A few months ago, I was building an AI product and hit a wall: Which LLM should we use?
The question sounds simple. But the pricing landscape for AI models shifts constantly and without warning. OpenAI has 69 models listed on their pricing page. Google has 18. Anthropic launched three new Claude versions in the span of weeks. By the time I had compared options and made a decision, I had no confidence the numbers were still current.
I checked pricing pages. I checked docs. I found no single place that told me:
So I automated it. Starting April 24, 2026, I built a daily scraper that pulls pricing data from every major provider. Here is what 8 weeks of data actually shows.
1,111models tracked
33providers
59days of data
5price changes detected
The headline finding is not dramatic: Anthropic, OpenAI, Google, Mistral, and xAI all held their prices flat for the entire 59-day period. Not a single price change detected across those five providers.
That is useful information on its own. If you are budgeting around these providers, your cost model from April is still accurate today.
Here is the current pricing snapshot for the most-used models, accurate as of June 22:
| Model | Input / 1M tokens | Output / 1M tokens | Note |
|---|---|---|---|
| openai / gpt-4-turbo | $10.00 | $30.00 | Stable since Apr 24 |
| openai / gpt-4o (chatgpt-4o-latest) | $5.00 | $15.00 | Stable since May 12 |
| anthropic / Claude Sonnet 4.5 | $3.00 | $15.00 | Stable since Apr 24 |
| anthropic / Claude Opus 4.5 | $5.00 | $25.00 | Stable since Apr 24 |
| anthropic / Claude Haiku 4.5 | $1.00 | $5.00 | Stable since Apr 24 |
| google / Gemini 2.5 Pro | $1.25 | $10.00 | Stable since Apr 26 |
| google / Gemini 2.5 Flash | $0.30 | $2.50 | Stable since Apr 26 |
| google / Gemini 2.0 Flash | $0.10 | $0.40 | Stable since Apr 26 |
| mistral / Codestral | $1.00 | $3.00 | Stable since Apr 24 |
| xai / Grok 4.3 | $1.25 | $2.50 | Stable since May 16 |
For teams budgeting: The major closed-model providers are behaving predictably right now. Your April cost model is still valid. That said, 8 weeks is a short window. Pricing for these providers has historically shifted without notice.
The bigger story is not volatility. It is the 600x price range that now exists across the tracked catalog, from sub-cent inference to frontier flagship models.
| Model | Input / 1M tokens | Output / 1M tokens | Note |
|---|---|---|---|
| groq / llama-3.1-8b-instant | $0.05 | $0.08 | Cheapest tracked |
| deepseek / deepseek-v4-flash | $0.14 | $0.28 | |
| google / Gemini 2.0 Flash | $0.10 | $0.40 | |
| deepseek / deepseek-v4-pro | $0.44 | $0.87 | |
| xai / Grok 4.3 | $1.25 | $2.50 | |
| anthropic / Claude Sonnet 4.5 | $3.00 | $15.00 | |
| openai / gpt-4-turbo | $10.00 | $30.00 | |
| openai / gpt-5.4-pro | $30.00 | $180.00 | Highest-cost frontier tracked |
Groq's Llama 3.1 8B costs $0.05 per million input tokens. GPT-4 Turbo costs $10.00 (200x). GPT-5.4 Pro costs $30.00 (600x). All three are in our dataset today. DeepSeek V4 Pro at $0.44 per million input tokens consistently ranks near the top on reasoning benchmarks at a fraction of the flagship price.
The question teams should be asking is not just “which model is best?” but “which model is best for this specific task given the cost?” For many production workloads, the answer is not the model you started with.
Real cost difference: A team running 100M input tokens per month on GPT-5.4 Pro spends $3,000. The same usage on GPT-4 Turbo costs $1,000. On DeepSeek V4 Pro it is $44. On Groq's Llama 3.1 8B it is $5. The right answer depends entirely on what the task requires, but most teams never run this comparison after their initial model choice.
Over 59 days, our scraper detected 5 pricing changes across all 1,111 models. Every single one was on Together AI, a hosting aggregator that runs third-party open-source models.
| Date | Model | Before | After | Change |
|---|---|---|---|---|
| May 27 | qwen37-max | $2.50 / $7.50 | $1.25 / $3.75 | -50% |
| Jun 2 | Qwen3.5-9B | $0.10 / $0.15 | $0.17 / $0.25 | +70% |
| Jun 2 | Llama-3.3-70B-Turbo | $0.88 / $0.88 | $1.04 / $1.04 | +18% |
| Jun 2 | Meta-Llama-3-8B-Lite | $0.10 / $0.10 | $0.14 / $0.14 | +40% |
| Jun 17 | DeepSeek-V4-Pro | $2.10 / $4.40 | $1.74 / $3.48 | -17% / -21% |
The biggest move: Qwen37-max dropped 50% overnight on May 27, with no public announcement that reached mainstream channels. Teams running that model through Together AI saw their costs cut in half. Teams not monitoring pricing had no idea.
Three models saw price increases on June 2: Qwen3.5-9B (+70%), Llama-3.3-70B-Turbo (+18%), and Meta-Llama-3-8B-Lite (+40%). Not all pricing movement favors the buyer.
Why this matters: None of these changes came with direct user notifications. No email, no in-dashboard alert, no API changelog. The only way to know was to check the pricing page, which almost nobody does after the initial setup.
Across all 5 detected changes, the announcement path was the same: nothing sent directly to users. One change (DeepSeek-V4-Pro via Together) appeared in a provider blog post days later. The rest were silent.
Compare that to how other infrastructure pricing works:
LLM providers are operating more like spot markets than enterprise software. Prices move when they move. Most teams find out on their invoice.
After a few weeks of running this manually, I automated it and opened it up. Token Prices tracks 1,100+ models across 33 providers daily and surfaces changes the moment they happen.
The tool gives you:
The free tier covers the 10 major providers. Paid plans add historical depth, more providers, and API access. tokenprices.io
Let us know at support@tokenprices.io
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。