# Anthropic and OpenAI: two strategies for the same race

**Anthropic OpenAI competition** began with a shared origin and diverged into two models of building AI. Anthropic was founded in 2021 by former OpenAI researchers, including Dario Amodei, and set out to develop systems capable of automating large portions of the economy while placing safety and alignment at the centre of the work. OpenAI scaled faster and reached a broader consumer audience first, and the two companies now compete for enterprise customers, funding and talent.

## Two approaches to the same problem

Anthropic's technical signature is Constitutional AI, a training method that gives a model a set of guiding principles and uses them to steer its responses. The company positions Claude as less likely to produce harmful output and easier to direct in conversation. OpenAI combined scale with a broad product suite and a large partner ecosystem, building its models into consumer and business tools through Microsoft. Both approaches aim at capable assistants; they differ in how openly they describe the training process and how they weigh release against control.

## The economics of frontier models

The race is expensive. Anthropic described a frontier model, provisionally named Claude-Next, that it expected to require on the order of 10^25 floating-point operations and about a billion dollars over 18 months to complete, with training on clusters of thousands of GPUs. The company raised capital in successive rounds and secured strategic backing from Google, which made Google Cloud its preferred provider. OpenAI had already secured a multibillion-dollar investment from Microsoft, which gave it compute and distribution. Access to computing power, not only to research talent, became the deciding factor.

## What the competition means for buyers

Competition between two large vendors benefits customers in some ways and creates risk in others. It accelerates capability and lowers prices, and it also produces lock-in when applications are built against a single proprietary interface. A company that makes a strategic choice today should verify data ownership, export formats, migration support and the security posture of the vendor, and it should prefer architectures that separate business logic from the specific model. The competition will continue to change which model leads on benchmarks. The terms of the relationship with the vendor change far less often.

The competition is not only commercial: it decides [the transparency debate around open models](/en/ecorner/2023/ai-transparency-democracy.html) and which systems can be audited in Europe.

## Frequently asked questions

**What is Constitutional AI?**
A training approach that uses a set of written principles to guide model behaviour and reduce harmful outputs, combined with human feedback.

**Why did Anthropic's founders leave OpenAI?**
They cited differences over direction and safety priorities, and they built a company focused on alignment research alongside commercial products.

**Does vendor choice lock a company in?**
It can, when integrations and data formats are proprietary. Open standards and exportable data reduce the risk.

**How should a business compare AI vendors?**
On documentation, data handling, security, audit rights, migration terms and total cost, alongside the model's quality for the specific task.

## Sources

Anthropic — company, Claude and research on Constitutional AI — https://www.anthropic.com/
Anthropic — Constitutional AI: harmlessness from AI feedback — https://arxiv.org/abs/2212.08073
OpenAI — research and company updates — https://openai.com/
Financial Times — Anthropic, Google Cloud and the AI partnership — https://www.ft.com/content/583ead66-467c-4bd5-84d0-ed5df7b5bf9c