ChatGPT 5.4 and 5.4 Pro: Everything New in OpenAI's Latest Models

Introduction

OpenAI has done it again — and this time, the release feels less like an incremental version bump and more like a deliberate repositioning of what ChatGPT is supposed to be. On March 5, 2026, OpenAI officially launched GPT-5.4 and GPT-5.4 Pro, rolling them out across ChatGPT, the developer API, and Codex simultaneously^[11]. The announcement landed with the kind of coordinated fanfare that signals OpenAI considers this a milestone release, not a maintenance update.

But what actually changed? And more importantly — should you care?

If you've been following the rapid-fire cadence of OpenAI's model releases over the past year, you'd be forgiven for feeling a certain fatigue. GPT-5.0 arrived, then 5.1, then 5.2, then 5.3 — each promising meaningful improvements, each delivering them unevenly. The writing quality debates that erupted around GPT-5.2 (more on that later) left a sour taste for creative users. The coding improvements were real but incremental. And through it all, Anthropic's Claude kept closing the gap, particularly on agentic and computer-use tasks.

GPT-5.4 is OpenAI's answer to all of that. It's a model that consolidates reasoning, coding, and agentic capabilities into a single architecture — rather than requiring users to switch between specialized models. It ships with up to 1 million tokens of context in the API and Codex. It introduces native computer-use capabilities that directly compete with (and, by OpenAI's benchmarks, surpass) Claude's offerings. And it comes in three distinct flavors: GPT-5.4 standard, GPT-5.4 Thinking (the version rolling out inside ChatGPT), and GPT-5.4 Pro, which targets maximum performance for the hardest professional tasks^[7].

The practitioner conversation on X has been immediate and substantive. People aren't just celebrating — they're benchmarking, comparing, stress-testing, and debating whether this release actually moves the needle in production workflows. This article breaks down everything new, what the benchmarks actually show, how the three versions differ, and what it all means if you're building with these models or relying on them daily.

Overview

The Three Versions: Standard, Thinking, and Pro

Let's start with the most important structural change. GPT-5.4 isn't a single model — it's a family of three, each optimized for different use cases and available through different surfaces.

GPT-5.4 (Standard) is the base model available through the API and Codex. It's what developers will call directly when building applications. Think of it as the workhorse: fast, capable, and cost-efficient for most tasks.

GPT-5.4 Thinking is the version rolling out inside ChatGPT itself. This is what Plus, Pro, and Team subscribers interact with in the chat interface. The "Thinking" designation means the model engages in extended reasoning — it takes more time to analyze complex queries, holds context better during long research sessions, and supports deeper web research workflows^[5]. OpenAI has been iterating on thinking modes since the o-series models, and GPT-5.4 Thinking represents the most refined version of that approach integrated into the main ChatGPT product.

GPT-5.4 Pro is the ceiling. It's designed for maximum performance on the most complex tasks — think multi-step scientific reasoning, intricate code generation across large repositories, and professional-grade document analysis. It targets users who need the absolute best output quality and are willing to pay for it (and wait slightly longer for responses)^[12].

This three-tier structure is a deliberate move by OpenAI. Rather than forcing users to choose between "fast but shallow" and "slow but deep," they're offering a spectrum. For developers building applications, this means you can route different types of requests to different model tiers based on complexity — a pattern that's becoming standard practice in production AI systems.

What's Actually New: The Feature Breakdown

1 Million Token Context Window

The headline number that's grabbing developer attention is the 1 million token context window, available in the API and Codex^[11]. To put that in perspective: GPT-4 Turbo launched with 128K tokens. GPT-5.2 expanded that significantly. Now, 1M tokens means you can feed the model an entire large codebase, a book-length document, or hours of conversation history — and it can reason across all of it.

This matters enormously for two use cases in particular. First, code analysis: developers working with large monorepos can now point GPT-5.4 at an entire codebase and ask it to trace bugs, understand architectural patterns, or suggest refactors with full context. No more carefully curating which files to include in your prompt. Second, extended agent runs: when you're building AI agents that operate over long workflows — browsing the web, calling APIs, processing documents — the agent accumulates context rapidly. A 1M token window means the agent can maintain coherent behavior across much longer task sequences without losing track of what it's doing.

That said, context window size and effective context utilization are different things. Previous models have shown degraded performance in the middle of very long contexts (the so-called "lost in the middle" problem). OpenAI claims GPT-5.4 handles long contexts more reliably, but practitioners will need to validate this in their specific use cases.

Native Computer-Use Capabilities

This is where GPT-5.4 makes its most aggressive competitive play. The model now has native capabilities for computer use — meaning it can see screens, click buttons, type text, navigate desktops, read emails, and edit spreadsheets^[3].

Anthropic's Claude pioneered this category with its computer-use features, and it's been a genuine differentiator. OpenAI is now directly challenging that lead. According to OpenAI's benchmarks, GPT-5.4 scores 75% on OSWorld (a benchmark for computer-use tasks), compared to Claude Opus 4.6's 72.7%^[11]. That's a narrow margin, but it represents OpenAI crossing the threshold from "catching up" to "competitive or better."

For developers building automation tools, RPA (robotic process automation) replacements, or AI assistants that interact with existing software, this is significant. Native computer use means you don't need to build custom integrations for every application — the model can interact with software the same way a human would, through the visual interface.

The API and Codex implementations include native Playwright browser control, which gives agents programmatic access to web browsing with the model's reasoning capabilities guiding the navigation^[10]. This is a meaningful step toward AI agents that can actually complete real-world tasks end-to-end.

Tool Search: Smarter Tool Selection

One of the less flashy but potentially most impactful features for developers is what OpenAI calls "Tool Search." In current agentic workflows, developers typically define all available tools in the system prompt — which consumes tokens and can confuse the model when the tool list gets long. Tool Search allows GPT-5.4 to dynamically search through a larger ecosystem of available tools and select the right ones for a given task^[11].

This is a practical quality-of-life improvement for anyone building complex agent systems. Instead of bloating your prompts with tool definitions, you can maintain a larger tool registry and let the model figure out which tools are relevant. It's the kind of infrastructure-level improvement that doesn't make headlines but meaningfully reduces the engineering overhead of building agentic applications.

Mid-Generation Steering

Here's a UX innovation that sounds small but changes how you interact with the model: you can now interrupt GPT-5.4 mid-response and redirect it with new instructions^[2]. If the model starts heading in the wrong direction — wrong tone, wrong approach, wrong level of detail — you don't have to wait for it to finish and then re-prompt. You can course-correct in real time.

This feature started rolling out on web and Android, with iOS coming soon^[7]. For power users who spend significant time in ChatGPT, this eliminates one of the most frustrating friction points: watching the model generate a long response you know is wrong, waiting for it to finish, and then crafting a correction prompt. Mid-generation steering makes the interaction feel more like a real collaboration and less like a turn-based game.

Improved Factuality: 33% Fewer Hallucinations

OpenAI is making a strong claim on factuality: GPT-5.4 produces 33% fewer false claims and 18% fewer responses containing any errors compared to GPT-5.2^[3]. In a field where hallucination remains the single biggest barrier to enterprise adoption, these numbers matter.

The 33% reduction is measured against GPT-5.2, which was already an improvement over earlier models. If these numbers hold up in real-world usage (always the caveat with benchmark claims), GPT-5.4 represents a meaningful step toward models you can actually trust for factual tasks — research, analysis, reporting, and knowledge work.

The Benchmarks: What the Numbers Actually Show

Let's look at the benchmark comparisons more carefully, because this is where the rubber meets the road.

The improvements from GPT-5.2 to GPT-5.4 are uneven but significant in key areas:

Coding (SWE-Bench Pro): 55.6% → 57.7%. A modest +2.1 point improvement. This benchmark tests the model's ability to solve real-world repository bug-fix tasks, so even small gains represent meaningful capability improvements. But this isn't a dramatic leap — it's iterative progress^[11].

Knowledge Work (GDPval): ~71% → 83% win/tie rate versus human professionals. This is a +12 point gain on tasks involving documents, spreadsheets, and structured analysis. For enterprise users, this is arguably the most important benchmark — it measures whether the model can do the kind of work that knowledge workers do daily^[9].

Computer Use (OSWorld-Verified): 47.3% → 75%. This is the standout number — a +27.7 point jump that surpasses the reported human baseline of ~72%^[11]. This suggests GPT-5.4's computer-use capabilities aren't just better than previous versions; they're better than what most humans achieve on these standardized tasks.

Science Reasoning (GPQA Diamond): 92.8%. This benchmark tests expert-level science questions, and GPT-5.4's score here is exceptional^[11]. For researchers and scientists using these models, this represents genuine utility for complex reasoning tasks.

Agentic Tool Use: 54.6% versus Claude's 44.8%. Web Browsing: 67.3%^[11]. These numbers reinforce OpenAI's positioning of GPT-5.4 as an "agentic" model — one designed not just for conversation but for taking actions in the world.

The competitive picture is clear: OpenAI is ahead on most agentic benchmarks, but the margins are narrow enough that Anthropic (and Google, with Gemini) could close them with their next releases. The AI capability frontier is advancing rapidly across all major labs.

Efficiency Gains: Faster and Cheaper

Beyond raw capability, OpenAI is emphasizing efficiency. GPT-5.4 uses fewer tokens to accomplish the same tasks, which translates directly to lower costs for API users and faster response times for ChatGPT users^[3].

OpenAI's pitch, as summarized by practitioners on X, is that GPT-5.4 is their "most factual and efficient model — fewer tokens, faster speed." For developers paying per token, efficiency improvements are as valuable as capability improvements. A model that solves the same problem with 30% fewer tokens effectively gives you a 30% cost reduction — without sacrificing quality.

The new pricing structure reflects this positioning. While specific pricing details vary by tier and usage, OpenAI has introduced new pricing options alongside the model launch^[4]. For enterprises evaluating the cost-benefit of upgrading, the combination of better performance and improved efficiency makes a compelling case.

The Agentic Shift: What It Means for Developers

The most significant strategic signal in the GPT-5.4 release isn't any single feature — it's the overall direction. OpenAI is explicitly positioning this model as being "built for agents"^[4].

What does "built for agents" actually mean in practice? It means the model is optimized for workflows where it:

1. Uses tools: Calling APIs, searching the web, executing code, interacting with databases
2. Takes actions: Clicking buttons, filling forms, navigating applications
3. Maintains state: Keeping track of multi-step tasks over long sequences
4. Handles errors gracefully: Recovering from failed tool calls, retrying with different approaches
5. Reasons about strategy: Deciding which tools to use, in what order, and how to verify results

This is a fundamentally different optimization target than "generate good text in response to a prompt." It requires the model to be reliable, consistent, and capable of operating semi-autonomously over extended periods. The 1M token context window, native computer use, Tool Search, and improved factuality all serve this agentic vision.

For developers, this means GPT-5.4 is the first OpenAI model that's genuinely competitive as the backbone of an autonomous agent system. Previous models could be coaxed into agentic behavior with careful prompting and scaffolding, but GPT-5.4 is designed for it from the ground up.

The Writing Quality Question: Has OpenAI Addressed the GPT-5.2 Backlash?

No discussion of GPT-5.4 would be complete without addressing the elephant in the room: writing quality. GPT-5.2 generated significant backlash from creative users who felt that OpenAI had sacrificed writing capabilities to prioritize coding performance.

This post captures a real and ongoing tension in OpenAI's user base. When Sam Altman acknowledged that GPT-5.2 had traded writing quality for coding improvements, it validated what many users had been experiencing. The subsequent restoration of GPT-4o access — after it had been deprecated — was an unusual admission that the newer model wasn't strictly better for all use cases^[7].

OpenAI's response has been to add more customization controls: base styles, tones like "Friendly," and adjustable warmth and enthusiasm settings^[5]. GPT-5.4 continues this trajectory, but the fundamental question remains: can a single model architecture excel at both rigorous technical reasoning and nuanced creative expression?

The early signals are mixed. GPT-5.4's improvements are heavily weighted toward reasoning, coding, and agentic tasks. Creative writing users should test carefully before assuming their experience will be better than with GPT-5.2. OpenAI's stated goal of giving users "more control and customization over how ChatGPT feels to use — not just what it can do" is the right aspiration, but it's an ongoing project, not a solved problem.

Codex Integration: The Developer Experience

For developers specifically, the Codex integration deserves special attention. GPT-5.4 is now the default model in Codex, OpenAI's coding-focused product, with full access to the 1M token context window and native computer-use capabilities^[11].

This means Codex can now:

• Analyze entire large codebases without chunking or summarization
• Execute multi-step coding tasks that involve reading code, understanding architecture, writing new code, and testing it
• Interact with development tools through native computer use — opening terminals, running commands, navigating IDEs
• Search through available tools and libraries using Tool Search to find the right approach for a given problem

For teams using Codex as part of their development workflow, these improvements could meaningfully accelerate development velocity. The combination of massive context and computer use means Codex can handle tasks that previously required significant human scaffolding.

How GPT-5.4 Compares to the Competition

The competitive landscape in March 2026 is more crowded than ever. Here's how GPT-5.4 stacks up:

vs. Anthropic Claude Opus 4.6:

• GPT-5.4 leads on computer use (75% vs 72.7% OSWorld)
• GPT-5.4 leads on agentic tool use (54.6% vs 44.8%)
• Claude has historically been preferred for nuanced writing and instruction following
• The gap is narrowing across the board

vs. Google Gemini:

• GPT-5.4's 1M token context matches Gemini's long-context capabilities
• Google has the advantage of native integration with its productivity suite
• GPT-5.4's agentic capabilities are more mature

vs. Open-source models (Llama, Mistral, etc.):

• The capability gap remains significant for complex reasoning and agentic tasks
• Open-source models offer cost and privacy advantages for simpler use cases
• GPT-5.4 Pro's performance tier has no open-source equivalent

The honest assessment: GPT-5.4 is the best model OpenAI has ever shipped, and it's competitive or leading on most benchmarks. But the era of any single lab having a dominant lead is over. The differences between frontier models from OpenAI, Anthropic, and Google are increasingly about character — how the model feels to use, what it's optimized for, how it handles edge cases — rather than raw capability gaps.

Pricing and Availability

GPT-5.4 Thinking is rolling out to ChatGPT Plus, Pro, and Team subscribers^[0]. The rollout is gradual, so not all users will see it immediately.

For API access, GPT-5.4 is available immediately with new pricing tiers^[13]. The 1M token context window is described as experimental in the API and Codex, suggesting OpenAI may adjust pricing or availability based on usage patterns^[6].

GPT-5.4 Pro is available to ChatGPT Pro subscribers and through the API at a premium price point^[12]. Given its positioning as the maximum-performance option, expect it to be significantly more expensive per token than the standard version.

For enterprise customers evaluating the upgrade, the key calculation is straightforward: GPT-5.4's efficiency improvements (fewer tokens per task) may partially or fully offset the per-token cost increase, depending on your workload. The improved factuality and agentic capabilities add value that's harder to quantify but potentially more impactful.

What Practitioners Should Test First

If you're evaluating GPT-5.4 for production use, here's a prioritized testing checklist:

1. Long-context reliability: Feed it your actual long documents or codebases. Test whether it maintains accuracy across the full context window, not just at the beginning and end.

1. Agentic task completion: Set up multi-step workflows with tool calls. Measure completion rates, error recovery, and consistency across runs.

1. Factuality on your domain: Test with questions where you know the ground truth. The 33% hallucination reduction is an average — your specific domain may see more or less improvement.

1. Computer-use tasks: If you're building automation, test the native computer-use capabilities against your actual target applications. Benchmark performance may not reflect your specific UI complexity.

1. Writing quality: If creative or nuanced writing is important to your use case, compare outputs carefully against GPT-5.2 and GPT-4o. Don't assume the upgrade is uniformly positive.

1. Cost efficiency: Run your typical workloads and compare token usage and costs against your current model. The efficiency gains should be measurable.

The Bigger Picture: OpenAI's Strategic Direction

GPT-5.4 tells us a lot about where OpenAI is heading. The emphasis on agentic capabilities, computer use, and tool integration signals that OpenAI sees the future of AI not as a better chatbot but as a capable digital worker — one that can operate software, complete tasks, and work semi-autonomously over extended periods.

This aligns with the broader industry trend toward AI agents. Every major lab is investing heavily in this direction, and enterprise customers are increasingly asking not "can AI answer my questions?" but "can AI do my work?" GPT-5.4 is OpenAI's most direct answer to that question yet.

The three-tier model structure (Standard, Thinking, Pro) also suggests OpenAI is moving toward a more nuanced product strategy. Rather than one model to rule them all, they're offering a spectrum that lets users and developers choose the right balance of speed, cost, and capability for each task. This is a mature product decision that reflects real-world usage patterns.

The 1M token context window, native computer use, Tool Search, and mid-generation steering aren't just features — they're infrastructure for a new class of AI applications. Applications where the AI doesn't just respond to prompts but actively participates in workflows, maintains context across long sessions, and interacts with the digital world the way humans do.

Whether GPT-5.4 delivers on this vision in practice — not just in benchmarks — is the question that will play out over the coming weeks and months as developers and users put it through its paces.

Conclusion

GPT-5.4 and GPT-5.4 Pro represent OpenAI's most coherent and ambitious model release in the GPT-5 era. The headline features — 1M token context, native computer use, 33% fewer hallucinations, and the three-tier model structure — are individually significant. Taken together, they signal a clear strategic shift toward AI as an autonomous agent rather than a conversational assistant.

The benchmarks are strong. The 75% OSWorld score, 92.8% GPQA Diamond, and 83% win/tie rate against human professionals on knowledge work are numbers that justify attention^[11]. The efficiency improvements — fewer tokens, faster responses — address the practical cost concerns that matter in production deployments.

But benchmarks aren't production. The real test of GPT-5.4 will come from developers building agent systems that need to be reliable over thousands of runs, from knowledge workers who need factual accuracy they can trust without verification, and from creative users who need a model that understands nuance and voice — not just logic and code.

The competitive landscape is tighter than ever. Claude, Gemini, and open-source alternatives are all advancing rapidly. GPT-5.4 gives OpenAI a clear lead on agentic benchmarks today, but that lead is measured in single-digit percentage points, not orders of magnitude. The era of any single model being the obvious choice for every use case is definitively over.

For practitioners, the advice is straightforward: test GPT-5.4 against your actual workloads. The improvements are real, but they're unevenly distributed across different task types. If you're building agents, this is likely the best foundation model available right now. If you're doing creative work, proceed with cautious optimism and keep your fallback models ready.

OpenAI called GPT-5.4 their "most capable and efficient frontier model for professional work." Based on the evidence available today, that claim is defensible. Whether it's enough — enough to justify the hype, enough to maintain OpenAI's market position, enough to deliver on the promise of truly useful AI agents — is the question the next few months will answer.

---

Sources

^[1] ChatGPT — Release Notes — https://help.openai.com/en/articles/6825453-chatgpt-release-notes

^[2] ChatGPT 5.4: All-new 'powerful' OpenAI update for Excel users — https://m.economictimes.com/news/new-updates/chatgpt-5-4-all-new-powerful-openai-update-for-excel-users-availability-usage-in-daily-life-and-more/articleshow/129136480.cms

^[3] ChatGPT New Features 2025–2026: Complete Updated List — https://aiinsider.in/ai-learning/chatgpt-new-features-2025-2026

^[4] GPT-5.4 is here — and OpenAI just made every other AI model look slow — https://www.tomsguide.com/ai/gpt-5-4-is-here-and-openai-just-made-every-other-ai-model-look-slow

^[5] New ChatGPT 5.4 Model Is 'Built for Agents.' Will It Lure Enterprises? — https://www.cnet.com/tech/services-and-software/openai-chatgpt-5-4-thinking-news

^[6] GPT-5.3 and GPT-5.4 in ChatGPT — https://help.openai.com/en/articles/11909943-gpt-52-in-chatgpt

^[7] OpenAI launches GPT-5.4 with Pro and Thinking versions — https://techcrunch.com/2026/03/05/openai-launches-gpt-5-4-with-pro-and-thinking-versions

^[8] GPT-5.3 and GPT-5.4 in ChatGPT — https://help.openai.com/en/articles/11909943-gpt-53-and-gpt-54-in-chatgpt

^[9] OpenAI, in Desperate Need of a Win, Launches GPT-5.4 — https://gizmodo.com/openai-in-desperate-need-of-a-win-launches-gpt-5-4-2000730268

^[10] OpenAI's new GPT-5.4 clobbers humans on pro-level work in tests — https://www.zdnet.com/article/openai-gpt-5-4

^[11] OpenAI launches GPT-5.4 Thinking and Pro — https://thenewstack.io/openai-launches-gpt-5-4

^[12] Introducing GPT-5.4 — https://openai.com/index/introducing-gpt-5-4

^[13] GPT-5.4 pro Model | OpenAI API — https://developers.openai.com/api/docs/models/gpt-5.4-pro

^[14] Models | OpenAI API — https://developers.openai.com/api/docs/models

^[15] ChatGPT just got another brain boost with GPT-5.4 Thinking — https://www.techradar.com/ai-platforms-assistants/chatgpt/chatgpt-just-got-another-brain-boost-with-gpt-5-4-thinking-and-its-built-for-bigger-more-complex-tasks

Further Reading

• [OpenAI Strikes $10B+ Compute Deal with Cerebras for AI Scaling](/buyers-guide/ai-news-openai-cerebras-compute-partnership-2) — OpenAI has forged a multibillion-dollar agreement with chip startup Cerebras Systems to acquire vast computing capacity, potentially exceeding $10 billion, to power its next-generation AI models. The deal, backed by OpenAI CEO Sam Altman who is also an investor in Cerebras, aims to address the growing compute demands for training advanced LLMs. This partnership highlights the intensifying race for AI infrastructure amid chip shortages and escalating costs.
• [OpenAI Unveils Prism: Free AI Tool for Scientific Writing](/buyers-guide/ai-news-openai-prism-launch) — OpenAI launched Prism on January 27, 2026, a free AI-powered workspace integrated with GPT-5.2 to assist scientists in drafting, revising, and collaborating on research papers. It features LaTeX support, diagram generation from sketches, full-context AI assistance, and unlimited team collaboration. Available to all ChatGPT users, it aims to accelerate scientific discovery through human-AI partnership.
• [OpenAI Unveils Prism: Free AI Workspace Powered by GPT-5.2](/buyers-guide/ai-news-openai-prism-workspace-launch) — OpenAI announced Prism on January 27, 2026, a free, AI-native workspace designed for scientists to draft, revise, and collaborate on research papers using LaTeX integration. Powered by the advanced GPT-5.2 model, it offers features like contextual editing, literature search, equation conversion from handwriting, and unlimited real-time collaboration. Available immediately to ChatGPT users, it aims to streamline fragmented research workflows.
• [OpenAI Launches Codex Mac App for Multi-Agent Coding](/buyers-guide/ai-news-openai-codex-app-release) — OpenAI released the Codex app for macOS on February 2, 2026, serving as a command center for developers to manage multiple AI coding agents. The app enables parallel execution of tasks across projects, supports long-running workflows with built-in worktrees and cloud environments, and integrates with IDEs and terminals. Powered by GPT-5.2-Codex model, it includes skills for advanced functions like image generation and automations for routine tasks.
• [OpenAI Unveils GPT-5.3-Codex: Coding AI Breakthrough](/buyers-guide/ai-news-openai-gpt-5-3-codex-release) — OpenAI released GPT-5.3-Codex, a advanced coding model achieving 57% on SWE-Bench Pro, 76% on TerminalBench 2.0, and 64% on OSWorld benchmarks. It introduces mid-task steerability, live updates, faster token processing (over 25% quicker), and enhanced computer use capabilities. This launch follows Anthropic's Claude Opus 4.6, intensifying competition in AI coding tools.

Related Articles

• OpenAI Accelerates GPT-5.2 Amid Gemini 3 Rivalry

  OpenAI has declared an internal 'code red' to fast-track the release of GPT-5.2, a new frontier model designed to surpass Google's Gemini 3 in speed, reliability, and reasoning. Polymarket odds indicate a high likelihood of launch by December 9, with insiders adjusting bets amid rumors of superior coding capabilities. This move signals intensified competition in the AI landscape.

• OpenAI Preps New Audio Model for Device Launch

  OpenAI is revamping its audio AI models with a new voice architecture releasing in Q1 2026 to power an upcoming audio-first personal device. The upgrades enable more natural, emotional speech, faster responses, and improved real-time interruption handling for proactive user assistance. Internal teams have merged to accelerate development, positioning this as a key step toward companion-style AI.

• OpenAI Develops Advanced Audio AI for Upcoming Device

  OpenAI is ramping up its audio AI efforts with a new model architecture set for Q1 2026 release, featuring more natural and emotive speech, faster responses, and improved real-time interruption handling. This upgrade supports an upcoming audio-first personal device expected in about a year, potentially including glasses and smart speakers, with collaboration from Jony Ive. The initiative merges internal teams to enhance voice interactions for proactive AI companions.

• SoftBank Pumps $40B into OpenAI for AI Dominance

  SoftBank announced a massive $40 billion investment in OpenAI, marking one of the largest funding rounds in AI history. This capital infusion aims to accelerate OpenAI's development of advanced AI models and infrastructure. The deal underscores growing investor confidence in OpenAI's leadership in generative AI.

• OpenAI Upgrades ChatGPT Deep Research with GPT-5.2 Model

  OpenAI announced on February 10, 2026, that its Deep Research feature in ChatGPT is now powered by the advanced GPT-5.2 model, enabling more sophisticated analysis and insights. The update includes enhanced controls for source selection and internal document integration, rolling out progressively to users. This builds on previous capabilities to make AI-driven research more accurate and customizable for complex queries.