OpenAI’s New GPT-4o Mini Is Giving Competitors A Run For Less Money, & More

Want more research from the ARK Team? Have feedback on our publications? Click here to help inform our content creation.

1. OpenAI’s New GPT-4o Mini Is Giving Competitors A Run For Less Money

By: Jozef Soja

Last week, OpenAI released GPT-4o mini, a smaller, more cost-efficient version of its flagship large language model, GPT-4o. On the HumanEval benchmark for python coding, GPT-4o mini scored 87.2%, compared to Anthropic's Claude 3 Opus’s 84.9%[1] and Google's Gemini 1.5 Pro’s 82.6%.[2] The larger version of GPT-4o scores 90.2%.[3]

GPT-4o mini’s performance is less expensive than competitors. Based on roughly three input tokens for every output token produced, GPT-4o mini costs only $0.26 per million tokens inferenced,[4] much less expensive than flagship models, some of which perform less well. GPT-4o and Gemini 1.5 Pro, for example, cost $7.50 and $5.25,[5] respectively, while small models like Gemini 1.5 Flash and older models like GPT-3.5 Turbo cost $0.53 and $0.75 per million tokens, respectively. OpenAI plans to replace GPT-3.5 Turbo with GPT-4o mini as the base model that will power free versions of ChatGPT.

As developers incorporate AI into more products at scale, maximizing performance could take a backseat to efficiency and cost-savings. GPT-4o mini demonstrates that OpenAI delivers not only best-in-class performance at the high end, but also models at costs low enough to justify wide-scale deployment. The impact of cost savings could be profound, as “good-enough” models at ~5% the cost of flagship models could encourage the deployment of agentic AI across workflows and organizations.

2. Interest In Polymarket Is Surging As The US Presidential Election Heats Up

By: Lorenzo Valente

Polymarket is a platform for decentralized predictions that allows users to bet on sports, geopolitics, and pop culture. Users can buy and sell shares associated with specific outcomes, their prices based on supply and demand. Priced from $0.00 to $1.00, a $0.50 share suggests that the probability of an event is 50%. If the event takes place, the share price will double to $1.00.

Founded by Shayne Coplan in 2018, Polymarket has gone parabolic in response to the US presidential race this year. Now that betting on the November election[6] has approximated $290 million, both daily volume and number of new active traders have hit all-time highs of $20 million and 30,000,[7] respectively. The leading global betting marketplace for the election,[8] Polymarket is powered by smart contracts on the Polygon Ethereum layer 2 network and has grown tremendously during the past three months, as shown below.

Prediction markets could have real staying power, thanks to platforms that not only are tamper-proof and transparent but also provide liquid bets with real-time consensus data, not to mention the “wisdom of crowds”[10] that provide highly accurate odds. As distrust of legacy media increases globally, prediction markets could become “information credibility” barometers that gauge public sentiment and surface truth.

3. TempusAI’s Technologies And Collaborations Are Paving The Way For Value-Based Healthcare

By: Nemo Marjanovic, PhD

Recently, TempusAI released its new AI-enabled Immune Profile Score (IPS), an algorithmic multimodal pan-cancer laboratory-developed test (LDT) that evaluates immunotherapy-related biomarkers to classify patients as either IPS-Low or IPS-High. Designed to support patient stratification, the test should inform treatment decisions by identifying which patients will benefit from immunotherapy. TempusAI also is collaborating with the Cleveland Clinic to develop other algorithmic tests leveraging comprehensive clinical and genomic data.

Interestingly, regulatory bodies are supporting these advanced approaches. The American Medical Association (AMA) recently granted a PLA (Proprietary Laboratory Analyses) code to Tempus’ PurISTSM algorithmic test, an important step toward reimbursement, underscoring the clinical utility of AI-driven diagnostics. Moreover, the FDA has granted 510(k) clearance[11] for Tempus’ ECG-AF, an AI-based algorithm that identifies patients at increased risk of atrial fibrillation. Clearly, the FDA recognizes the transformative potential of AI/ML technologies in healthcare.

In our view, TempusAI is leading the charge in algorithmic testing, working closely with the providers and regulatory bodies essential to achieving payor acceptance and reimbursement. In the short term, algorithmic testing should improve efficiency and patient care significantly. Longer term, it could catalyze a paradigm shift toward value-based healthcare.

Want more research from the ARK Team? Have feedback on our publications? Click here to help inform our content creation.

[1] Anthropic. 2024. “Introducing the next generation of Claude.”

[2] Gemini Team Google. 2024. “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.” arXiv.

[3] OpenAI. 2024. “Hello GPT-4o.”

[4] OpenAI. 2024. “Pricing.”

[5] Google AI for Developers. “Gemini API Pricing.”

[6] Polymarket. 2024. “Presidential Election Winner 2024.”

[7] Dune Analytics. 2024. “Polymarket Open Interest.” “Polymarket Monthly New Accounts.”

[8] Merchant, M. 2024. “Polymarket Is World's Largest 2024 Presidential Election Prediction Pool: Bernstein.” Decrypt.

[9] Dune Analytics. 2024. “Polymarket Open Interest.” “Polymarket Monthly New Accounts.”

[10] Halton, C. 2022. “Wisdom of Crowds: Definition, Theory, Examples.” Investopedia.

[11] According to the U.S. Food and Drug Administration, “The 510(k) clearance process involves a comprehensive review of safety and performance data for the device, which may include scientific, non-clinical, and clinical data, as appropriate, to determine if a new device is substantially equivalent to a device that is already on the market (that is, a predicate device).” See U.S. Food and Drug Administration. 2024. “Medical Device Safety and the 510(k) Clearance.” Process.