The AI price war has started. Here is what to do with your budget
Something changed in the last few weeks, and it matters more to a small business than any benchmark score. The frontier labs stopped competing purely on capability and started competing on what it costs you to run a task.
What actually happened
| Event | The cost angle |
|---|---|
| Anthropic released Claude Opus 5 on 24 July 2026 | $5 per million input tokens and $25 output — unchanged from Opus 4.8, and half its own Fable 5 flagship, while claiming near-flagship capability |
| OpenAI's GPT-5.6 family reached general availability on 9 July 2026 | Marketed explicitly on efficiency: OpenAI's own GPT-5.6 page leads with state-of-the-art results using fewer tokens at lower estimated cost |
| Both shipped effort or mode controls | You can now dial how hard the model thinks, trading cost against quality per request |
Sam Altman framed it directly in comments to CNBC around the GPT-5.6 launch, saying enterprises are now thinking about spend and the value they get for it. When the vendor's own launch messaging is about cost efficiency rather than raw intelligence, the market has shifted.
Why capability-per-rupee is the metric that changed
For two years the story was "the new model is smarter." That is useful but hard to act on. Cost per completed task is different, because it changes the arithmetic of what is worth automating.
A workflow that costs ₹40 per document to process is uneconomic if the human alternative costs ₹25. At ₹12 it becomes obviously worth doing. Falling prices do not just save money on what you already run — they move a set of workflows across the line from "too expensive" to "worth building." That is the real opportunity, and it arrives quietly.
What this means for how you buy
1. Stop signing annual AI commitments
An annual prepay locks you into today's price in a market where the price is falling and the capability is rising. The discount is rarely worth the lost optionality. Monthly, until a workflow has proven itself for two quarters — the same rule from our software audit guide.
2. Keep the model swappable
If switching models means a rewrite, you cannot benefit when a cheaper option appears. Keep prompts and business logic separate from the specific model, and keep an evaluation set of 20-30 real tasks you can re-run against any candidate. When a new release lands, adopting it should be a config change plus a test run.
3. Measure on your work, not the leaderboard
Benchmarks measure curated tasks. Your invoices, your customers, your edge cases are not in them. Run your own sample and count usable outputs per rupee. Two days of measurement beats every launch article, this one included.
4. Use the cheap tier deliberately
Most labs now ship a fast/cheap variant and a flagship. Classification, extraction, summarising, and formatting rarely need the expensive model. Routing simple work to the cheap tier is often a larger saving than switching vendors, and it takes an afternoon.
What has not changed
- Verification still costs human time. A cheaper model that needs the same review produces the same bottleneck. The saving shows up in token spend, not in your team's calendar.
- Data handling still matters. Cheaper access does not change retention terms or what you should paste into a chat window. Our AI policy guide covers the classification to apply.
- Fit beats price. The cheapest model that cannot do the job costs infinity per completed task.
- Vendor claims are still marketing. "Comes close to" is doing real work in that sentence, and Anthropic still recommends its flagship for the hardest autonomous jobs.
The honest summary for a small business
If you are not currently paying for AI by the token, none of this changes your week. Consumer chat subscriptions are priced flat and update on their own schedule.
If you are running real volume through an API, do three things this month: re-check your current provider's price against alternatives, route your simple tasks to a cheaper tier, and confirm you could switch models without a rewrite. That is the whole action list.
And if you have been holding off on automating a workflow because the numbers did not work, run the arithmetic again. It may work now. Our framework for choosing AI tools is the sane way to test that without buying another unused subscription, and the Opus 5 analysis shows how to read a launch announcement without being swept along by it.
No company paid for placement in this article. Verify current prices and terms with each provider before buying.