Should my business lock into one AI vendor while prices keep falling?

Last updated 2026-09-29

TL;DR

No. In the week of September 22, 2026, Anthropic cut Opus prices by 20% and OpenAI halved GPT-6 prices, and Epoch AI measures the cost of a given performance level falling about 47% a quarter. Sign short AI contracts, keep your prompts and data portable across models, and re-test vendors at least once a year.

Do not lock your business into one AI vendor for years while model prices are falling this fast. On September 22, 2026, Anthropic and OpenAI both released new flagship models priced below the ones they replaced, and engineers on Hacker News spent the week comparing bills and switching providers. An AI contract signed today at today's prices will look expensive within a year. Buy flexibility instead of a discount.

1. 1. Short termsMonthly or annual AI commitments, no multi-year volume deals
2. 2. Swappable modelPrompts, data, and logs that work with more than one provider
3. 3. Your own test set50 to 100 real examples with known right answers
4. 4. Yearly re-testRun the test set on new and cheaper models every year

How much did AI model prices fall this week?

Anthropic priced Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% below Claude Opus 5. OpenAI priced GPT-6 Sol at $2 and $10, half of GPT-5.6 Sol. GPT-6 Luna costs $0.10 and $0.50, down from $0.20 and $1.20 for GPT-5.6 Luna. Both launches landed on September 22, 2026.

The headline cut understates the change for some workloads. Anthropic's Opus 5.5 announcement cuts cached input from $0.50 to $0.20 per million tokens, 60% less, and says the model costs 40% less than Opus 5 on typical workloads at default settings. OpenAI's API pricing page also halves every rate for batch jobs that can wait.

How fast do AI costs fall per performance level?

Epoch AI, an independent research group, measured the cost of reaching a fixed level of AI performance falling about 47% per quarter since 2023, or roughly 13 times per year. Performance that has just become state of the art gets cheaper fastest: 66% per quarter. Two years later, the same level still falls about 32% per quarter.

The study, The Plunging Price of Thought by Luke Emberson and David Roodman, was published on September 22, 2026, the same day as both launches. It tracks five benchmarks in math, science, and games. The authors warn that benchmark performance is not the same thing as useful work, so treat the rate as a direction, not a forecast for your invoice.

The direction is what matters for a contract. If the task you are automating today needs a top model, the same quality will likely cost a fraction of the price within a year. A commenter on the Hacker News thread about "Tokens too cheap to meter" pointed readers to the Epoch study, while others argued that the fall cannot continue forever. Both points argue for a short commitment.

What are engineers on Hacker News doing about prices?

Engineers on Hacker News are comparing providers task by task and moving work to cheaper models when quality holds. In the 1,129-comment thread on Claude Opus 5.5 and the 346-comment thread on Sonnet 5.5, some developers said they had moved to cheaper Chinese models such as DeepSeek, while others kept paying for Claude or OpenAI.

In the Opus 5.5 thread, one commenter said DeepSeek V4.1 was a "dirt cheap" model that did the work. In the Sonnet 5.5 thread, another said Claude cost 20 times more than the Chinese models they used, and that their employer would no longer pay for it. Those are individual reports, not measurements. DeepSeek's own pricing page lists V4.1 Flash at $0.30 per million input tokens and $1.20 per million output at peak hours.

A second theme matters more to an owner. In the GPT-6 thread, one developer wrote that getting attached to one model made them feel "professionally vulnerable to the labs." Your business carries the same exposure when its workflows depend on one model behaving one way.

Should a business sign a long AI contract now?

A business should avoid multi-year AI commitments while prices fall this quickly. Prefer monthly or annual terms, pay per use where possible, and refuse volume minimums priced at today's rates. A three-year deal signed in 2026 locks in 2026 prices while every competitor on a short term gets the next price cut.

Ask a reseller or AI consultancy four questions before signing. Which model runs underneath, and can you change it? What happens to your price when the provider cuts theirs? Who owns the prompts, the test examples, and the logs? What is the exit fee? A vendor that cannot answer in writing is pricing its own lock-in into your contract.

Models also expire. Anthropic's model deprecations page promises at least 60 days' notice before retiring a public model, and lists Claude Opus 4 as retired on June 15, 2026. A long contract tied to one model version will outlive the model. Budget for a migration every year or two whether you switch vendors or not.

What does a model-agnostic AI setup look like?

A model-agnostic AI setup keeps everything you paid to build outside the model: prompts in your own repository, company documents in your own database, a test set of real examples, and a log of every request. The model call sits behind one small piece of code, so switching from Anthropic to OpenAI or DeepSeek is a configuration change.

1. You own itPrompts, documents, test examples, request logs
2. Vendor owns itThe model, its price, and its retirement date
3. The seamOne connection point your developer can repoint in a day

Do not overbuild today what will be cheap next year. Custom fine-tuning, a private GPU server, or an elaborate cost-saving layer can each make sense at scale, but each one ties you to today's prices and today's models. Start with the simplest version that meets the need, measure it, and spend the savings from each price cut on the parts you own.

Related guides

• What to know before signing a three-year SaaS contract
• The contract clauses that create the most lock-in
• When is an AI API wrapper enough, and when do you need more?
• The Second Opinion: an independent review of an AI contract or initiative
• Epoch AI: The Plunging Price of Thought
• Anthropic: model deprecations and retirement dates

Key takeaways

• Anthropic cut Opus prices 20% and OpenAI halved GPT-6 prices in the same week of September 2026.
• Epoch AI measures the cost of a fixed AI performance level falling about 47% per quarter since 2023.
• Keep AI contracts monthly or annual, with no volume minimums set at today's prices.
• Own your prompts, documents, test examples, and logs so the model can be swapped in a day.
• Re-run your own test set on new and cheaper models at least once a year.

Frequently asked questions

Will AI model prices keep falling?

The trend is strong but not guaranteed. Epoch AI found the cost of a fixed performance level fell about 47% per quarter from 2023 to 2026, across five benchmarks. Commenters on Hacker News noted that no curve falls forever. Plan for further cuts without assuming a specific rate in your budget.

Is it safe to switch my business to a cheaper Chinese AI model?

Test it on your own examples first and check the legal side. Ask where the provider processes and stores your data, what its terms allow, and whether your customers or regulators accept that location. Some teams run open models on their own servers to avoid sending data abroad, which adds hosting cost.

How often should a business re-test its AI vendor?

At least once a year, and whenever a major provider releases a new model or cuts prices. Keep 50 to 100 real examples with known correct answers, run them against each candidate model, and compare accuracy and cost per task. A repeatable test turns a vendor switch into a measured decision.

What should an AI contract say about price cuts?

It should pass provider price cuts through to you or let you leave without penalty. Ask for per-use pricing tied to the underlying model's published rates, a named model you can change, ownership of your prompts, data, and logs, and a short notice period. Have a lawyer review the final terms.

Back to Blog