GPT-6.2 Astra Price Cut by 20%, Qwen4 Free, LLMs Running in Browsers—The “Cost of AI” is Crumbling Weekly. How Will SMEs Respond?
Related Articles
Conclusion First: The Era of Estimating AI Utilization Costs on an Annual Basis is Over
In July 2025, a series of events unfolded within a single week:
- OpenAI releases “GPT-6.2 Astra” with a 20% reduction in API token prices.
- Alibaba makes “Qwen4-Max” (a 2.4 trillion parameter sparse MoE) available for free via API, reducing average task costs by 62%.
- Research is published showing a small LLM of about 50MB can operate within a browser tab, eliminating the need for APIs and communication costs.
All three of these events occurred in the same week. Not on a monthly basis, but weekly.
The cost of AI operations that was estimated at 100,000 yen six months ago has dropped to 60,000 yen this week and may fall to 30,000 yen next month. This is not an exaggeration; it reflects the actual speed of price disruption.
For small and medium-sized enterprise (SME) owners, the question has become simple: “It’s not about whether to use AI, but how to ride the wave of this cost collapse.
—
GPT-6.2 Astra—The True Impact of the “20% Reduction”
OpenAI’s new flagship model, “GPT-6.2 Astra,” has lowered the token price for both input and output by 20%. At first glance, this might seem like a mere 20% discount, but the reality is much more profound.
The key point is the compression of output token numbers. Astra is designed to return answers using fewer tokens for the same tasks compared to previous models. This means not only has the price per token decreased, but the amount of tokens used has also reduced. The effects compound.
Let’s calculate this concretely.
For instance, consider a business making 100,000 API calls per month, which previously cost 150,000 yen.
- Token price reduction of 20% → 150,000 yen × 0.8 = 120,000 yen
- Average output token count reduction of 15% (estimated from OpenAI’s published benchmarks) → 120,000 yen × 0.85 = approximately 102,000 yen
The monthly cost drops from 150,000 yen to 102,000 yen, representing an effective 32% cost reduction.
This is the reality behind the “20% reduction.” Judging solely by the headline figure can lead to a misunderstanding of the true impact.
Moreover, this change can be implemented immediately by switching from GPT-4o. It requires just a few lines of code to be rewritten, with zero capital investment. For SMEs, this experience of “reducing costs by 30% just by rewriting code” is unprecedented.
—
Qwen4-Max Free Release—The Implications of “2.4 Trillion Parameters for Free”
Alibaba has made Qwen4-Max available for free. This is a sparse MoE model with 2.4 trillion parameters.
Let’s break down the numbers:
- Average task cost: 62% reduction (compared to Qwen3)
- API usage fee: Free (with a certain call limit)
- Performance: Scores equal to or exceeding GPT-4.5 on major benchmarks
“For free, you can access performance on par with GPT-4.5.” This is the reality as of July 2025.
Think back to a year ago. Companies were paying hundreds of thousands of yen per month to use the GPT-4 API, saying, “This is already cheaper.” Now, equivalent or better performance is available for free.
For SMEs, this means:
The cost of “trying out AI” has become almost zero.
Companies that previously hesitated to adopt AI due to uncertainty about monthly costs now find that barrier has disappeared. They can experiment with Qwen4-Max’s free tier and switch to the paid GPT-6.2 Astra for production use. This pathway of “free experimentation followed by paid production” is now accessible to anyone.
What’s crucial is the reason Alibaba made this free. It’s not charity. They are aiming for an ecosystem. By allowing free usage, they can attract developers and monetize through higher-tier services. Thus, this free offering is not a temporary campaign but a structural price disruption. OpenAI will have no choice but to respond. Prices will continue to fall.
—
50MB LLM Running in Browsers—Entering a World Where “APIs Aren’t Even Needed”
The third piece of news is understated but structurally fascinating.
Research has been published showing that a small language model of around 50MB can operate within a browser tab. Utilizing WebGPU and WebAssembly, it can perform text generation and classification tasks without any communication with servers.
What does this mean?
- API call cost: Zero (as it operates locally)
- Communication latency: Zero
- No external data transmission (company data remains in-house)
While it doesn’t yet match the versatility of GPT-6.2, it is sufficiently practical for routine classification, summarization, and extraction tasks.
Consider this in the context of SMEs. For example, automating the classification of inquiry emails, extracting amounts from quotes, or summarizing daily reports. Tasks that are done daily but don’t require much thought can now be automated simply by opening a browser. Zero monthly costs, zero risk of data leakage.
The previous notion of “using AI = paying for external APIs” is beginning to crumble.
While many of these solutions are still in the research phase, the direction is clear. The execution environment for AI is moving from the cloud to the edge, and from the edge to the browser. This is a movement that will drastically lower costs.
—
Should SMEs “Wait” or “Act”?
Having read this far, some may wonder, “Wouldn’t it be better to wait a little longer for even lower prices?” To be honest, that’s true. Prices will drop further next month and even more in three months.
However, the strategy of “waiting for lower prices” has already failed.
There are two reasons:
1. Competitors are not waiting.
Your neighboring company might try Qwen4-Max for free this month and start operating with GPT-6.2 Astra next month. By the time you say, “Maybe it’s time to implement it,” in three months, your competitor will have already accumulated three months’ worth of operational data and know-how. The differentiation in AI adoption will not come from the tool’s performance but from the tuning and operational accumulation tailored to your business. This cannot be gained by waiting.
2. With the “cost of experimentation” being almost zero, there’s no reason to wait.
Qwen4-Max is free. The browser LLM is also free. GPT-6.2 Astra can be tried for just a few thousand yen. The rationale of “waiting due to cost concerns” no longer holds. If it doesn’t work out, you can simply stop. The only thing you risk losing is a few hours of testing time.
So, what should you do specifically?
—
Three Things SMEs Should Do This Week
① Automate one routine task in your company using the free API of Qwen4-Max.
Whether it’s email classification, meeting minutes summarization, or FAQ responses, anything will do. Since it’s free, there’s no cost even if you fail. The important thing is to experience “what AI can do” firsthand. Reading catalog specifications won’t get you anywhere.
② Recalculate the costs of existing AI operations with GPT-6.2 Astra.
Companies already using OpenAI’s API may see an immediate cost reduction of around 30% by switching to Astra. It’s as simple as rewriting the model name. Check your invoice this month and calculate how much you will save next month.
③ Add the trends of “browser LLMs” to your watchlist.
While it may be too early for full-scale deployment, it’s highly likely that they will reach practical levels in six months. Especially for companies handling “tasks that cannot send data outside,” this technology could provide a solution. It’s worth listing which tasks might fit this category now.
—
What Happens After the “Cost of AI” Crumbles?
Finally, let’s discuss what lies ahead.
As the cost of utilizing AI approaches zero, “using AI” will no longer be a differentiator. Since everyone can use it, that’s only natural.
So what will become the differentiator?
The sense of “what to do with AI” and the design ability to “integrate it into your own operations.”
This is an area where SMEs have an advantage over large corporations. Decision-making is faster. The distance between the field and management is shorter. “Let’s change it starting next week” can genuinely mean a change next week. While large corporations are still going through bureaucratic processes, SMEs can complete three cycles of trial and error.
The collapse of AI costs is not a threat to SMEs. It’s a historic opportunity where the handicap of “financial power differences” disappears.
The question is whether they will recognize this opportunity and take action. That’s all.
Costs are crumbling on a weekly basis. New models will emerge again next week. There’s no time to wait. Start by creating an account for Qwen4-Max today.
JA
EN