What AI Credit Pricing Reveals About the Future of Free Information

Many AI product pricing pages now quote credits instead of familiar units such as dollars, minutes or documents. There is no industry standard for what a credit buys, so two products that each offer “500 credits a month” can deliver very different amounts of work. Large vendors now sell both ways: OpenAI’s pricing page, for example, states that credit-based and token-based pricing are both available on its Enterprise plans (OpenAI enterprise pricing).

Credits can reflect both a vendor’s costs and its chosen business model. Underneath many text-based credits sits the token, a unit that can represent a word, part of a word, or punctuation. Tokens are priced differently even within one model. When checked on November 6, 2026, OpenAI’s standard API price list showed gpt-6.1-sol at $2.00 per million input tokens, $0.10 per million cached input tokens and $10.00 per million output tokens (OpenAI API pricing), and Anthropic listed Claude Sonnet 5.5 at $2 per million input tokens, $0.10 for cache hits and $10 for output (Claude API pricing). Cache writes, processing modes, tools, and other features can carry additional charges. Reasoning models add a further layer: they generate internal reasoning tokens that the user never sees but that OpenAI bills as output tokens (Reasoning token billing). When a bill reports “1,240 tokens consumed,” it describes the vendor’s billing unit, not the amount of work you got done.

Why AI Pricing Is Hard to Predict

Usage is uneven and hard to forecast

The situation becomes more complex inside organizations. Usage is rarely even: a few heavy users, or one automated workflow, can consume a large share of a shared credit pool, and teams end up watching dashboards and adding approval steps. Buyers have noticed. In a 2026 survey of 50 technical AI buyers by the venture firm Andreessen Horowitz, 27 preferred credits tied to a recognizable unit of work and only 14 preferred token-based pricing (AI buyer pricing survey). A small survey run by investors is not a representative sample, but it suggests that some buyers want to pay for outcomes they can recognize.

Budgeting habits that help

A few habits help:

• Set a spending cap where supported, add alerts, and run a representative pilot to estimate workload costs. Confirm whether limits are hard stops or delayed billing alerts.
• Prefer tools that price around finished work, such as documents processed or searches run, when the work is repetitive.
• Reserve reasoning modes for tasks that need them, since their hidden reasoning tokens are billed as output.
• Read the fair-use terms of “unlimited” plans; they describe the limits that actually apply.

Prices for a given capability are falling fast

At the same time, the price of a given level of AI capability has fallen quickly. Stanford’s 2025 AI Index reports that the cost of querying a model that performs at GPT-3.5’s level on the MMLU benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a drop of more than 280-fold (Stanford AI Index). Epoch AI found that the price of reaching a fixed performance level fell by between 9 and 900 times per year depending on the benchmark, and by about 40 times per year for GPT-4-level performance on PhD-level science questions. Epoch cautions that the fastest declines are recent and may not last (Inference price trends). Cheaper capability does not automatically mean free access, but it lowers the cost of the basic tiers that free and education plans are built on. Public programs also help: the U.S. National Science Foundation’s National AI Research Resource, established as a pilot in 2024, gives researchers and educators access to computing, data, models and training (NSF research resource).

Why AI Answers Can Cost More to Serve Than Search Results

A common question is why ChatGPT cannot simply copy Google Search’s free, ad-supported model. One important consideration is inference cost, although actual costs vary by model, infrastructure, caching, and workload. Search engines amortize crawling and indexing across many queries, while still incurring serving costs. Generating an AI answer generally requires inference at request time, although prompt caching can reuse prior computation. Cost can grow with prompt length, output length, reasoning effort, and tool use.

What the early estimates showed

Estimates of the gap are rough and now dated. In February 2023, the analyst firm SemiAnalysis estimated that ChatGPT cost about 0.36 cents per query to run, against about 1.61 cents of revenue that Google earned per search in 2022. On those assumptions, serving every Google search with a ChatGPT-style model would have cut Google’s operating income by about $36 billion a year (Early inference cost analysis). This compares estimated AI serving cost with search revenue, not two directly measured serving costs. The estimate predates the price falls described above, so read it as a scenario based on early-2023 assumptions, not as today’s cost.

Less room for advertising

A second challenge is advertising design. A traditional results page carries several paid links alongside organic results. An AI answer consolidates information into a single response, which changes where ads can appear and makes clear separation of sponsored content an important design requirement.

Who Pays: Subscriptions, Advertising and Commerce

The industry is testing several approaches. Perplexity began testing ads in the U.S. in November 2024 as sponsored follow-up questions and paid placements beside an answer, stating that brands do not write or edit the answers (Perplexity advertising announcement). OpenAI launched Instant Checkout in ChatGPT in September 2025, letting U.S. users buy from Etsy sellers inside a chat; merchants pay a fee on completed purchases, and OpenAI stated that product results were unsponsored (ChatGPT Instant Checkout). Microsoft launched Copilot Checkout for in-chat purchases in the U.S. in January 2026 (Copilot Checkout reporting). Microsoft Advertising reported, from its own data for February to May 2025, 73% higher click-through rates and 16% stronger conversion rates for ads in Copilot than in traditional search; treat those vendor-reported comparisons as context, not a guaranteed advertiser outcome (Microsoft advertising results).

Advertising in chat interfaces is being tested and monetized. In January 2026, OpenAI announced that it would test ads for logged-in adult users on its Free and Go tiers in the U.S. It said ads would be labeled, separate from answers and kept away from sensitive or regulated topics such as health, mental health and politics (OpenAI advertising policy). In March 2026, an OpenAI spokesperson told Reuters that the U.S. pilot had passed $100 million in annualized revenue within six weeks of launch (ChatGPT ad pilot reporting). Alphabet reported total revenues of $402.8 billion for 2025, more than 70% from online advertising, and $82.3 billion of Google advertising revenue in the fourth quarter alone (Alphabet earnings release; Alphabet annual filing). Privacy is an important concern. In February 2026, OpenAI researcher Zoë Hitzig resigned and argued in a New York Times essay that personal disclosures to chatbots create unusual risks when advertising is involved (ChatGPT advertising privacy debate).

Engineering Efficiency Versus Agentic Demand

Cheaper ways to serve a model

Engineers are cutting model-serving costs in several ways. Distillation trains a smaller model to imitate a larger one: the original DistilBERT was 40% smaller and 60% faster than BERT while retaining about 97% of its language-understanding performance (DistilBERT research paper). Prompt caching lets a provider reuse the processed form of a long, repeated prompt, such as a reference document or standing instructions, instead of recomputing it. Anthropic said caching could cut costs by up to 90% and latency by up to 85% for long prompts (Claude prompt caching), and current price lists charge cache hits at a fraction of the normal input rate. Prompt caching reuses processed input, not finished answers. Routing simple queries to cheaper models and reserving large models for difficult problems can lower average costs further.

More calls per task

AI products are also becoming more agentic, often making many model calls in the background to complete a single task. Falling per-call prices can be offset by more calls per task. The balance between these trends will shape future costs, so measure the complete workflow rather than relying on one headline price.

What This Means for Research Teams

Budget for verification, not just generation

For research teams, the useful unit is the verified answer. A tool that is cheap per query but needs extensive checking can cost more in staff time than a pricier tool that produces more usable results. All consequential claims still need verification. In a 2023 audit of four AI search engines, only about half of generated sentences were fully supported by their citations (Generative search citation audit). For a fuller method, see AI answer verification guide.

Compare tools by cost per verified answer

A repeatable test makes tools comparable regardless of how they price:

1. Pick a fixed test set of 20 to 30 real questions from your work, each with a known answer or primary source.
2. Run every tool on the same set and record credits or tokens used, converted to money at your plan’s actual rate. Include allocated subscription and tool fees where relevant.
3. Check each answer and citation against primary evidence and record verification time.
4. Calculate cost per verified answer: (usage cost + verification minutes × staff cost per minute) ÷ answers that were correct and properly sourced. Record rejected or unanswered questions separately.
5. Repeat after major model or pricing changes, and date your results.

For example, if a tool uses $30 of credits on 25 questions, verification takes 100 minutes at $60 an hour, and 20 answers check out, the cost per verified answer is ($30 + $100) ÷ 20 = $6.50. These figures are illustrative. If no answers qualify, report the test’s cost and failure rate rather than dividing by zero.

Access and equity

Credit pricing also affects access. Limits, model availability, and advertising differ by vendor and plan; OpenAI’s January 2026 ad-test announcement covered Free and Go, while excluding Plus, Pro, Business and Enterprise (OpenAI advertising policy). Check current terms before purchase. Equal access is not the same as equal ability to use a tool. A 2025 mixed-methods study of 135 academic staff at 13 Norwegian higher-education institutions, with follow-up focus groups, found that most respondents expected AI to create divides based on AI literacy, critical thinking and precise instructions as well as access (AI digital divides study). The sample was self-selected. Teams should record the tier and model used, and libraries and research offices can support access alongside verification training.

Key Takeaways

• Credits have no standard value; convert them to money and units of work before comparing tools.
• Check input, output, cache-write, cache-hit, reasoning, and tool charges. Date any price you rely on.
• Prices for fixed benchmark capability have fallen sharply, but agents can make more calls per task.
• Advertising, subscriptions, and in-chat commerce are different ways of funding access; none makes computation costless.
• Read advertising and privacy policies before relying on a free tier.
• Measure cost per verified answer, including staff checking time, rather than cost per query alone.

The AOFIRS research report The Pricing of AI Intelligence explores these questions in more depth, and AI versus traditional search compares how the tools find information.

References

1. Alphabet Inc. (2026a, February 4). Alphabet announces fourth quarter and fiscal year 2025 results. Alphabet earnings release.
2. Alphabet Inc. (2026b). Form 10-K for the fiscal year ended December 31, 2025. Alphabet annual filing.
3. Anthropic. (2025, August 14). Prompt caching with Claude. Claude prompt caching.
4. Anthropic. (2026). Pricing. Claude API documentation. Accessed November 6, 2026. Claude API pricing.
5. Axios. (2026, January 8). Microsoft turns Copilot chats into a checkout lane. Copilot Checkout reporting.
6. Babu, J. (2026, March 27). OpenAI’s US ad pilot exceeds $100 million in annualized revenue in six weeks. Reuters, via The Star. ChatGPT ad pilot reporting.
7. Cottier, B., Snodin, B., Owen, D., & Adamczewski, T. (2025, March 12). LLM inference prices have fallen rapidly but unequally across tasks. Epoch AI. Inference price trends.
8. Erten, T., & Wang, S. (2026, August 27). You are not a model. Don’t price per token. Andreessen Horowitz. AI buyer pricing survey.
9. Farrell, J. (2026, February 11). OpenAI researcher quits over slippery slope of ChatGPT ads. SiliconANGLE. ChatGPT advertising privacy debate.
10. Hodøl, H.-O., Kjærgård, P. I., & Økland, Ø. (2025). From access to ability: Educator’s perspectives on AI and new digital divides in higher education. Digital Society, 4(3), Article 78. AI digital divides study.
11. Klein, K., & Longo, P. (2025, August 6). 73% higher CTRs: Why advertisers need to pay attention to conversational AI. Microsoft Advertising. Microsoft advertising results.
12. Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. Generative search citation audit.
13. National Science Foundation. National Artificial Intelligence Research Resource. Retrieved November 6, 2026. NSF research resource.
14. OpenAI. (2025, September 29). Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol. ChatGPT Instant Checkout.
15. OpenAI. (2026a). Pricing. Accessed November 6, 2026. OpenAI enterprise pricing.
16. OpenAI. (2026b). API pricing. Accessed November 6, 2026. OpenAI API pricing.
17. OpenAI. (2026c). Reasoning models. Accessed November 6, 2026. Reasoning token billing.
18. OpenAI. (2026d, January 16). Our approach to advertising and expanding access. OpenAI advertising policy.
19. Patel, D., & Ahmad, A. (2023, February 9). The inference cost of search disruption. SemiAnalysis. Early inference cost analysis.
20. Perplexity. (2024, November 12). Why we’re experimenting with advertising. Perplexity advertising announcement.
21. Sanh, V., Debut, L., Chaumond, J., & Wolf, T. (2019). DistilBERT, a distilled version of BERT: Smaller, faster, cheaper and lighter. DistilBERT research paper.
22. Stanford Institute for Human-Centered Artificial Intelligence. (2025). The 2025 AI Index report. Stanford AI Index.