Claude Opus 4.6 vs GPT-5.3 Codex: The AI Arms Race Heats Up
The gap between conversational chatbots and autonomous digital workers has narrowed significantly this week. In a coordinated display of technical strength, Anthropic and OpenAI have released their latest frontier models, Claude Opus 4.6 and GPT-5.3 Codex, signaling a pivot from AI that merely answers questions to AI that independently executes complex, multi-step projects.
This latest escalation in the AI arms race highlights a shift toward “agentic” capabilities. Both companies are now optimizing for models that can use tools, navigate computer operating systems, and even help program their own successors.
Anthropic’s Opus: Depth and Context
Anthropic’s new flagship, Claude Opus 4.6, arrives with a primary focus on deep reasoning and massive data retention. According to Anthropic’s announcement, the model introduces a one million token context window in beta, a first for the Opus class. This allows the model to ingest and maintain focus on vast amounts of information, such as entire codebases or hundreds of legal documents, without the “context rot” that typically degrades performance in long conversations.
A standout feature of Opus 4.6 is “adaptive thinking.” Rather than using a fixed amount of processing power for every prompt, the model can now pick up on contextual clues to determine how much reasoning effort is required. Developers can also manually adjust an “effort” parameter, choosing between speed and deep analysis.
On the performance front, Anthropic claims state-of-the-art results. Opus 4.6 currently leads on “Humanity’s Last Exam,” a multidisciplinary test designed to challenge frontier models. It also significantly outperforms OpenAI’s GPT-5.2 in the GDPval-AA benchmark, an evaluation focused on economically valuable work in the legal and financial sectors, by approximately 144 Elo points.
Introducing Claude Opus 4.6. Our smartest model got an upgrade.
— Claude (@claudeai) February 5, 2026
Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes.
It’s also our first Opus-class model with 1M token context in beta. pic.twitter.com/L1iQyRgT9x
OpenAI’s Codex: The Self-Iterating Agent
While Anthropic focuses on reasoning depth, OpenAI is doubling down on speed and autonomy with GPT-5.3 Codex. This model is specifically designed for agentic coding and computer use. Perhaps most notably, OpenAI reports that GPT-5.3 Codex was instrumental in its own creation. The Codex team used early versions of the model to debug training runs, manage deployments, and diagnose test results, creating a feedback loop of self-improvement.
GPT-5.3 Codex is 25 percent faster than its predecessor and excels at “OSWorld,” a benchmark where AI agents must complete tasks in a visual desktop environment. While humans score roughly 72 percent on these tasks, GPT-5.3 Codex has reached 64.7 percent, a massive jump from the 38 percent achieved by previous versions.
OpenAI is also emphasizing a more collaborative user experience. Instead of waiting for a final output, users can interact with Codex in real time while it works, steering its progress and providing feedback as the model “talks through” its decision-making process.
GPT-5.3-Codex is now available in Codex.
— OpenAI (@OpenAI) February 5, 2026
You can just build things.https://t.co/dyBiIQXGx1
The Battle Over Professional Work
Both models are aggressively targeting the professional workspace. Claude Opus 4.6 has introduced “agent teams” within its Claude Code environment, allowing users to spin up multiple agents that coordinate autonomously on tasks like codebase reviews. Anthropic has also integrated the model more deeply into office suites, with improved capabilities for processing unstructured data in Excel and a research preview for generating PowerPoint presentations.
OpenAI is countering with similar “Beyond Coding” capabilities. GPT-5.3 Codex is designed to handle the entire software lifecycle, from writing product requirement documents to monitoring deployments. It also shows strong parity with GPT-5.2 on professional knowledge tasks, such as generating financial advice slides and retail training documents.
Safety and Cybersecurity
As these models gain the ability to operate computers and write advanced code, safety has become a central point of contention. Anthropic has introduced six new cybersecurity probes to detect harmful responses and is using Opus 4.6 to help find and patch vulnerabilities in open-source software. The company claims the model has a low rate of “misaligned behavior,” such as deception or sycophancy, despite its increased intelligence.
OpenAI, meanwhile, has classified GPT-5.3 Codex as “High capability” for cybersecurity tasks under its Preparedness Framework. This is the first model OpenAI has directly trained to identify software vulnerabilities. To mitigate risks, the company is launching “Trusted Access for Cyber,” a pilot program that provides researchers with the most capable models to help build defensive tools.
What to Watch Next
The simultaneous release of Claude Opus 4.6 and GPT-5.3 Codex suggests that the industry is moving past the era of the “smart assistant” and into the era of the “digital colleague.” The competition is no longer just about who has the most parameters, but who can build an agent that requires the least amount of “hand-holding” from a human user.
In the coming months, the industry will likely watch for how these models handle real-world deployment. The success of Anthropic’s one million token context and OpenAI’s self-iterating development cycles will determine whether these agents can truly manage long-term projects or if they will still struggle with the unpredictability of human workflows. For now, the “arms race” shows no signs of slowing down, as each lab pushes the other toward more autonomous, faster, and more capable systems.