Signals Inbox·July 28, 2026·Frontier AI

Will cheap AI models kill OpenAI and Anthropic?

Cheap AI models will not kill OpenAI or Anthropic, but they are already dismantling the comfortable economics of charging frontier prices for work that increasingly capable alternatives can do for a fraction of the cost.

Q1Why do cheap AI models suddenly look dangerous?

Cheap AI models look genuinely dangerous today because their prices have collapsed while their capabilities have moved much closer to the frontier.

A study covering 318 commercial models and more than 3,000 historical model releases estimated that token prices fell about 600-fold between 2020 and 2026. The decline also accelerated after 2024, suggesting that competition between providers is now pushing prices down alongside technical improvements.

Chinese laboratories no longer look like distant followers either. Moonshot AI’s Kimi K3 quickly reached the top group on independent coding and reasoning rankings. Demand became so intense that Moonshot temporarily stopped accepting new subscriptions. DeepSeek, Z.ai and Alibaba have also released models that developers can run cheaply, modify or host themselves.

Corporate behavior is changing at the same time. Recent reporting by The Wall Street Journal found that companies including Zoom, Hex and Telnyx were becoming much more selective about when they used expensive frontier models. Amazon has also been reducing Alexa’s dependence on Anthropic’s Claude models to control inference costs.

Cheaper models are getting capable, switching is becoming easier and buyers are finally asking whether every task needs the best available model. That is a much bigger threat than another low-priced chatbot entering the market.

Q2What would it take to actually “kill” OpenAI or Anthropic?

Cheap models would kill OpenAI or Anthropic only if they broke the economics of staying at the frontier and pushed the companies away from their customers.

Taking routine summarization, translation or classification work would hurt API revenue, but both companies could remain enormous businesses while handling fewer tokens. Premium models could still dominate difficult coding, research, financial analysis and autonomous work.

Corporate failure would require several problems to arrive together. Customers would need to stop paying extra for frontier performance. Products such as ChatGPT and Claude Code would need to lose their grip on users. Revenue would then have to fall behind the cost of training models and operating data centers, while investors became unwilling to finance the gap.

We judge the threat through three practical questions. Can cheaper models replace GPT and Claude on valuable work? Can OpenAI and Anthropic keep direct relationships with users? Can their revenue support the extraordinary cost of remaining competitive?

Anything short of that means margin pressure, declining market share or a weaker strategic position. Painful, yes. Corporate death is a much higher bar.

Q3How much cheaper are the alternatives right now?

The cheapest serious alternatives currently undercut the main GPT and Claude models by roughly six to 35 times on generated output.

DeepSeek V4 Pro costs $0.435 per million input tokens and $0.87 per million output tokens. OpenAI charges $2.50 and $15 for GPT-5.6 Terra, while Anthropic charges $3 and $15 for Claude Sonnet 4.6. At the premium end, GPT-5.6 Sol costs $5 for input and $30 for output, while Claude Opus 5 costs $5 and $25.

List prices do not reveal the complete cost of finishing a task. A weaker model may write more, require several attempts or create extra review work. Caching and batch processing can also reduce the published prices substantially.

Still, a gap above 15 times gives engineering teams plenty of room to experiment. A cheap model can waste half its calls and remain less expensive.

OpenAI can fight back inside its own model portfolio. GPT-5.4 Nano costs only $0.20 for input and $1.25 for output. The pricing war reaches beyond Chinese models: OpenAI and Anthropic can release cheaper tiers themselves, though doing so moves customers away from their higher-priced products.

Current model API list prices

Q4Are cheap AI models already good enough?

Cheap AI models are already good enough for a large share of everyday business work. They still lose too often on difficult, long-running tasks.

Stanford’s latest AI Index found that the performance gap between the leading American and Chinese models had narrowed to 2.7% on its selected comparison. The two countries had traded the lead several times since early 2025. That is a much tighter race than the public distinction between “frontier” and “second-tier” models suggests.

Kimi K3 provides a recent example. Independent rankings placed it close to the leading Claude and GPT models overall, with particularly strong performance on frontend coding. Developers can download its weights, use another company to host it or operate it on their own infrastructure.

An academic study of more than 100 trillion tokens processed through OpenRouter found substantial adoption of open-weight models. The usage went well beyond experimentation. Developers were sending production coding, role-playing, application and agent workloads through a wide range of providers.

For extraction, document tagging, basic customer support, translation, first drafts and easily tested code, a model does not need to win every benchmark. It needs to complete the task cheaply enough and accurately enough.

That describes a huge portion of current AI usage. OpenAI and Anthropic can lead at the difficult edge while watching much of the surrounding volume move elsewhere.

Q5Do small benchmark gaps still matter in real work?

Small benchmark gaps can produce large differences once a model must complete several steps without human help.

Imagine that one model succeeds on 90% of individual steps while another succeeds on 85%. Across a ten-step process, the first would finish every step correctly about 35% of the time. The second would manage roughly 20%, assuming each step carries an independent chance of failure.

Current agent benchmarks show the problem clearly. Stanford’s AI Index found that success on OSWorld, which tests agents on real computer tasks, rose from roughly 12% to 66.3%. That is rapid progress, but it still leaves about one failure in every three attempts.

METR reaches a similar conclusion from another direction. Its latest task-horizon research suggests that leading agents can reach around 80% reliability on work that takes a skilled human roughly one hour. Reliability drops as the work becomes longer and less structured.

The gap can become wider outside familiar benchmark settings. SWE-Bench Mobile tested 22 model and agent combinations on work taken from a production iOS codebase. The best result was only 12% task success. Another benchmark involving real 5G software found that models could diagnose more than 91% of bugs but actually resolve only 10% to 30%.

Cheap models can look almost equal on short tests and then need much more supervision during complicated work. Customers will keep paying for better models when that extra reliability removes expensive human intervention.

Q6Are companies genuinely switching away from OpenAI and Anthropic?

Companies are moving routine workloads away from the most expensive GPT and Claude models, though most are mixing providers rather than making a clean break.

Amazon offers the clearest example. Internal documents reviewed by Business Insider projected that Alexa’s AWS costs could reach $1.7 billion during 2026. Amazon responded by routing more requests through its own models, using deterministic systems for predictable answers, improving caching and avoiding unnecessary Claude calls. The combined changes were expected to support more than four times as many customer transactions per unit of computing capacity.

Recent Wall Street Journal reporting found the same cost discipline at Zoom, Hex and Telnyx. These companies were using combinations of premium, open and cheaper models instead of automatically choosing one frontier provider for every request.

Mozilla’s chief technology officer has also described using Chinese systems such as Kimi K3 and GLM-5.2 for productivity work. His reasoning was practical: the models were slightly weaker in some areas but offered much better value.

Ramp’s latest business-spending data adds a broader view. Some of the fastest-growing AI vendors on its platform were inference services that give companies access to inexpensive open models.

Examples of companies becoming more selective about model use

Q7Will AI routers turn GPT and Claude into backup models?

AI routers will push GPT and Claude toward premium fallback roles in many applications. Their share of total tokens could fall sharply.

A router examines each request and sends it to an appropriate model. A simple classification can go to a tiny model. A normal customer question can go to a cheap general model. A difficult coding problem can still reach Claude Opus or GPT-5.6 Sol.

The technical friction keeps falling. DeepSeek supports API formats compatible with OpenAI and Anthropic, allowing developers to change the model behind an application without rebuilding the entire system. Tools such as vLLM Semantic Router can choose among OpenAI, Anthropic, Gemini, locally hosted models and other providers through the same routing layer.

Products such as Cursor already give users access to several model families inside one interface. The product owns the subscription and user relationship. The underlying model can change from one task to the next, and many users will barely notice.

Being the premium fallback can remain lucrative. The difficult 5% of requests may generate more revenue than the easy 95% because premium models cost much more. But the router decides how often those models are called and collects the data showing where each one succeeds.

Power moves toward applications, cloud platforms and orchestration companies. OpenAI and Anthropic are responding by building their own agents and work products, because owning the final experience is safer than waiting for someone else’s router to choose them.

Q8Why do large companies still pay for GPT and Claude?

Large companies still pay for GPT and Claude because the purchase includes security controls, support, predictable capacity and someone accountable when production breaks.

OpenAI offers zero-data-retention arrangements, data residency, encryption, identity controls, audit features and dedicated enterprise support. Anthropic provides similar administrative controls and distributes Claude through Amazon Web Services, Google Cloud and Microsoft’s enterprise infrastructure.

These features rarely appear on model leaderboards, yet they can decide whether a bank, hospital or government agency approves a deployment. A legal team may care more about where data is processed and how activity is logged than a five-point difference on a reasoning test.

Ramp’s latest AI Index shows how strong the established providers remain. Anthropic reached 34.4% paid adoption among businesses visible in Ramp’s data, narrowly passing OpenAI at 32.3%. Overall business adoption of paid AI products reached 50.6%.

The result is striking because cheaper open models were already widely available. Companies clearly have not chosen price alone. Anthropic’s adoption quadrupled over the preceding year, while OpenAI remained one of the two dominant paid vendors.

Open models will gain ground as cloud hosts add better governance, support and compliance. Today, OpenAI and Anthropic still have a large advantage among customers that need a complete production service rather than access to model weights.

Q9Does cheaper AI mean OpenAI and Anthropic will make less money?

Cheaper AI has not reduced OpenAI or Anthropic’s revenue so far because usage is growing much faster than prices are falling.

OpenAI says it now generates approximately $2 billion in revenue each month, compared with $1 billion per quarter at the end of 2024. Its APIs process more than 15 billion tokens per minute.

Anthropic says its annualized revenue run rate has crossed $47 billion, up from roughly $9 billion at the end of 2025. These are company-reported figures rather than audited public-company results, but the direction and scale are hard to dismiss.

The broader market is expanding alongside them. Gartner recently increased its estimate for worldwide spending on AI models and platforms from $39 billion in 2025 to $64 billion in 2026. It expects spending on generative AI models to more than double.

Lower prices make previously uneconomic uses possible. A company that once summarized 1,000 documents may process its entire archive. A coding assistant that answered occasional questions can inspect repositories continuously. Agents also generate several rounds of reasoning, tool calls and verification for one user request.

OpenAI and Anthropic can charge less for each unit of intelligence while selling vastly more units. The threat begins when customers’ efficiency efforts reduce spending faster than new uses expand it. We have seen early examples of that behavior, but no overall revenue collapse.

Q10Can exploding AI usage rescue their margins?

Exploding usage can protect revenue. It cannot rescue margins unless inference costs fall at least as quickly as prices.

Financial documents reported by The Wall Street Journal suggest that OpenAI and Anthropic expect to spend nearly $65 billion combined during 2026 on training and operating their models. That total is projected to rise to $127 billion the following year and approach $250 billion by 2029.

Separate documents reviewed by Ars Technica indicated that OpenAI’s cost of revenue rose from $2.65 billion in 2024 to $7.5 billion in 2025. Much of that increase appears connected to serving a rapidly expanding number of users.

AI agents make the problem harder. A chatbot may produce one answer. An agent can read files, search, write code, test it, inspect the result and try again. The user sees one completed task while the provider processes a pile of model calls behind it.

Caching, better chips, smaller models and more efficient architectures can lower the cost of each call. GPT-5.6 and Claude Opus 5 also use effort controls that reduce computation on simpler work. Those improvements help, although every provider can adopt similar techniques.

Revenue growth gives an incomplete picture. OpenAI and Anthropic need useful paid activity to grow faster than the combined cost of training, inference, staff and infrastructure. The boom proves that demand exists. It has not yet proved that frontier AI will produce durable software-like margins.

Q11Can ChatGPT protect OpenAI from cheaper models?

ChatGPT gives OpenAI the strongest protection in this fight because it owns a vast audience that rarely compares token prices.

OpenAI says ChatGPT is approaching one billion weekly active users. More than nine million paying business users rely on it for work, and over one million organizations purchase OpenAI business products.

That audience gives OpenAI several ways to respond to commoditization. It can route easy questions through inexpensive internal models, reserve GPT-5.6 Sol for demanding tasks and hide the calculation inside a monthly subscription. Most users will keep calling the product ChatGPT regardless of which model handled the request.

OpenAI also has direct distribution for search, voice, image generation, coding and workplace agents. Codex has reached three million weekly users, while enterprise activity already contributes more than 40% of company revenue.

The company is developing revenue streams around advertising, commerce and compute too. Each one reduces its dependence on charging developers a high price for API tokens.

ChatGPT’s audience could still move if another assistant becomes consistently better. Google, Anthropic and specialist applications control important distribution channels of their own. For now, OpenAI has a serious cushion: a cheap rival must win people’s habits and workflows, not merely publish a lower price.

Q12Can Claude Code protect Anthropic from cheaper coding models?

Claude Code currently gives Anthropic a real moat because developers pay for completed software work through a familiar workflow.

Anthropic says Claude Code reached an annualized revenue run rate above $2.5 billion less than a year after its public launch. Weekly users doubled from the beginning of 2026, while business subscriptions quadrupled. Enterprise customers now generate more than half of Claude Code revenue.

An external analysis cited by Anthropic estimated that Claude Code was producing around 4% of public GitHub commits worldwide. That figure cannot capture private repositories and is best treated as directional, but the growth suggests Claude Code has moved far beyond a niche developer tool.

The product’s value comes from exploring repositories, editing files, executing commands, checking results and maintaining context over longer jobs. A cheaper model must work reliably inside that full process. Matching Claude on isolated code generation would leave much of the product advantage untouched.

Developers are demanding and quick to switch. They already use multiple agents, and coding output is easier to test than many other forms of knowledge work. Anthropic could lose usage quickly if a cheaper system completes the same tasks with similar reliability.

Right now, Claude Code gives Anthropic a direct customer relationship and a reason to keep paying beyond access to the Claude API. It is the company’s clearest answer to model commoditization.

Q13When is paying 20 times more for GPT or Claude actually worth it?

Paying 20 times more for GPT or Claude makes sense when the better model saves more human time than the small dollar difference between two individual tasks.

Token-price comparisons can sound more dramatic than the final bill. Generating 10,000 output tokens costs about $0.009 with DeepSeek V4 Pro and $0.25 with Claude Opus 5. The premium is roughly $0.24.

If Claude avoids two minutes of review by an employee earning $60 per hour, it has already paid for itself. A prevented software bug, faulty contract clause or incorrect financial calculation could be worth much more.

The calculation changes for high-volume, low-value work. Generating one million output tokens raises the difference to more than $24. A company processing billions of tokens through customer support or document extraction can save millions by moving routine requests to a cheaper provider.

Frontier models should retain pricing power where quality affects labor, risk or revenue. Cheap models should dominate tasks that are easy to check, repeat frequently and carry a low cost of failure.

How the cost difference changes with output volume

Q14Could compute bills hurt them more than cheap models?

Compute commitments could hurt OpenAI and Anthropic more than cheap rivals if the companies lock in huge capacity while model prices keep falling.

The Wall Street Journal recently reported that OpenAI’s planned cloud and computing spending through 2030 had climbed to about $750 billion. OpenAI is also developing a $20 billion data center in Georgia with 3.2 gigawatts of contracted power, alongside large agreements with Microsoft, Oracle, Amazon, Nvidia, AMD and Broadcom.

Anthropic has secured up to five gigawatts of Amazon capacity for training and serving Claude. It has also signed for several gigawatts of future Google and Broadcom capacity and is building additional infrastructure with Fluidstack.

These commitments can strengthen both companies. Large reserved fleets may lower unit costs, prevent capacity shortages and allow models to serve millions of users without slowing down.

The danger appears when the selling price of intelligence falls faster than the cost of owning or renting that infrastructure. Data centers, energy agreements and chips remain expensive even when customers switch their easy workloads to DeepSeek or an internally hosted model.

Both companies have raised enormous amounts of capital, giving them more room than a normal startup. Capital buys time. It does not repair weak unit economics. Their survival could eventually depend less on who builds the smartest model and more on who keeps expensive hardware busy with valuable paid work.

Q15Could cheap models seriously weaken them without killing them?

Cheap models are very likely to weaken both companies even if neither comes close to failing.

Routine API work will become harder to price at a premium. Customers can test several providers quickly, and procurement teams now have enough credible alternatives to negotiate discounts.

OpenAI and Anthropic will also lose some control to routers, cloud platforms and applications. Those intermediaries can decide when GPT or Claude is worth the extra cost and replace either model when performance changes.

Both companies must consequently spend more on products, integrations, enterprise administration and distribution. They are entering markets already occupied by Microsoft, Google, Amazon, Salesforce, Atlassian, Adobe, Cursor and dozens of specialist software companies.

OpenAI enters this period with stronger consumer distribution. Anthropic has built a particularly strong position in coding and enterprise work, though those customers are highly capable of testing alternatives and watching costs.

The likely outcome is a less comfortable business than investors once imagined. OpenAI and Anthropic may remain huge, influential and fast-growing while earning lower margins, sharing customers with many models and spending heavily to defend their products.

Cheap models can remove much of the economic reward attached to being the best general-purpose model for a few months. That alone would reshape both companies.

Q16What would have to go wrong for OpenAI or Anthropic to actually fail?

OpenAI or Anthropic would face a real survival problem only if four bad developments arrived together.

First, cheaper models would need to match the frontier on long-running professional work, including the reliability that remains missing from current agents. Second, large companies would need to become comfortable buying governance, support and security from open-model hosts.

Third, OpenAI and Anthropic would have to lose their product advantages. ChatGPT users would move elsewhere, while developers would replace Claude Code with cheaper agents. Finally, revenue and funding would need to weaken while both companies remained committed to enormous infrastructure bills.

We see movement toward the first two conditions. The third is far from established, and both companies still have exceptional access to capital.

What would need to happen before cheap models became an existential threat

Q17Will cheap AI models kill OpenAI and Anthropic?

Cheap AI models will not kill OpenAI or Anthropic.

They are already destroying the easy version of their business model, where the leading laboratories could release the smartest model, charge a large premium and assume customers would stay.

Today, cheap and open models can complete a large share of routine work. Companies are routing tasks more carefully, Chinese models are reaching the frontier group and token prices continue falling. API margins and customer loyalty will face sustained pressure.

OpenAI has the better protection through ChatGPT’s enormous audience, expanding workplace use and several possible revenue streams. Anthropic has built a powerful position through Claude Code and enterprise adoption, but its business remains more exposed to developers and companies that can compare models automatically.

The harder risk sits in their cost structures. Both companies must continue funding frontier research and vast computing infrastructure while customers learn to consume intelligence more efficiently. Strong revenue growth gives them room to adapt, though revenue alone does not guarantee the economics will eventually work.

Our judgment is direct: the prediction of corporate death is exaggerated. The business-model disruption is already happening. OpenAI and Anthropic can survive cheap AI models, but only by owning products, workflows and customer relationships that remain valuable after the underlying models become widely available and inexpensive.