The AI model price war: what changes for companies that automate

What was announced

The figures below are API prices per million tokens, as published by the vendors and compiled by technical publications during launch week.

• GPT-6 Sol (OpenAI, Sep 22): US$2 input and US$10 output, about half the price of GPT-5.6 Sol, which cost US$4 and US$20.
• GPT-6 Luna (OpenAI, Sep 22): US$0.10 input and US$0.50 output, also about 50% below the previous version. It is one of the cheapest models OpenAI has ever released.
• Grok 4.7 (xAI, Sep 21): US$2 input and US$6 output for prompts of up to 200,000 tokens — the same input price as Sol, with output 40% cheaper.
• Claude Opus 5.5 (Anthropic, Sep 22): US$4 input and US$20 output, 20% below Opus 5.0, with cache reads 60% cheaper.

Why prices are falling

The reduction follows two trends. The first is competitive: with several labs shipping models with similar performance, price has become a direct selling point. The second is technical: more efficient models use less compute to produce the same answer, and part of those savings reaches the list price.

The practical effect is that tasks that required a frontier model a year ago can now be handled by a mid-tier or small model, at a fraction of the cost.

The model is rarely the largest cost

In AI automation projects, token usage is usually a smaller share of the total cost. What weighs more is the work of integrating with existing systems, organizing data, defining business rules, monitoring and maintenance over time.

Automated customer service on WhatsApp, for example, can exchange thousands of messages a month and still have a low model cost compared with the effort of keeping the CRM, inventory and calendar integrations working correctly.

That is why a 50% drop in token price does not mean a project that is 50% cheaper. It means the variable portion got smaller — which matters mainly in high-volume operations.

What is worth reassessing now

Even with that caveat, the change opens the door to a few concrete reviews:

• Projects shelved because of cost: high-volume automations that did not add up may have become viable, especially with small models like Luna.
• Choosing the model per task: classifying a message or extracting a data point does not require the same model that drafts a proposal. Splitting the steps and using the right model for each one cuts cost without losing quality.
• Using caching: with cheaper cache reads, instructions and context repeated in every conversation cost much less.
• Reliance on a single vendor: with prices changing every few months, an architecture that lets you switch models without rewriting the system is worth more than ever.

A lower price does not replace evaluation

Switching models just because one got cheaper, without testing, is a risk. Different models behave differently given the same instructions, and a cost gain can come with more errors in cases specific to your business.

The safe path is to keep a set of real test cases and run each candidate against it before any switch in production. That way the decision is based on results, not on a price sheet.

Frequently asked questions

Should I switch the model I use today for a cheaper one?

Only after testing. Ideally, run the new model against a set of real cases from your operation and compare quality and cost. If results hold, the switch makes sense; if they get worse in important cases, the savings can end up being expensive.

Does the price drop make AI implementation cheaper?

It reduces the variable cost of usage, but not the cost of implementation. System integration, data organization and monitoring are still most of the investment. The effect is greatest in operations with a high volume of messages or documents.

Will prices keep falling?

The trend in recent years has been downward, driven by competition and technical efficiency, but there is no guarantee of pace. That is why it is more prudent to design systems that allow switching models than to bet on a specific future price.