Anthropic and OpenAI Ship Opus 4.6 and GPT-5.3-Codex on the Same Day, February 2026

On 5 February 2026 the two leading frontier labs released competing flagship models within hours of each other. Anthropic launched Claude Opus 4.6, available on claude.ai, the API and major cloud platforms at 5 dollars per million input tokens and 25 dollars per million output tokens, with higher rates above 200,000 tokens. The model introduces a one-million-token context window in beta, output of up to 128,000 tokens, "adaptive thinking" that decides when extended reasoning is worthwhile, and automatic compaction of older context. In Claude Code it adds agent teams, a research preview in which several agents split a task and coordinate with each other rather than working in sequence. Anthropic reports the top score on Terminal-Bench 2.0 and a lead on Humanity's Last Exam, along with a rise to 76 percent on the MRCR v2 long-context test from 18.5 percent for its predecessor. It also says the model has the lowest over-refusal rate of any recent Claude release.

The same day OpenAI introduced GPT-5.3-Codex, which it calls its most capable agentic coding model. OpenAI cites 56.8 percent on SWE-Bench Pro, 77.3 percent on Terminal-Bench 2.0 and 64.7 percent on OSWorld-Verified, and says the model is about 25 percent faster than its predecessor. Notably, OpenAI states that GPT-5.3-Codex is "the first model we classify as High capability" for cybersecurity tasks under its preparedness framework. That classification triggers safety training, automated monitoring and a new Trusted Access for Cyber programme, alongside 10 million dollars in API credits for defensive research. The model shipped in the Codex app, CLI, IDE extension and web for paid plans, with API access to follow.

The convergence is striking. Both labs now sell the same proposition: models that run long, multi-step engineering and analysis tasks with limited supervision, with parallel agents as the next step. Anthropic's Opus 4.5 had appeared only in November 2025; the three-month cycle between flagship releases is now the norm. Benchmark comparisons are difficult because each vendor selects its own tests, and the two announcements overlap on only one headline benchmark.

For buyers, the more important developments are structural. A million-token context and 128,000-token outputs make it feasible to hand entire repositories or document sets to a model in one call, which expands both capability and the surface exposed to prompt injection. OpenAI's cyber classification is an explicit acknowledgement that offensive capability now comes bundled with coding skill.

What it means for leaders

• Update threat models for agent teams. Multiple coordinating agents with tool access require per-agent scoping, spend limits and audit trails; the OWASP LLM Top 10 items on excessive agency and insecure output handling apply directly.
• Treat cyber-capability ratings as procurement inputs. Ask vendors how "High capability" classifications affect monitoring, access controls and incident reporting for your tenancy.
• Re-evaluate data exposure at million-token scale. Larger contexts mean more sensitive material per request; confirm retention, logging and regional processing terms before enabling them.
• Standardise evaluation across vendors. With releases every quarter, maintain an internal benchmark suite tied to your use cases rather than reacting to vendor charts.
• Document human oversight for autonomous coding. For regulated software, code review, provenance and change-management evidence remain mandatory regardless of how the code was produced.