AIFEATURE

What Is Multimodal AI? GPT-4o and Gemini Show Two Paths to Unified Understanding

N
NathanTechnology Editor · Technical Lead
Published · Updated
Multimodal AI systems process text, code, audio, image and video within one model instead of stitching together separate tools for each format. OpenAI's GPT-4o, announced May 13, 2024, trains a single neural network end-to-end across text, vision and audio, replacing a three-model voice pipeline that lost tone and speaker information. Google's Gemini, announced December 6, 2023, was pre-trained natively across modalities from the start. Both companies say the approach aids complex reasoning, though OpenAI initially limited GPT-4o's audio and video features to a small group of trusted API partners.

What Is Multimodal AI, and How Does It Differ From Traditional Pipelines?

Multimodal AI refers to a system built to generalize across and combine different types of information — including text, code, audio, image and video — rather than handling each format with a separate toolCITE:E1.

Google says it built Gemini "from the ground up to be multimodal," meaning the model can generalize and seamlessly understand, operate across, and combine text, code, audio, image and videoCITE:E1. Google made that design goal explicit in its official announcement dated December 6, 2023CITE:E1.

The contrast with a non-unified approach is concrete: before GPT-4o, OpenAI's Voice Mode ran as a pipeline of three separate models — one transcribed audio to text, GPT-3.5 or GPT-4 processed that text and produced text output, and a third model converted the text back into audioCITE:E2. OpenAI states that because the main source of intelligence in that pipeline only ever read text, it could not directly observe tone, multiple speakers, or background noise, and could not output laughter, singing, or express emotionCITE:E2. That gap is why a text-only, pipeline-based setup loses non-textual information that a model designed to be multimodal from the outset is meant to retain.

Why Build One End-to-End Model Instead of Chaining Multiple Models?

OpenAI trained GPT-4o as a single new model end-to-end across text, vision and audio, so one neural network processes every input and outputCITE:E5.

"With GPT-4o, we trained a single new model end-to-end across text, vision, and audio, meaning that all inputs and outputs are processed by the same neural network," OpenAI said in its May 13, 2024 announcementCITE:E5. This directly answers the information loss described in the earlier three-model voice pipeline: collapsing transcription, reasoning and synthesis into one system removes the intermediate step where tone and speaker information were previously discardedCITE:E2CITE:E5.

How Do GPT-4o and Gemini Represent Different Multimodal Approaches?

GPT-4o accepts any combination of text, audio, image and video as input and generates any combination of text, audio and image as output, while Gemini was pre-trained natively across modalities from the start and later fine-tunedCITE:E3CITE:E4.

ModelCompanyInput modalitiesOutput modalitiesDesign approachAnnounced
GPT-4oOpenAIText, audio, image, videoText, audio, imageSingle model trained end-to-end across text, vision, audioCITE:E52024-05-13CITE:E3
GeminiGoogleText, code, audio, image, videoNot specified in the cited evidencePre-trained natively on different modalities from the start, then fine-tuned with additional multimodal dataCITE:E42023-12-06CITE:E1

Notably, GPT-4o's output set does not include video generation, only text, audio and imageCITE:E3. Google describes its own approach in similar terms of built-in design: "We designed Gemini to be natively multimodal, pre-trained from the start on different modalities. Then we fine-tuned it with additional multimodal data to further refine its effectiveness," the company saidCITE:E4.

What Practical Advantage Does Multimodal Reasoning Provide?

Google says Gemini's multimodal reasoning capabilities can help make sense of complex written and visual informationCITE:E6.

In its December 6, 2023 announcement, Google stated: "Gemini 1.0's sophisticated multimodal reasoning capabilities can help make sense of complex written and visual information"CITE:E6. This framing links back to the model's native multimodal pretraining described above — the ability to combine text and visual inputs is presented as a direct outcome of that design choiceCITE:E4CITE:E6.

Why Is There a Gap Between Announced Capabilities and What's Actually Available?

OpenAI said at GPT-4o's launch that its new audio and video capabilities would go first to a small group of trusted partners in the API, not to all users immediatelyCITE:E7.

"We plan to launch support for GPT-4o's new audio and video capabilities to a small group of trusted partners in the API in the coming weeks," OpenAI said in the same May 13, 2024 announcement that described the end-to-end unified modelCITE:E7. The statement was published alongside, not separately from, the architecture claim, meaning the single-model design and the limited rollout were disclosed togetherCITE:E5CITE:E7.

What This Means

Across the two companies' own statements, a consistent pattern emerges: both frame multimodality as something built into the model from the start rather than added afterward — Google through "natively multimodal" pretraining and OpenAI through a "single model end-to-end" trained across text, vision and audioCITE:E1CITE:E4CITE:E5. That architectural claim is distinct from availability: OpenAI's own announcement that introduced the unified model also disclosed that its audio and video capabilities would initially reach only a small group of trusted API partners, not the full user baseCITE:E7. Read together, the evidence supports a technical claim about how the models are built, but not a claim about how broadly their newest capabilities were accessible at launch.

📊 Evidence

FAQ

What Is Multimodal AI, and How Does It Differ From Traditional Pipelines?

Multimodal AI refers to a system built to generalize across and combine different types of information — including text, code, audio, image and video — rather t…

Why Build One End-to-End Model Instead of Chaining Multiple Models?

OpenAI trained GPT-4o as a single new model end-to-end across text, vision and audio, so one neural network processes every input and outputCITE:E5.

How Do GPT-4o and Gemini Represent Different Multimodal Approaches?

GPT-4o accepts any combination of text, audio, image and video as input and generates any combination of text, audio and image as output, while Gemini was pre-t…

What Practical Advantage Does Multimodal Reasoning Provide?

Google says Gemini's multimodal reasoning capabilities can help make sense of complex written and visual informationCITE:E6.

📎 Sources

  1. blog.google
  2. openai.com

Related data

Author's TakeNathan

The concrete technical fact here is architectural, not conversational: GPT-4o's shift to a single end-to-end network removes the three-model voice pipeline that previously stripped out tone, multiple speakers, and background noise before the reasoning model ever saw them. That is a specific, disclosed design change, not a vague upgrade claim. Gemini's native pretraining across modalities describes the same underlying principle — build the model to handle multiple formats from the start rather than bolt them together afterward — even though Google and OpenAI use different language for it. The detail worth tracking going forward is OpenAI's own rollout caveat: audio and video access was limited to a small group of trusted API partners at launch, so the real test of the unified-model approach is how broadly those capabilities move from an announced architecture to a generally available feature.

N
NathanTechnology Editor · Technical Lead

Related

BRIEF

GreenTrans Unveils GT5X, GT3X Quadruped Robots, Targets 100% Taiwan-Made Content by 2027

GreenTrans (綠捷), the robotics subsidiary of China Motor (中華車), unveiled quadruped robots GT5X and GT3X at SEMICON Taiwan 2026, targeting 100% Taiwan-made content by 2027. The robots combine an in-house-designed control unit and battery management system, NVIDIA's Jetson Orin and Isaac Lab platforms, and a new LFP battery developed with Formosa Smart Energy (台塑新智能), while GreenTrans's inspection robots are already deployed in semiconductor fabs.

EffectStory 編輯部 ·
BRIEF

Nvidia Confirms $12.93 Billion Acquisition of Hugging Face

Nvidia confirmed on September 3, 2026 that it agreed to buy Hugging Face for $12.93 billion, exactly $12,930,300,000, gaining the open-source AI hosting platform used by over 18 million developers. CEO Jensen Huang pledged the platform will stay open, with no Nvidia compute required to build on or deploy through it.

EffectStory 編輯部 ·
BRIEF

NVIDIA to Subscribe US$3.5 Billion of MediaTek's Record US$3.9 Billion Convertible Bond

NVIDIA will subscribe US$3.5 billion of MediaTek's US$3.9 billion offshore convertible bond offering, the largest such issuance in Taiwan's capital market history, deepening cooperation in AI infrastructure, edge AI computing, and automotive platforms while marking NVIDIA's first major investment in a Taiwanese company.

EffectStory 編輯部 ·