# Frontier AI models: March 2026

> OpenAI answered with GPT-5.4 and its 'Thinking' mode, extending frontier reasoning into finance, law and strategy work.

## Direct answer

OpenAI released GPT-5.4 on March 5, 2026, introducing 'Thinking' capabilities - the ability to reason through a multi-step problem before answering. Anthropic followed with Claude Cowork on March 24, an agentic memory layer for teams. March 2026 is the month deliberate reasoning reached the frontier, which mattered most to financial modelling, legal analysis and strategic planning rather than to coding.

## Released in March 2026

| Date | Model | Lab | Capability | Effect on automation |
| --- | --- | --- | --- | --- |
| 5 March 2026 | GPT-5.4 | OpenAI | Flagship 'Thinking' mode - deliberate multi-step reasoning | A step change for complex knowledge work: financial modelling, legal analysis and strategic planning became partially automatable at a quality that was out of reach a year earlier. |
| 24 March 2026 | Claude Cowork | Anthropic | Agentic memory layer carrying context across a team | Shifted AI from a per-person tool to shared team infrastructure, which is the precondition for automating coordination work rather than individual tasks. |

## Leading models as of March 2026

| Lab | Leading model | Strongest at |
| --- | --- | --- |
| OpenAI | GPT-5.4 | Deliberate multi-step reasoning |
| Anthropic | Claude Opus 4.6 | Agentic coding and multi-file refactors |
| Google | Gemini 3.1 Pro | Multimodal reasoning and throughput |

## What this month changed for careers

Until March, the safest description of AI's limit was 'it cannot think through a problem, only pattern-match one'. GPT-5.4 retired that sentence. Roles whose defensibility rested on analytical reasoning - analyst, associate, strategist - lost their cleanest argument in March 2026.

## Questions

### What AI models were released in March 2026?

GPT-5.4 from OpenAI on March 5, 2026, and Claude Cowork from Anthropic on March 24. GPT-5.4 was the month's frontier release.

### What is GPT-5.4 Thinking mode?

'Thinking' is GPT-5.4's ability to reason through a multi-step problem internally before producing an answer, rather than responding in a single pass. It raised performance most on tasks with several dependent steps, such as financial modelling and legal analysis.

---

**Source:** Job Security Meter — https://jobsecuritymeter.com/ai-model-releases/march-2026
**Human-readable version:** https://jobsecuritymeter.com/ai-model-releases/march-2026
**Last updated:** March 2026
**Attribution:** Free to quote and cite with attribution to Job Security Meter.
**Full site specification for language models:** https://jobsecuritymeter.com/llms-full.txt