Cut Your AI Bill by Up to 80%: A Price-War Playbook

The AI price war is the best deal you will get this year, and most people have not touched their settings. Frontier labs are cutting prices against each other, cheap open models keep pushing from below, and switching between them is now one line of config. Here is the playbook for turning that into a smaller bill.

What happened

On October 25, developer Max Woolf published an analysis of LLM pricing. His main point: OpenAI and Anthropic cut prices by up to 80% within a few months, one after the other, under pressure from cheap open-weight models and secondary inference providers. Models are close to commodities, because tools like OpenRouter make switching easy, and that limits how much anyone can charge.

One detail matters more than the headline. In agentic coding, cache reads make up 96 to 99% of the token spend. So the cache-read price, not the headline price, decides your real bill.

Here are the prices per million tokens, input and output (US dollars):

• GPT-6 Luna: $0.10 in, $0.50 out, cache read $0.01
• DeepSeek V4.1 Flash: $0.15 in, $0.60 out on DeepSeek’s own API. Other hosts differ: OpenRouter listed $0.02 in and $1.00 out when I checked
• GPT-6.1 Sol: $2 in, $10 out, cache read $0.10
• Claude Sonnet 5: $2 in, $10 out, cache read $0.20. Sonnet 5.5 has the same price with a $0.10 cache read, per Anthropic’s page
• Kimi K3.1 Pro: $2 in, $10 out, cache read $0.20
• Claude Opus 5.5: $4 in, $20 out, cache read $0.20

These numbers come from the analysis, Anthropic’s pricing page and OpenRouter. They move often, so check before you commit.

How to cut your bill

Seven steps. None of them needs code skills beyond changing a model name.

1. Measure first. Pull last month’s usage per task type: tokens in, tokens out, how much was cached. You cannot cut what you have not counted.
2. Sort tasks into easy and hard. Easy: classification, summaries, formatting, data extraction, first drafts. Hard: multi-file code changes, planning, anything where a wrong answer costs real money.
3. Send easy tasks to a cheap model. GPT-6 Luna or DeepSeek Flash. Keep hard tasks on a frontier model.
4. Turn on prompt caching. Keep your system prompt and prefix identical between calls. Cache reads are the biggest line on the bill.
5. Use one gateway. OpenRouter or a similar router gives you one key and one format, so changing models is a config line, not a rewrite.
6. Test before you switch. Run 20 real tasks through the cheap model and score the results yourself. Cheap is only cheap if the output survives.
7. Re-check prices every month. They have moved several times since June, and the author expects them to keep falling.

Do the math: assume a month with 30 million input tokens and 6 million output tokens, no caching, to keep it simple (my assumption).

• All on Claude Opus 5.5: 30 × $4 = $120, plus 6 × $20 = $120. Total: $240 a month.
• All on Claude Sonnet 5.5: $60 plus $60. Total: $120.
• All on GPT-6 Luna: $3 plus $3. Total: $6.
• Mixed: 80% of the work on Luna ($4.80), 20% on Opus ($48). Total: $52.80.

The mixed setup saves $187.20 a month against all-Opus, or about $2,246 a year. That is a 78% cut, right in line with the headline. Real caching changes the exact figures, but the shape stays the same.

Where to start this week

Pick one workflow, not ten. Support-reply drafts are a good first candidate: short inputs, short outputs, and low risk if a human reads the draft before it goes out.

Move only that workflow to a cheap model for a week. Keep the old setup running next to it and compare cost and quality side by side. If the quality holds, move the next workflow. If it does not, you lost an afternoon, not a month of money.

Two habits make this easier. First, log the model name next to every request in your own records, so when the invoice arrives you can see which task burned the money instead of guessing. Second, set a monthly budget alert in your provider dashboard. It takes two minutes and it is the cheapest insurance in this whole article.

Limits and weak spots

• Cheap models fail on hard work. Luna is not Opus. Sending everything to the cheapest model is how you get cheap garbage.
• Benchmark scores are not your task. The analysis cites an index where DeepSeek V4.1 Flash scores 45, Kimi K3.1 Pro 55 and Opus 5.5 58. Your own tests matter more.
• Prices differ by host and change fast. DeepSeek is the clear example above.
• Data location. Check where each provider processes your data before you send customer information to it.
• The author’s favorite is an opinion. He likes Opus 5.5 for value at the top end. That is experience, not a formal benchmark.

Alternatives

• OpenRouter: one key, many models, easy switching.
• Self-hosting an open model: the token bill drops to $0, but you pay for the hardware.
• Direct negotiation: large buyers are reportedly getting steep discounts from both labs. That is unconfirmed, but it is a reason to ask for a better price instead of accepting the list.

What it means

When models become commodities, loyalty stops paying. The labs fight for your usage, so you get the discount. The real cost is lock-in: if your code is written around one provider, you cannot move when a better price appears. Keep your stack portable and let them compete.

Source: Max Woolf · Anthropic

II

About the author

Ivan Ivanov

Founder ofHawkValue

If it doesn’t save you time or money, you won’t see it here

• Facts, not hypeEvery price and number is checked against the official source
• Always the real priceYou see what it costs before you click
• I say when it’s weakEven when it’s OpenAI or Google

All articles by Ivan →