What Changed in GPT-4o, Gemini, DeepSeek and MCP?

TL;DR

GPT-4o improved substantially in coding, instruction following, creativity, and image generation, while Gemini 2.5 Pro emerged as a fast thinking model with strong coding performance and a million-token context window. DeepSeek V3 gained stronger math and coding capabilities under an MIT license, and MCP adoption by Zapier, OpenAI, and Microsoft signaled growing support for a common agent-tool standard.

Transcript

You want to hear something crazy? GPT40 got some massive updates. In fact, it is not only the best at image generation now, but it's also the number one non-reasoning coder on the market. And this is verified by independent benchmarks. Look at this artificial analysis. Today's GPT40 update is actually big. It leapfrogs Claude 3.7 sonnet non-reasoni... Read More

Key Insights

• GPT-4o is presented as the leading non-reasoning coding model in the cited Artificial Analysis comparison, rising from an intelligence index score of 41 for its November 2024 version to 50 after the update, just behind the recent DeepSeek V3 model.
• GPT-4o's update improves detailed instruction following, especially for prompts containing multiple requests, while also strengthening complex technical and coding problem solving, intuition, and creativity. The model reportedly uses fewer emojis, and paid users received immediate access while free access was scheduled over subsequent weeks.
• GPT-4o image generation demand exceeded OpenAI's expectations, leading the company to introduce rate limits. The presenter also observed that GPT-4o had become very slow for ordinary queries, highlighting response speed as an important factor when selecting a general-purpose or agentic AI model.
• Gemini 2.5 Pro is a thinking model that the presenter considers the best coding model he has used, particularly because it combines strong capability with fast responses. Its million-token context window may help it understand entire codebases, and it became available through both Windsurf and Cursor.
• DeepSeek V3 is an updated checkpoint rather than a completely new model, with strong performance in coding, mathematics, and logic. It remains fast, downloadable, and open weight, while its switch to the permissive MIT license gives users broad freedom to experiment with the model.
• ARC-AGI-2 is designed around abstraction tasks that humans can solve easily but current AI systems find difficult. The human panel scored 100 percent on ARC-AGI-2, while o3 Low achieved 4 percent at a reported cost of $200 per task, revealing a substantial capability and efficiency gap.
• MCP is becoming a widely adopted method for connecting agents to external tools. Zapier announced an MCP offering covering access to 10,000 tools, OpenAI added MCP support to its agents API, and Microsoft introduced the protocol in Copilot Studio after Anthropic established the standard.
• New image and visual reasoning models are expanding beyond GPT-4o. Reve Image 1.0 ranked highly in quality based on 100,000 user votes, Ideogram 3.0 emphasized controls such as remixing, upscaling, and style preferences, and Qwen's QVQ-Max combined vision with evidence-based reasoning.

Questions & Answers

Q: What improved in the latest GPT-4o update?

The updated GPT-4o improved detailed instruction following, particularly when a prompt includes multiple requests. It also became better at complex technical and coding problems, intuition, creativity, and image generation, while reportedly using fewer emojis. Artificial Analysis placed it ahead of Claude 3.7 Sonnet in non-reasoning mode and Gemini 2.0 Flash in its intelligence comparison, making it the leading non-reasoning coding model discussed.

Q: Why did OpenAI continue improving GPT-4o?

The transcript connects continued GPT-4o development to limited computing capacity and rising demand. OpenAI reportedly could not obtain enough GPUs to fine-tune and improve GPT-4.5 at the desired scale. The presenter frames this through Jevons paradox: when computing or AI capabilities become cheaper and more accessible, users demand more of them, so overall compute requirements continue increasing rather than declining.

Q: How does Gemini 2.5 Pro help with coding projects?

Gemini 2.5 Pro combines thinking capabilities, fast responses, strong coding performance, and a million-token context window. The presenter expects that large context capacity to help the model understand a codebase as a whole, particularly for projects containing fewer than a million tokens of code. Its availability in Windsurf and Cursor also makes it practical to test within agentic and coding workflows.

Q: What changed in the new DeepSeek V3 checkpoint?

The new DeepSeek V3 release is a checkpoint within the existing V3 series, not a completely new model. It performs strongly in coding, mathematics, and logic while operating as a fast, non-thinking model. It is also open weight, downloadable, and licensed under the permissive MIT license. However, its large size may make local operation difficult for users without sufficient computing resources.

Q: What does ARC-AGI-2 measure?

ARC-AGI-2 measures whether a system can infer abstract patterns from examples and apply those patterns to unfamiliar cases. Tasks may present grids with colors, shapes, separators, and transformations, requiring the solver to identify the underlying relationship and reproduce it. The benchmark is intentionally easy for humans but difficult for AI, making the gap between human and model performance central to its purpose.

Q: How did humans and AI models perform on ARC-AGI-2?

The human panel scored 100 percent on ARC-AGI-2, compared with 98 percent on ARC-AGI-1. The highest model result shown was o3 Low, which scored 4 percent on ARC-AGI-2 after reaching 75.7 percent on ARC-AGI-1. Its reported cost was $200 per task, while the human cost was $17 per task, showing large performance and cost differences.

Q: Why is MCP becoming important for AI agents?

MCP provides a common method for connecting AI agents to tools and external services. Zapier's implementation can expose 10,000 tools after users configure their desired applications and receive an MCP server URL. OpenAI added MCP to its agents API, and Microsoft introduced it in Copilot Studio. These adoptions suggest agents can use the same connection standard across multiple platforms.

Q: What new image and visual reasoning models were announced?

Reve Image 1.0 demonstrated accurate text rendering, varied styles, realistic subjects, and strong quality rankings based on 100,000 user votes. Ideogram 3.0 emphasized hyperrealistic output and controls for remixing, upscaling, styles, and preferences. Qwen also previewed QVQ-Max, a visual thinking model that can analyze images and videos, reason about their contents, and address tasks involving mathematics, programming, everyday questions, and artistic creation.

Summary & Key Takeaways

• GPT-4o received major improvements to coding, detailed instruction following, technical problem solving, intuition, creativity, and image generation. Artificial Analysis ranked it as the leading non-reasoning coding model discussed. Demand for its image capabilities exceeded expectations, prompting rate limits, while the presenter reported that slow response times were diminishing its usefulness for routine queries.

• Gemini 2.5 Pro impressed the presenter through strong coding performance, fast reasoning, and a million-token context window. Its availability in Windsurf and Cursor enables broader testing on complete software projects. DeepSeek also released a stronger V3 checkpoint that performs well in coding, mathematics, and logic while remaining downloadable under a permissive MIT license.

• ARC-AGI-2 exposed a large gap between human and model performance, while MCP expanded through Zapier, OpenAI, and Microsoft. New image systems from Reve and Ideogram emphasized accurate text, visual quality, and controllability. The update also covered OpenAI's projected revenue growth, leadership changes, and QVQ-Max, a visual reasoning model from Qwen.