Three separate AI models are now marketed under the name 'o3' or nearly identical variants, causing widespread mix-ups in benchmarks, pricing comparisons, and product evaluations. OpenAI's o3 is a reasoning language model released in April 2025. Kuaishou's Kling O3 is a video generation tool announced in February 2026. And MiniMax's H3 is frequently called 'Hailuo 03' by the community, even though the company never uses that name. The confusion has real consequences: coding benchmarks are being attributed to video models, and procurement teams are receiving quotes based on entirely different pricing structures.

Why Are These Models Getting Confused?

The naming collision stems from a combination of vendor naming changes, community shorthand, and search engine algorithms that reward similar-sounding names. MiniMax's own video product line has cycled through four different naming schemes since February 2025. The company originally called its video models "T2V-01-Director" and "I2V-01-Director," then switched to "MiniMax-Hailuo-02," then "MiniMax-Hailuo-2.3," and finally settled on "MiniMax H3" in July 2026. When MiniMax followed that pattern, the community reasonably expected the next version to be called "Hailuo-03." That prediction collided directly with Kuaishou's actual product naming.

OpenAI's o3 was named with a lowercase "o" specifically to avoid a trademark clash with the O2 mobile network. Kuaishou, meanwhile, does not officially call its video model "Kling O3" at all; the company's own announcement from February 2026 refers to it as "Video 3.0 Omni." The "O3" shorthand is a back-formation created by third-party hosting platforms and the broader AI community, even though it is defensible given that Kuaishou does use O-numbering elsewhere in its product line.

What Are the Real-World Consequences of This Confusion?

The naming mix-up has created measurable problems in three critical areas: benchmark attribution, product procurement, and search engine results. MiniMax's M3 is a text and coding model released on June 1, 2026, with a 1,000,000-token context window. It scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench, and 74.2% on MCP Atlas. These are software engineering and tool-use benchmarks that measure coding ability. However, because MiniMax's model card tags M3 with the word "video" (referring to its ability to understand video input, not generate it), search results routinely show these coding benchmarks next to "H3" in video model comparisons. A video generation model cannot be meaningfully scored on coding benchmarks, yet procurement teams and researchers regularly encounter these misattributed numbers.

Pricing confusion creates even more serious problems. MiniMax M3 is billed per token processed, while MiniMax H3 is billed per second of generated video. A quote built on the wrong model is not slightly inaccurate; it is priced on a completely different unit. An enterprise comparing costs could easily receive a quote for M3's per-token pricing and apply it to H3's per-second model, resulting in a cost estimate that bears no relationship to actual expenses.

How to Distinguish Between the Three Models

• OpenAI o3: A reasoning language model released in April 2025 that processes text and code. It is the only one of the three whose vendor actually publishes it under the name "o3." It was called o3 rather than o2 to avoid trademark issues with the O2 mobile network.
• Kuaishou Kling O3 (officially Video 3.0 Omni): A video generation model announced February 5, 2026, by Kuaishou Technology, a Hong Kong-listed Chinese social platform. It generates video clips up to 15 seconds long with multilingual audio support and accent variation. The company does not officially use the name "O3."
• MiniMax H3 (sometimes called Hailuo 03 by the community): A video model released July 31, 2026, that generates 4 to 15 second clips at up to 2K resolution and 24 frames per second with native stereo audio. MiniMax never uses the name "Hailuo 03" or "Hailuo 3.0" in any official documentation. The model is accessed via the API string "MiniMax-H3."

The simplest rule for MiniMax products: at MiniMax, the letter "M" indicates a text and coding model, while the letter "H" indicates a video model. One letter separates a coding tool from a video generator, and the two operate behind different API endpoints on the same host.

What Distinguishes MiniMax M3 and H3 From Each Other?

MiniMax M3 and MiniMax H3 are fundamentally different products despite their similar names and shared company origin. M3 is a language model for text, code, and long-context work. It accepts a 1,000,000-token context window, meaning it can process roughly 1 million words at once. It can read images and video as input to understand them, but it does not generate video. M3 was released on June 1, 2026, and uses synchronous chat completion calls via the endpoint "/v1/chat/completions" or "/anthropic/v1/messages." It is billed per token.

H3 is a video generation system released on July 31, 2026, that generates video with synchronized audio. It accepts a 7,000-character prompt limit and can take video as a style or motion reference input. H3 operates asynchronously, meaning users create a task, receive a task ID, and then poll that ID to download the finished video. It is billed per second of generated video. H3 is described by MiniMax as a system of three parts: H3-Context-IR (which interprets complex multimodal prompts), H3-Base (which generates at 768p resolution), and H3-Regenerate-2K (which upscales output to 2K). The open-weight release covers only the base generation variants.

How Should Researchers and Procurement Teams Verify Model Identity?

• Check the API string: MiniMax M3 uses the API identifier "MiniMax-M3," while MiniMax H3 uses "MiniMax-H3." OpenAI's o3 is identified as "o3" (lowercase). Kuaishou's model is "Video 3.0 Omni" in official documentation.
• Verify the vendor's official documentation: All claims should be cross-referenced against vendor-controlled surfaces such as official documentation, release notes, license files, or company press releases, not aggregator websites. MiniMax's official surfaces never use the names "Hailuo 3.0," "Hailuo 03," or "Hailuo-03." Kuaishou's official announcement does not use "Kling O3."
• Match benchmarks to modality: Coding benchmarks like SWE-Bench Pro, Terminal-Bench, and MCP Atlas apply only to language models, not video models. If a page attributes these scores to a video model, the wires have been crossed. Video models are evaluated using blind side-by-side human preference ratings expressed as Elo scores.
• Confirm the pricing unit: M3 is priced per token, H3 is priced per second of video, and OpenAI's o3 uses its own pricing structure. A quote that does not specify the unit is incomplete and may be based on the wrong model entirely.

The source material was verified against primary vendor sources on August 4, 2026, and the verification methods described above remain valid even as vendor naming conventions continue to evolve.