make economic sense suddenly do, the market often gets bigger.  

For Anthropic, the bull case comes down to elasticity. Token volume, customer adoption, premium workloads, and lower serving costs need to grow fast enough to make up for lower revenue per unit. Sonnet 5 illustrates the approach: a lower price than Sonnet 4.6, better capability, and a wider menu of cost-performance options.^⁵ Cheaper access can widen distribution and embed Anthropic more deeply in the software customers already use.  

The Bear Case: Volume Has to Outrun Compression  

Lower prices also make it easier to shop around. Customers can route workloads across providers, push routine work to smaller models, use open-weight alternatives, or save frontier systems for a narrower set of tasks. That can squeeze revenue per token and raise the stakes on differentiation.  

Costs can also be high. Developers have to fund training, inference, research, safety work, and infrastructure access. More tokens don’t guarantee enough gross profit if price competition heats up or compute costs fall too slowly. The bear setup is that usage grows, just not fast enough to offset price compression and the resources it takes to serve all that usage.  

Who Wins When AI Gets Cheaper?  

Enterprise customers come out ahead first, since cheaper pricing lowers the bar for deployment. Model developers can win back part of that value through more volume, premium capabilities, cached workflows, and deeper integrations. Cloud providers and chipmakers can benefit if total inference demand grows, though newer hardware and smarter algorithms can also shrink the compute each task needs.  

For Anthropic, traders will be watching token-volume growth; the mix across Haiku, Sonnet, Opus, and premium models; input versus output usage; cache adoption; enterprise pricing; and compute cost per token. Together, those signals can show whether cheaper prices are growing the market or shifting value elsewhere.  

Trading Anthropic’s Token Economics  

The Direxion Daily Anthropic Bull 2X ETF (Ticker: ANTO) and Direxion Daily Anthropic Bear 2X ETF (Ticker: ANTD) are expected to begin trading on the second trading day after Anthropic’s Class A common stock begins trading. ANTO will seek daily investment results, before fees and expenses, of 200% of the daily performance of Anthropic Class A common stock. ANTD will seek daily investment results, before fees and expenses, of 200% of the inverse, or -200%, of that daily performance. Holding the funds for longer than a single day can produce returns materially different from 200% or -200% of the stock’s performance over that period. 

The funds will seek leveraged or inverse results for a single trading day, and will track one company’s stock rather than a diversified index, which can increase the effects of company-specific volatility. 

Token prices can keep falling while the token market keeps growing. The dividing line is whether more usage, premium workloads, and lower serving costs can outrun the drop in revenue per unit. Cheaper intelligence opens up the opportunity. It also forces every provider to prove where its pricing power comes from.