99% Cheaper AI Models Put OpenAI's IPO Math at Risk

The uncomfortable question for AI companies is no longer whether their best model is smart enough. It’s whether customers are about to realize they’ve been overpaying for intelligence they don’t always need.

Can the AI boom survive if “good enough” gets much cheaper?

The AI industry has mostly sold one idea: bigger models win. That logic justified premium pricing, heavy infrastructure spending, and the habit of sending too many workloads to the most advanced model available. Now that logic is under pressure, according to TechCrunch, because companies are starting to test whether smaller and cheaper AI models can handle real work without degrading quality.

The sharpest version of the case is the idea that a large share of routine AI work may not need the most powerful model available. If cheaper models can handle enough of those tasks, the AI market doesn’t just get cheaper. It gets repriced.

For OpenAI and Anthropic, the risk is direct: if the same tasks can be handled by lower-cost models, premium labs may capture less of the spending that currently flows to frontier systems. That doesn’t mean frontier models stop mattering. It means their premium must be earned task by task, not assumed by default.

XOOMAR analysis: the next phase of AI won’t be decided only by who builds the largest model. It will be decided by who can match the task to the least expensive model that still clears the quality bar.

Where does the new AI cost pressure actually show up?

The cost debate starts with inference, not training. Training frontier models is still expensive, but customers feel the pain every time a product calls a model inside a daily workflow. The more AI moves from demos into document review, coding assistants, search-like interactions, customer support, internal analytics, and compliance checks, the more each token matters.

The additional context supplied from Forbes captures the pricing tension. OpenAI’s ChatGPT and Microsoft’s GitHub Copilot helped set a $20/month psychological anchor for AI tools. But more capable systems can cost far more to run. The same Forbes analysis says OpenAI’s o1 costs $60 per million output tokens, while o1-pro costs $600 per million output tokens. It also cites ChatGPT Pro at $200/month, and says Sam Altman acknowledged OpenAI was “losing money on OpenAI Pro subscriptions” because “people use it much more than we expected.”

That matters because agentic workflows multiply usage. A coding agent may search files, pull context, edit code, and reprocess the expanded context after each tool call. Larger context windows also raise costs. The Forbes source says Gemini 2.5 Pro has a 1 million token context window, while Claude models offer up to 200K tokens.

The cost curve looks simple from a distance: cheaper tokens, cheaper AI. Up close, it’s messier.

• More context: Better answers often require more input, which raises token volume.
• More tool use: Each tool result can force the model to reprocess added context.
• More reliability: Premium models may use extra processes to reduce errors.
• More usage: Better AI invites users to ask it to do more.

XOOMAR analysis: cheaper models matter most when they cut inference spend without forcing users to shrink prompts, reduce context, or abandon workflows. If they only save money by making the product worse, enterprises won’t switch at scale.

Why are smaller models suddenly credible for enterprise work?

The strongest argument for smaller models is not that they beat frontier models across the board. It is that many enterprise tasks may not require the same level of intelligence every time. Some workflows need maximum reasoning, while others need speed, consistency, and a low enough cost to run repeatedly.

That is the real story. Not “small beats big.” The story is selective use.

This reframes AI procurement. A company doesn’t need one perfect model for everything. It needs a system that knows when a cheaper model is enough and when the expensive one is justified.

TechCrunch also warns against framing this only as a fight between proprietary labs and open-weight alternatives. The more important split is large models versus small models. A company might save money by moving a task from a frontier system to a cheaper independently served model, but a smaller proprietary model from an established provider could also be enough.

That distinction matters. If small proprietary models, open-weight models, and independently served alternatives all compete for the same lower-cost workloads, premium frontier pricing faces pressure from several directions at once.

XOOMAR analysis: model routing turns AI from a one-model product into a cost-control stack. The hard part won’t be finding cheaper models. It will be proving, with evaluations, that the cheaper model gives the right answer often enough for the job.

Does cheaper AI weaken the case for frontier models?

Not completely. Even the cheaper-model argument leaves room for latest-generation systems where maximum capability is important. TechCrunch makes the same point indirectly: the question isn’t whether frontier models vanish, but whether they remain the default.

The industry got here through a scaling-first mindset. TechCrunch points to the “bitter lesson,” the idea that broad progress in AI has come from throwing more compute at general methods. Labs leaned into that lesson by training the most compute-intensive models they could. Customers, while prices were heavily subsidized by investors, had little reason to choose anything but the most advanced option.

That logic is now harder to take for granted. Even if token prices keep falling on an apples-to-apples basis, real-world AI bills can still rise when products use more context, more tool calls, and more automated steps. Users are now facing cost pressure in production systems, not just comparing model price sheets.

This creates three possible responses:

• Switch models: Move many tasks to smaller, cheaper models.
• Use less AI: Make fewer calls or reduce context.
• Cut weak deployments: Drop projects that don’t justify their cost.

Only the first outcome is bullish for cheaper-model adoption. The other two would reduce demand without proving that smaller models can take over.

XOOMAR analysis: the bear case for big labs isn’t that nobody needs frontier intelligence. It’s that frontier intelligence becomes a premium tier used selectively, while routine workloads migrate elsewhere.

Who benefits if enterprises stop defaulting to the biggest model?

The immediate winners are AI app companies that can lower inference costs while preserving output quality. If an app pays less for the same customer-visible result, its unit economics improve.

Enterprise buyers also gain negotiating power. Once a vendor admits that different models can handle different tasks, procurement teams can ask a sharper question: why is this workflow priced as if every prompt needs the most expensive model?

Developers get a different mandate. Hard-coding around one provider becomes risky if model prices and quality keep shifting. Applications need model flexibility, fallback paths, observability, and task-level evaluation. The product should know when to spend and when not to.

For model providers, the pressure is harsher. TechCrunch says there is already an active price war between in-house inference from big labs and independently served open-weight models. If cheaper models preserve quality across a large share of workloads, premium providers must defend their pricing with measurable performance, not brand gravity.

XOOMAR analysis: this is where the AI market starts to look more like software procurement. Buyers won’t just ask, “Does it work?” They’ll ask, “Does it work at the lowest defensible cost?”

Which question won’t be answered for months?

The unresolved question is whether enterprises will actually switch. TechCrunch is careful here. Cost pressure might push users toward smaller models, but it could also make them use less context, make fewer calls, or abandon marginal AI deployments.

That uncertainty is the center of the story.

The cheaper-model prediction is bold. Forbes’ pricing examples show why the pressure is real. But broad migration depends on evidence inside production systems, not benchmark charts or vendor promises.

The evidence to watch is practical:

• Quality retention: Cheaper models must match required accuracy in live workflows.
• Routing success: Systems must reliably identify when a task needs a premium model.
• Usage behavior: Lower costs should increase useful deployment, not just expose weak demand.
• Provider pricing: Big labs must decide whether to cut prices, push mini models, or protect premium tiers.
• Enterprise contracts: Buyers will look for task-level cost transparency.

The cheaper-model era won’t kill demand for advanced AI. It may make AI more common by forcing the industry to stop wasting expensive intelligence on routine work. The companies to watch are the ones that make that substitution invisible: same answer, lower bill, fewer excuses.

The Bottom Line

• AI customers may cut spending by routing routine tasks to cheaper models.
• Premium AI labs like OpenAI and Anthropic may face pressure to justify higher prices.
• The next AI advantage may come from matching each task to the lowest-cost model that works.