AI Agent News
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Latest industry news
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Key events timeline
OpenClaw erupts on GitHub
OpenClaw hits global GitHub Top 10 in 10 days, outpacing the Linux kernel star growth
Meta acquires Manus for $2B
Meta acquires Manus AI for $2B, locking in the general-purpose Agent race
DeepSeek-V3 open-sourced
The value king, at just 5% of GPT-4 cost
Manus goes viral overnight
The world's first general-purpose AI Agent draws unprecedented attention
OpenAI Deep Research
OpenAI ships a deep-research Agent that generates professional reports in one click
MCP Servers pass 500
The MCP ecosystem erupts — 500+ servers built in 3 months
DeepSeek-R1 stuns the world
Open-source reasoning model at just 3% of OpenAI cost, reshaping the global AI landscape
MCP protocol born
Anthropic releases the Model Context Protocol, the de facto standard for Agent interfaces
Claude Computer Use
Anthropic lets AI directly control the computer screen for the first time, opening a new paradigm
Replit Agent full-stack automation
Natural language to a shipped product, aimed at non-engineers
Cursor ARR passes $100M
The fastest-growing SaaS ever, the new king of AI coding tools
Claude 3.5 tops SWE-bench
The strongest coding AI, bug-fixing at a junior engineer level
Devin launches
The world's first autonomous AI software engineer, able to complete full coding tasks on its own
GPT-5 Launch: 5 Major Breakthroughs and Real User Reviews
## GPT-5 Officially Released, OpenAI Calls It the 'Biggest Leap' In May 2026, OpenAI officially released GPT-5, the most significant model update since GPT-4. The company claims five major breakthroughs. This article provides an independent evaluation based on public benchmarks and early user feedback. ## 5 Official Breakthroughs ### 1. Comprehensive Multimodal Capabilities GPT-5 natively supports text, image, audio, and video input, eliminating the need to switch between 'GPT-4o' and 'DALL-E'. **Test**: Uploaded a product demo video, asked to 'extract key selling points and generate marketing copy'—GPT-5 completed it fully, with video understanding quality close to Gemini 2.5 Pro. ### 2. Reasoning Ability Close to o3 | Benchmark | GPT-4o | GPT-5 | o3 (Reference) | |-----------|--------|-------|---------| | AIME 2024 | 13.4% | 72.3% | 96.7% | | GPQA Diamond | 53% | 79.1% | 87.7% | | SWE-bench | 38% | 57.6% | 71.7% | For most users, there's no longer a need to choose between 'fast model' and 'reasoning model'—GPT-5 handles most scenarios directly. ### 3. Context Window Expanded to 256k Increased from 128k to 256k, handling longer documents and codebases. ### 4. Tool Calling Reliability Greatly Improved Function Calling success rate increased from 84% to 96%—significant for AI Agent applications. ### 5. Price Same as GPT-4o Despite major capability improvements, GPT-5 pricing is comparable to GPT-4o ($2.5/1M input tokens). --- ## Real User Feedback **Developer**: > 'Writing complex business logic, one-shot success rate went from 60% to 80%+. Function Calling is much more stable.' **Content Creator**: > 'Chinese writing quality has improved significantly, especially coherence and logical structure in long texts.' **Researcher**: > 'Math reasoning is weaker than o3, but sufficient for most research tasks. No need to wait for o3's slow responses.' --- ## When to Use GPT-5? When to Use o3? **Use GPT-5**: Daily work tasks, multimodal tasks, real-time conversations requiring fast responses, Agent tool calls **Continue with o3**: Math proofs, high-precision code debugging, research tasks requiring highest reasoning quality --- ## Pressure on Claude and Gemini - Claude 3.5 Sonnet's writing advantage is significantly narrowed by GPT-5 - Gemini 2.5 Pro's multimodal advantage remains, but GPT-5 has entered the same competitive tier Anthropic is expected to release Claude 4 in Q3, and Google will accelerate Gemini updates. --- ## Conclusion GPT-5 is a truly meaningful iteration. For most users, it can replace GPT-4o + separate image generation + most o3 use cases in the existing toolchain. If you use only one AI tool, GPT-5 will be the most reasonable primary choice in the second half of 2026.
OpenAI o3 and o4-mini Hands-On Analysis: Which Tasks Need Reasoning Models? A Selection Guide
## Direct Answer **What tasks are o3/o4-mini suitable for?** Reasoning models (o-series) are ideal for: mathematical proofs, complex code debugging, logic puzzles, scientific reasoning—any task requiring multi-step verification. **What are they not suitable for?** Everyday writing, quick Q&A, creative tasks—GPT-4o is faster and cheaper for these, with comparable results. **One-sentence distinction**: GPT-4o is "smart intuition," o3 is "rigorous reasoning." ## o3 vs o4-mini vs GPT-4o Selection Guide | Scenario | Recommended Model | Reason | |----------|-------------------|--------| | Math competition problems/proofs | o3 | Deepest reasoning, highest accuracy | | Complex algorithm design | o3 | Strong multi-step planning | | Code bug debugging | o4-mini | Sufficient and 6x cheaper | | Everyday code generation | GPT-4o | Fast, cost-effective | | Scientific paper analysis | o3 | Rigorous logic, accurate citations | | Copywriting | GPT-4o | Better creativity; reasoning models can be rigid | | Quick Q&A | GPT-4o / GPT-4o-mini | Reasoning models have long wait times | ## Differences Between o3 and o4-mini ### o3 (Flagship Reasoning Model) - **Capability**: Strongest reasoning depth, best for hardest tasks - **Speed**: Slow (30 seconds–3 minutes per query, depending on complexity) - **Price**: $15/1M input, $60/1M output - **Best for**: Research, high-precision code, strategic planning ### o4-mini (Lightweight Reasoning Model) - **Capability**: About 80% of o3's reasoning ability - **Speed**: Faster than o3 (10–30 seconds per query) - **Price**: $1.1/1M input, $4.4/1M output (1/14 of o3) - **Best for**: Daily reasoning tasks, cost-sensitive scenarios ## Hands-On: Performance on 6 Typical Tasks ### Task 1: Math Competition (AMC/AIME Problems) - **o3**: Accuracy **91%** - **o4-mini**: Accuracy 84% - **GPT-4o**: Accuracy 67% → **Winner: o3** ### Task 2: Python Code Debugging (Complex Bugs) - **o3**: First-fix success rate 78% - **o4-mini**: First-fix success rate 71% - **GPT-4o**: First-fix success rate 58% → **Winner: o4-mini** (best value) ### Task 3: Creative Copywriting - **o3**: Content quality 6.8/10 (logical but rigid) - **GPT-4o**: Content quality 8.4/10 (more fluent, creative) → **Winner: GPT-4o** ### Task 4: Scientific Paper Interpretation - **o3**: Clearly superior in accuracy and depth; can identify logical flaws in papers → **Winner: o3** ### Task 5: SQL Query Optimization - **o4-mini** performs on par with o3, but is 14x cheaper → **Winner: o4-mini** (best value) ### Task 6: Strategic Planning (Business Proposals) - **o3**: Most complete structure, considers most dimensions - **GPT-4o**: More creative, but slightly less rigorous logic → **Depends on needs** ## API Usage Tips for Reasoning Models ```python from openai import OpenAI client = OpenAI() # Use o4-mini with controlled reasoning depth response = client.chat.completions.create( model='o4-mini', messages=[ {'role': 'user', 'content': 'Prove that there are infinitely many prime numbers, requiring a rigorous mathematical proof'} ], # reasoning_effort: 'low' | 'medium' | 'high' controls reasoning depth and cost reasoning_effort='high' ) ``` **Cost-saving tips**: - Use `reasoning_effort='low'` for quick validation, `'high'` only for final output - Use Batch API for batch tasks (50% cheaper than real-time calls) - Start with o4-mini; switch to o3 only if unsatisfied ## FAQ **Q: Can o3 get simple questions wrong due to "overthinking"?** A: Yes, this is known as "overthinking." Passing simple problems to reasoning models can sometimes lead to errors from over-analysis. It's recommended to use o3 only for truly complex tasks. **Q: o3 has long wait times; any optimization methods?** A: Use streaming output (streaming=True) to see partial output while o3 is thinking, improving user experience. **Q: Will the o-series be replaced by GPT-5 in the future?** A: GPT-5 already has built-in reasoning mode, but o3's extreme reasoning capabilities (for research, etc.) will remain relevant for some time. ## Related Resources - Agent reasoning mode comparison: [aiskillnav.com/tutorials/agent-reasoning-vs-streaming-tradeoff](https://aiskillnav.com/tutorials/agent-reasoning-vs-streaming-tradeoff) - Full AI model comparison: [aiskillnav.com/models](https://aiskillnav.com/models)
GPT-5 Officially Released: In-Depth Analysis of Capabilities, Pricing, and Impact on the AI Industry
## Direct Answer **GPT-5's 3 Most Important Upgrades:** 1. **Native Multimodal Reasoning**: Not just seeing images, but joint reasoning across images, text, and code. 2. **Adaptive Computation**: Automatically adjusts compute based on task complexity—quick answers for simple questions, deep reasoning for complex ones. 3. **Native Agent Capabilities**: Built-in tool calling, memory management, and multi-step planning without extra frameworks. **GPT-5 vs GPT-4o: Is It Worth Upgrading?** For everyday users: Not much difference; GPT-4o is already good enough. For developers and professional users: Significant improvements in code and reasoning, with complex task success rates up by 30%+. ## GPT-5 Release Background OpenAI released GPT-5 in 2026, about 2 years after GPT-4. This is OpenAI's largest model upgrade: - Training compute: ~10x that of GPT-4 - Parameter count: Not disclosed, but leaked info suggests 1-2 trillion - Training data cutoff: End of 2025 - Release format: Claude.ai web version + API ## Core Capabilities Explained ### Multimodal Reasoning (True Cross-Modal Understanding) GPT-4o can "see and talk"; GPT-5 can "reason across modalities": **Example**: Give GPT-5 a circuit diagram + a piece of code, ask "Which circuit module does this code control? Where might the bug be?" → GPT-5 simultaneously analyzes the diagram and code, providing a cross-modal diagnostic answer. → GPT-4o can only analyze separately, unable to perform joint reasoning. ### Adaptive Computation (o1 Capabilities Built-In) GPT-5 integrates o1/o3 reasoning modes into the same model: - Simple questions: Direct answer (fast, cheap) - Complex questions: Automatically triggers "deep thinking" mode (slow but accurate) - Users can manually control reasoning depth (`thinking_level: auto/low/high`) ### Native Agent Architecture GPT-5's API natively supports: - **Persistent Memory**: Remembers user preferences and context across conversations. - **Enhanced Tool Calling**: Calls multiple tools in parallel, reducing round trips. - **Plan-Execute Mode**: Generates a complete plan first, executes after user confirmation. ## Benchmark Comparison | Benchmark | GPT-4o | GPT-4.5 | GPT-5 | |-----------|--------|---------|-------| | MMLU | 87% | 90% | 95% | | HumanEval (Code) | 90% | 93% | 97% | | SWE-bench | 48% | 55% | 69% | | MATH | 76% | 84% | 93% | | GPQA (Science) | 61% | 70% | 82% | ## Pricing (API) | Model | Input | Output | |-------|-------|--------| | GPT-4o | $2.5/1M | $10/1M | | GPT-5 Standard | $8/1M | $25/1M | | GPT-5 Deep Reasoning | $15/1M | $60/1M | **Conclusion**: GPT-5 is 3-4x more expensive than GPT-4o, suitable for high-value tasks; for daily tasks, sticking with GPT-4o is more cost-effective. ## Impact on the AI Industry **Pressure on Anthropic (Claude)**: GPT-5 narrows the gap with Claude 3.5 in coding, pushing Anthropic to accelerate the Claude 4 series release. **On Open Source Community**: GPT-5's release will spur Meta to speed up Llama 4 development; the gap between open-source and commercial models is shrinking. **On Developers**: GPT-5's native agent architecture lowers the barrier to building AI agent applications, expecting an explosion of agent apps in 2026. ## FAQ **Q: Can ChatGPT free users use GPT-5?** A: Initially, only ChatGPT Plus/Team/Enterprise users. Free users can expect a limited version in about 6 months. **Q: Will GPT-5 replace the o1/o3 series?** A: In the long term, yes. OpenAI's strategy is to integrate reasoning capabilities into the main series models; the o-series will gradually phase out. **Q: Can users in China access it?** A: OpenAI is not directly accessible in China; users need to use API proxies or third-party platforms that support OpenAI. ## Related Resources - Complete AI Model Comparison: [aiskillnav.com/models](https://aiskillnav.com/models) - OpenAI vs Anthropic vs Google Comparison: [aiskillnav.com/news/openai-vs-anthropic-vs-google-2025-ai-battle](https://aiskillnav.com/news/openai-vs-anthropic-vs-google-2025-ai-battle)
Claude 4 Full Series Deep Dive: Opus 4, Sonnet 4 Capabilities and Usage Guide
Anthropic has officially released the Claude 4 series, including Opus 4 (top-tier reasoning) and Sonnet 4 (high cost-performance). This article provides an in-depth analysis of the core capability improvements of both models, comparisons with the previous generation, real-world performance, and guidance on which model to choose for different scenarios. ## Quick Answer **The three most important upgrades in Claude 4:** 1. **Extended Thinking 3.0**: Significantly improved reasoning depth, with math/coding benchmarks exceeding 95% 2. **200K→500K Context**: Opus 4 supports 500K tokens, equivalent to 400 pages of PDF 3. **Tool Call Stability**: Multi-tool concurrent call success rate increased to 98%, with notable improvement in agent task completion ## Claude 4 Release Background In May 2026, Anthropic officially launched the Claude 4 series at its annual developer conference, about 10 months after the Claude 3.5 series. This is the largest model upgrade in Anthropic's history, with simultaneous releases of: - Claude Opus 4 (flagship reasoning model) - Claude Sonnet 4 (high cost-performance workhorse) - Claude Haiku 4 (ultra-fast lightweight model) - Claude Code 2.0 (coding agent designed for developers) ## Opus 4 vs Sonnet 4: How to Choose | Aspect | Opus 4 | Sonnet 4 | |--------|---------|----------| | **Positioning** | Top-tier reasoning, complex tasks | Daily workhorse, cost-effective choice | | **Context** | 500K tokens | 200K tokens | | **Speed** | Moderate (deep thinking) | Fast (2-3x) | | **Price** | $15/M input tokens | $3/M input tokens | | **Best for** | Math proofs, long document analysis, complex code refactoring | Daily writing, code generation, conversation | **Recommendation**: 90% of daily tasks can be handled by Sonnet 4; only tasks requiring deep reasoning (research reports, complex algorithm design) need Opus 4. ## Benchmark Data | Benchmark | Claude 3.5 Sonnet | Claude Sonnet 4 | Claude Opus 4 | |-----------|-------------------|-----------------|---------------| | SWE-bench | 49% | 62% | 74% | | MATH | 71% | 83% | 92% | | GPQA | 59% | 68% | 78% | | HumanEval | 92% | 95% | 97% | ## Key Changes for Developers ### API Level - New `thinking_budget` parameter (controls reasoning depth, balancing cost and quality) - Tool calls support streaming output (significantly reduces time-to-first-token) - New `computer_use_2.0` tool type (enhanced interface manipulation capability) ### Claude Code 2.0 - Supports simultaneous understanding of multiple code repositories (up to 5 repos) - New "Planning Mode": outputs a complete modification plan first, then executes after user confirmation - Test-driven development: automatically generates tests → runs them → modifies code based on failures, iterating in a loop ## Common User Feedback (First Week After Release) **Positive**: - "Sonnet 4's coding ability is noticeably stronger than 3.5, with higher one-shot generation success rate" - "Extended Thinking provides clearer steps for math problems, significantly reducing error rates" **Areas for Improvement**: - "Opus 4 is expensive; medium tasks don't need it" - "Image generation still relies on third parties; hope for native image capabilities" ## FAQ **Q: Can I still use Claude 3.5 Sonnet?** A: Yes, Anthropic promises to support it for at least 12 months. However, from a cost-performance perspective, Sonnet 4 offers similar pricing with stronger capabilities, so gradual migration is recommended. **Q: Has Claude 4 improved Chinese language support?** A: Yes, significantly. Chinese comprehension accuracy has improved by about 15%, and generated Chinese text is more natural and fluent, with fewer awkward translation artifacts. ## Related Resources - AI Model Comparison: [aiskillnav.com/models](https://aiskillnav.com/models) - Claude Code Usage Guide: [aiskillnav.com/tutorials/claude-code-vs-cursor-2026-complete-comparison](https://aiskillnav.com/tutorials/claude-code-vs-cursor-2026-complete-comparison)
OpenAI vs Anthropic vs Google: Who Leads the 2025 AI Tool War?
The AI race is heating up. We compare the latest products from three major players to help you choose the best AI tool for your needs in 2025. ## Quick Answer **2025 Three Major AI Platforms Comparison Summary**: - **Best Overall Experience**: OpenAI (ChatGPT ecosystem is most mature) - **Strongest Technical Capability**: Anthropic Claude (top for coding/analysis, highest safety) - **Best Enterprise Integration**: Google Gemini (seamless integration with Google Workspace) - **Best Value**: DeepSeek (open-source, cost is 1/10 of mainstream models) ## In-Depth Comparison ### OpenAI — Most Complete Ecosystem, Largest User Base **Core Products**: ChatGPT, GPT-4o, o3, DALL-E 3, Sora **Advantages**: - Most user-friendly interface, suitable for non-technical users - Richest plugin ecosystem (Plugins/GPTs) - Most mature multimodal capabilities (text+image+voice+video) - Most comprehensive API ecosystem with the most third-party integrations **Disadvantages**: - High cost for top-tier models - Lags behind Claude in certain specialized tasks (coding, analysis) **Best for**: General users, creative professionals, teams needing multimodal capabilities ### Anthropic Claude — Strongest Technical Capability, Most Popular with Developers **Core Products**: Claude 3.5 Sonnet, Claude Opus, Claude Code **Advantages**: - Industry-leading coding and code analysis - Ultra-long context (200K tokens), best for processing large documents - Constitutional AI, highest safety and controllability - Claude Code changes the way AI-assisted development works **Disadvantages**: - Image generation relies on third parties - Slightly conservative in some creative writing scenarios **Best for**: Developers, researchers, professional users who need to process complex documents ### Google Gemini — Best Enterprise Integration, Strongest Data Capabilities **Core Products**: Gemini 1.5 Pro, Gemini 2.0 Flash, Gemini Advanced **Advantages**: - Deep integration with Google Workspace (Gmail, Docs, Sheets) - 1 million token context, handles ultra-long documents effortlessly - Google Search integration, best real-time information retrieval - Free tier is more generous than competitors **Disadvantages**: - Still lags behind GPT-4o and Claude in complex reasoning tasks - Inconsistent quality in creative content generation **Best for**: Google ecosystem users, enterprise teams, scenarios requiring ultra-long document processing ## Use Case Selection Guide | Scenario | Recommended Platform | Reason | |----------|----------------------|--------| | Daily conversation/writing | ChatGPT | Easiest to use, most natural responses | | Code development | Claude | Best coding capabilities | | Document processing | Gemini | Longest context | | Image generation | ChatGPT (DALL-E) | Most mature text-to-image | | Academic research | Claude | Rigorous logic, accurate citations | | Cost-sensitive scenarios | DeepSeek | Open-source, ultra-low cost | ## FAQ **Q: How do the data privacy policies differ among the three?** A: All three offer enterprise versions where data is not used for training. For personal versions, Anthropic's privacy policy is relatively the most transparent. **Q: Who will win the AI war in the future?** A: It's likely not a winner-takes-all scenario. Different scenarios will have different optimal choices, and the diversified landscape of AI tools will persist. ## Related Resources - View all AI model comparisons: [aiskillnav.com/models](https://aiskillnav.com/models) - Stay updated on the latest AI news: [aiskillnav.com/news](https://aiskillnav.com/news)
Anthropic Launches Claude Opus 4.1: 50% Cost Reduction, Top-Tier Reasoning Performance
Anthropic officially released Claude Opus 4.1, the most cost-effective version to date. Opus 4.1 scores 72.8% on SWE-bench (coding tasks), surpassing all open-source models and matching GPT-4o. Meanwhile, API pricing drops from $15 to $7.50 per million tokens, a 50% reduction. More importantly, Opus 4.1 introduces a 'Progressive Reasoning' mode, allowing agents to verify and correct intermediate results in real-time during reasoning, improving accuracy by 23% compared to traditional one-shot output. Cursor and Windsurf have announced they will use Opus 4.1 as the default backend.
Multimodal AI Tools Explained: What They Are, Why They Matter, and Real-World Applications in 2025
Multimodal AI can see, hear, and read. Learn how multimodal AI tools work and their most valuable use cases across industries. ## Quick Answer **What is multimodal AI?** Multimodal AI refers to AI systems that can process multiple types of input simultaneously (text, images, audio, video). The most representative multimodal AI tools in 2025: GPT-4o (text+image+voice), Gemini 1.5 Pro (text+image+audio+video), Claude 3.5 (text+image). **Why is multimodal important?** Because real-world information is not just text—over 80% of data is unstructured images, video, and audio. Multimodal AI enables AI to truly handle complex real-world information. ## Core Multimodal AI Tools ### GPT-4o (OpenAI) - **Modalities**: Text, image, voice (real-time) - **Highlights**: Real-time voice conversations with extremely fast response; can "see" images and answer questions - **Best Use Cases**: Real-time voice assistants, image analysis, visual question answering ### Gemini 1.5 Pro (Google) - **Modalities**: Text, image, audio, video - **Highlights**: Can process over 1 hour of video content; 1 million token context - **Best Use Cases**: Video content analysis, long multimedia document processing ### Claude 3.5 Sonnet (Anthropic) - **Modalities**: Text, image - **Highlights**: Best combination of image understanding and code generation (analyze UI screenshots → generate code directly) - **Best Use Cases**: Generating code from design mockups, document image analysis ## Industry Applications ### Healthcare - Medical imaging assisted diagnosis (X-ray, CT scan analysis) - Combined analysis of medical record photos and text - Real-time surgical video guidance ### Education - Photo-based problem solving: students take a photo of a math problem, AI provides step-by-step solutions - Handwritten homework grading: upload handwritten documents, AI recognizes and grades - Automatic subtitling and translation for multilingual videos ### E-commerce - Automatic product description generation from product images - Competitor price monitoring (screenshot analysis) - User uploads photo, AI recommends similar products ### Creative Design - Sketch to finished design: upload hand-drawn sketch, AI generates detailed design - Brand visual audit: batch analyze images for brand compliance - Video content summarization: 1-hour video → 3-minute highlight summary ### Enterprise Operations - Contract image/PDF scanning to extract key clauses - Automatic structured meeting minutes from meeting recordings - Intelligent security surveillance analysis (anomaly behavior detection) ## How to Choose a Multimodal AI Tool | Need | Recommended Tool | Reason | |------|-----------------|--------| | Real-time voice interaction | GPT-4o | Lowest latency, most natural | | Video content analysis | Gemini 1.5 Pro | Only one supporting long videos | | Image to code | Claude 3.5 | Strongest code+vision combination | | Batch image processing | GPT-4o API | Most mature API ecosystem | ## FAQ **Q: How accurate is multimodal AI in image recognition?** A: In standard image understanding tests, GPT-4o and Gemini achieve 85-95% accuracy, but in specialized fields (medical imaging, industrial inspection), human verification is still required. **Q: Are API calls for multimodal AI expensive?** A: Image input increases token consumption. One image is roughly equivalent to 500-1000 text tokens, increasing costs accordingly. Video processing is the most expensive; it's recommended to validate ROI in small-scale scenarios first. **Q: Are there free multimodal AI tools?** A: Yes, ChatGPT free version (with usage limits), Google Gemini free version, and Bing Copilot (based on GPT-4o) all offer free multimodal capabilities. ## Related Resources - Compare all AI model capabilities: [aiskillnav.com/models](https://aiskillnav.com/models) - Explore multimodal AI use cases: [aiskillnav.com/usecases](https://aiskillnav.com/usecases)
Claude 4 Resets the Agent Benchmark: SWE-bench Breaks 72%, Surpassing Junior Human Programmers
Anthropic releases Claude 4, achieving a score of 72.5% on the software engineering benchmark SWE-bench Verified, significantly surpassing the average performance of junior human engineers (around 60%) for the first time. Claude 4 introduces an "Extended Thinking" mode, allowing the model to engage in internal reasoning for several minutes before delivering a final answer, particularly excelling in agent tasks that require multi-step planning. AI IDEs like Cursor and Windsurf have announced priority access to Claude 4 as their default agent engine.
GPT-4o Real-Time Voice + Function Calling: Agents Enter a New Era of 'Listen, Speak, Act'
OpenAI has significantly upgraded GPT-4o's real-time API, introducing mid-conversation function calling capabilities. This allows agents to query databases, call tools, and execute code during live voice conversations, with results seamlessly flowing back into the dialogue. This breakthrough removes the final barrier for voice agents: they are no longer just 'talkative assistants' but executors that can complete orders, check balances, and control systems in real-time during calls. Several domestic voice SaaS providers have announced integration.