Claude Opus 4.6 and GPT-5.3-Codex land on the same day

What shipped

Anthropic

Claude Opus 4.6

Anthropic's new top model, stronger at long agentic work and offering a 1M-token context window in beta.

OpenAI

GPT-5.3-Codex

OpenAI's latest Codex model, released the same day as Opus 4.6 and aimed at long-running coding tasks.

Alibaba Qwen

Qwen3-Coder-Next

An open coding model built for agents, light on active parameters so it can run on modest hardware.

StepFun

Step 3.5 Flash

StepFun's fast open reasoning and agent model.

Mistral

Voxtral Transcribe 2

A new speech-to-text family from Mistral, including a real-time model with open weights.

OpenAI

Codex changelog (Codex app for macOS)

OpenAI launched a desktop Codex app for running several coding agents side by side.

Thursday 5 February was the busiest launch day I've logged yet.

Opus 4.6 against GPT-5.3-Codex

Anthropic released Claude Opus 4.6 and OpenAI released GPT-5.3-Codex within hours of each other. I do not think that was an accident. Both are aimed at the same buyer, the person who wants an agent to take a big piece of work and bring it back finished.

Opus 4.6 is Anthropic's strongest model so far, and the feature I noticed first is the 1M-token context window, available in beta. For my kind of work that is the difference between feeding an agent one chapter of a tender and feeding it the whole pack. Anthropic also says the model plans more carefully and keeps going longer on agentic tasks.

GPT-5.3-Codex pushes the same direction from OpenAI's side. It is a coding model, but coding models are doing more than code now. They read files, run commands, and fix their own mistakes, which is most of what an operations task looks like.

I will not rank them off launch-day benchmarks. I have seen too many numbers shift when the work gets real. My plan is the usual one: run both on the same internal task, compare the output, and keep whichever one wastes less of my time.

Agent tooling around the models

OpenAI also launched a Codex desktop app for macOS, built for running several agents in parallel and reviewing what each one did. Mistral released Voxtral Transcribe 2, with an open-weights real-time model for live transcription.

Open weights

Qwen3-Coder-Next is a coding model that activates few parameters per token, so it runs on smaller machines. StepFun's Step 3.5 Flash is another fast open model aimed at reasoning and agents. The open side keeps pace on cost even when it trails on the very top scores.

A word on the comparison trap. Launch-day coverage turns into a scoreboard within hours, and I understand why, but my own decisions run on three numbers: how often the agent finishes without me, how much I have to fix, and what it costs per finished job. A model that is two points better on a public benchmark and twice as expensive per job can still be the wrong choice. I track these quietly and let the noise pass.

What I'd watch next: independent comparisons of Opus 4.6 and GPT-5.3-Codex on long, messy, real-world tasks rather than benchmark suites.