OpenAI Has Model Updates, Lower Prices, and New Requests from Anti-trust Regulators
It's a very different OpenAI than 12 months ago
OpenAI announced several new updates this week regarding its large language models (LLM) and API services. This is a sign of the maturing of the OpenAI technology stack, which supports production features and cost management instead of strictly focusing on model performance.
In addition, OpenAI has been brought into a new inquiry from the U.S. Federal Trade Commission (FTC). The U.S. agency is responsible for anti-trust monitoring and enforcement. OpenAI’s close association with Microsoft has ensnared it in a regulatory investigation that will surely begin to expose confidential messages from within the company and between it and Microsoft.
New Models and Services
OpenAI unilaterally reduced API access prices to its models twice in 2023. It is starting off 2024 with a similar strategy.
The company positions these price reductions as passing along efficiency cost savings to customers. However, there are other benefits. Lower prices encourage more system usage by budget-constrained software partners and enterprise organizations. They also make it harder for competitors. In addition to closing the performance gap with OpenAI, they must also contend with a cost gap.
Model updates also include performance gains. The API monitoring dashboard will be a welcome addition for any organization managing multiple models and use cases. Here is a quick summary.
GPT-3.5 Turbo price cuts. The February launch of gpt-3.5-turbo-0125 will have a 50% price reduction for token inputs and 25% for token outputs over the current model. It will also include some accuracy improvements and a fix for UTF-8 encoding bug.
GPT-4 Turbo performance updates. The new gpt-4-1025-preview “completes tasks like code generation more thoroughly than the previous preview model and is intended to reduce cases of “laziness” where the model doesn’t complete a task.”
There are two new text-embedding models. Availability for embedding-ada-002 will continue and now be complemented by text-embedding-3-small and text-embedding-3-large. The “small” version is actually larger than the “ada” model but “substantially more efficient,” according to OpenAI. The result is better performance at one-fifth the price. The “large” model offers even better performance at a 30% higher price than “ada.”
While GPT-3 and GPT-4 are optimized for text generation, text embedding models are typically optimized for “text search, code search, and sentence similarity tasks.”
The moderation models are also receiving updates. “The free Moderation API allows developers to identify potentially harmful text. As part of our ongoing safety work, we are releasing text-moderation-007, our most robust moderation model to date.”
New observation and management tools for controlling API use. This fills a gap for many users who typically must use third-party tools today to monitor that token usage. “The usage dashboard and usage export function now expose metrics on an API key level after turning on tracking. This makes it simple to view usage on a per feature, team, product, or project level, simply by having separate API keys for each. In addition, users can now restrict keys for use in a particular function.
Regulatory Inquiry
You know your company has arrived in the big leagues when anti-trust regulators start asking for materials as part of an investigation. Reuters reported Friday:
The U.S. Federal Trade Commission said on Thursday it had ordered OpenAI, Microsoft, Alphabet, Amazon, and Anthropic to provide information on recent investments and partnerships involving generative AI companies and cloud service providers…
Deals among the small number of powerful players and Big Tech have raised antitrust concerns.
The FTC orders will allow the agency to scrutinize the inner workings of deals between Microsoft, Google and Amazon and AI providers to help the antitrust and consumer protection agency understand how these deals have affected competition.
…
The extensive document request seeks details on how the partnerships with Big Tech influence strategy and "decisions around the pricing of products and services; decisions around the granting of access to products and services; and decisions around personnel."
The agency, which also wants information on exclusivity agreements, said the companies have 45 days to respond to the orders.
The FTC regularly monitors the activities of tech giants. Given its modest size compared to the cloud hyperscalers, OpenAI would normally escape this type of scrutiny. However, the FTC is very interested in the development of generative AI, and OpenAI’s financial deal with Microsoft likely created a reasonable pretext to include it in the discovery.
Including Anthropic would be even harder for regulators to justify under normal circumstances. However, the large investments by Google and Amazon, along with the hosting deals through Google Cloud and AWS, made it a ripe target. The FTC will want to know what terms were associated with the multi-billion dollar checks from Anthropic’s cloud partners.
This process is early and will take a long time to play out. There is no implied action by the FTC. It is the start of an investigation that may simply provide the regulator with information for future inquiries or actions. The FTC is starting its preparation now, so it is ready if it wants to take action in the future.
Generative AI Competition
Many of OpenAI’s and Anthropic’s competitors will welcome this activity. Synthedia was among the first analysts (maybe the first) to identify the LLM wars as proxy battles for the broader cloud wars.
Foundation model developers who have not been “anointed” by Azure, AWS, or Google Cloud worry that they will have difficulty competing against models developed internally by the leading cloud providers and their favored investments, OpenAI and Anthropic. The situation represents a competitive disadvantage for companies such as AI21 Labs, Aleph Alpha, Cohere, Databricks, Mistral, and TII.
The tight product and financial alliances also have the potential to impact the adoption of foundation models from other tech giants—notably Meta and NVIDIA. In some ways, it is surprising that these companies, with such large generative AI footprints today, were excluded from the latest inquiry. However, they are not cloud-hosting giants. That may be the common thread behind the FTC’s move.
New models, new tools, new pricing, and new requests from regulators. This is a much different open AI than one year ago.