Claude Opus 4.6 vs GPT-5.3-Codex: Depth Meets Speed

On February 5th, 2026, Anthropic and OpenAI did something nobody expected: they launched their flagship models within 15 minutes of each other. Not by accident. TechCrunch confirmed both companies originally planned a 10:00 AM PST release, but Anthropic moved up by 15 minutes to steal the opening shot. Three days later, both bought Super Bowl ads. Two days before that, Anthropic’s Cowork legal plugin triggered a $300B software stock selloff. This was the most chaotic week in AI history — and it revealed something the benchmark charts don’t show.

Claude Opus 4.6 bets on depth: agent teams, a 1M token context window, and 500+ zero-day vulnerability discoveries. GPT-5.3-Codex bets on speed: 25% faster inference, self-training capabilities, and OpenAI’s first “High” cybersecurity classification. Neither model is better. They’re barely competing on the same axis.

Claude Opus 4.6: Anthropic’s Enterprise Depth Play

The headline feature is agent teams — a research preview enabling multiple Claude instances to work in parallel on different aspects of a complex project. Think contract review, financial impact analysis, and regulatory compliance running simultaneously on the same deal. This isn’t multi-agent debate or chain-of-thought routing. It’s true parallel execution with shared context, and for document-heavy industries like law, finance, and healthcare, it represents a fundamentally different architecture for how AI work gets done.

The 1M token context window (in beta) backs this up with hard numbers. Opus 4.6 scored 76% on the 8-needle MRCR v2 retrieval test at 1M context. For comparison, Claude Sonnet 4.5 scored 18.5% on the same test. That’s not an incremental improvement — it means Opus 4.6 can reliably find and use information buried deep in massive document repositories where previous models were essentially guessing. For anyone processing multi-thousand-page regulatory filings or contract repositories, this is the difference between a useful tool and an expensive autocomplete.

Enterprise benchmarks tell the same story. Opus 4.6 leads GDPval-AA (the finance and legal reasoning benchmark) by 144 Elo points over GPT-5.2 — a 190 Elo jump from its own predecessor, Opus 4.5. It scored 80.8% on SWE-Bench Verified and 84.0% on BrowseComp (up from 67.8%). These are substantial gaps on the exact tasks enterprise customers are paying for.

Where Opus 4.6 falls short: Terminal-Bench 2.0, where it scored 65.4% against GPT-5.3-Codex’s 77.3%. That 11.9-point gap is real and significant — it reflects a genuine weakness in autonomous execution tasks. Anthropic optimized for reasoning depth, and the tradeoff shows.

API pricing is $5/$25 per million tokens (input/output), or $10/$37.50 above 200K context. Transparent, predictable, and notably consistent with Anthropic’s positioning: charge a premium, deliver measurably superior enterprise reasoning, and let the benchmarks justify the price tag.

GPT-5.3-Codex: OpenAI’s Speed and Autonomy Gambit

Here’s the line that matters: GPT-5.3-Codex is OpenAI’s first model that “helped build itself.” Early versions debugged their own training runs, managed deployment, and diagnosed evaluation results. Sam Altman called it “a sign of things to come.” To be clear — this is not recursive self-improvement. The model assisted human engineers; it didn’t autonomously create itself. But the trajectory is unmistakable. If models can accelerate their own development cycles, the compounding advantages get very real very fast.

The benchmarks reflect what self-training optimizes for: execution speed and autonomous task completion. GPT-5.3-Codex hit 77.3% on Terminal-Bench 2.0 (vs Opus 4.6’s 65.4%), 56.8% on SWE-Bench Pro, and 64% on OSWorld. These aren’t reasoning benchmarks — they’re action benchmarks. The model doesn’t just think about what to do. It does it, quickly and autonomously.

Add 25% faster inference over GPT-5.2-Codex, and speed becomes a compounding advantage for agents and real-time coding tools. Faster inference means faster agent iterations means more value extracted per token dollar spent. For teams building autonomous systems — CI pipelines, automated debugging, live code assistance — this speed gap translates directly into cost savings and user experience. GitHub Copilot general availability for GPT-5.3-Codex followed on February 9th.

At launch, GPT-5.3-Codex was available with paid ChatGPT plans across Codex surfaces: the app, CLI, IDE extension, and web. API access was not enabled at launch but OpenAI confirmed it’s being worked on — a notable restriction for a model positioned as developer-first.

The Cybersecurity Paradox: Same Power, Opposite Story

Both models can analyze code and find vulnerabilities at a deep level. What’s fascinating is how each company chose to frame that same capability.

Anthropic went offensive-defensive: Opus 4.6 discovered 500+ previously unknown zero-day vulnerabilities in open-source code using “out-of-the-box capabilities” — no specialized tooling, no custom prompting. Findings were validated by Anthropic’s red team and external security researchers. The message: AI should find bugs before attackers do. Enterprises can now audit their own codebases with a model that catches what human reviewers miss.

OpenAI went precautionary: GPT-5.3-Codex received the company’s first-ever “High” cybersecurity classification under the Preparedness Framework, triggering delayed API access. Their language: “We don’t have definitive evidence the new model can fully automate cyberattacks, but we’re taking a precautionary approach.” They committed $10 million in API credits to accelerate cyber defense — carrots alongside the stick of restricted access.

Neither company is wrong. Anthropic says: the capability exists, let’s weaponize it defensively. OpenAI says: the capability exists, let’s restrict access carefully. Two legitimate approaches to the same technological reality — and a preview of how the entire industry will navigate increasingly powerful dual-use AI systems.

So Which Model Wins?

Wrong question. The right question is: which problem are you solving?

If you’re in finance, law, or healthcare — industries where decisions require synthesizing thousands of pages of context from multiple perspectives — Opus 4.6’s agent teams and 1M context window aren’t just features. They’re architectural advantages that no amount of inference speed compensates for. The GDPval-AA dominance (144 Elo over GPT-5.2) directly measures the kind of complex reasoning these industries are willing to pay premium pricing for.

If you’re a development team optimizing for shipping speed, autonomous agents, and real-time coding assistance — GPT-5.3-Codex’s Terminal-Bench lead, 25% speed advantage, and self-training capabilities point toward a future where the model gets better at helping you ship without waiting for the next release. The self-training precedent alone changes the calculus on long-term platform bets.

This isn’t hedging. It’s the market finally segmenting. Anthropic is pursuing enterprise defensibility through depth. OpenAI is pursuing developer ubiquity through speed. Both are winning on their chosen metrics, and the Super Bowl ad war — where Anthropic mocked ChatGPT ads and Altman called their campaign “clearly dishonest,” adding that “Anthropic serves an expensive product to rich people” — shows both companies know the competition has moved well beyond benchmarks.

The Real Test Begins Now

February 5th, 2026 was a watershed moment — not because one model won, but because the AI industry stopped pretending there would be one model to rule them all. Anthropic bets depth beats speed. OpenAI bets speed compounds into dominance. The self-training milestone alone signals that the development cycles that shaped December’s model wars are about to accelerate in ways we haven’t fully mapped.

The days of a single dominant AI model are over. What replaces it is more competitive, more specialized, and requires actual decision-making about which platform bet aligns with your use case. For anyone building with these tools, that’s far more interesting — and far more consequential — than a simple horse race ever was.