AI News March 2026: Claude Sonnet 4.6, Perplexity Computer, NVIDIA Nemotron & the OpenAI Fallout

March 2026 has been one of the most eventful months in AI since ChatGPT launched. A new model from Anthropic just redefined what a mid-tier AI can do. OpenAI is haemorrhaging users. NVIDIA dropped a fully open 120B model. And Perplexity launched a product that might be the closest thing yet to a genuine AI agent for everyday work. Here's the full picture.

Anthropic

OpenAI

NVIDIA

Perplexity

Anthropic: Claude Sonnet 4.6 Changes Everything

On 17th February 2026, Anthropic released Claude Sonnet 4.6 — and it immediately became the most talked-about model release of the year. What makes it significant isn't just raw benchmark performance, but what it represents: the blurring of the line between "efficient" and "frontier" AI.

Sonnet 4.6 ships with a 1 million token context window (in beta), adaptive thinking that dynamically decides when to reason step-by-step, and near-human-level performance on complex computer use tasks — navigating spreadsheets, filling multi-step web forms, and executing real office workflows. Early developer testing found it outperforming even Claude Opus 4.5 (released November 2025) on many real-world tasks.

Sonnet 4.6 is now the default model for all Claude.ai Free and Pro users, which matters enormously for adoption. It's also now generally available in GitHub Copilot — putting it directly in the hands of millions of developers daily.

Anthropic's Revenue Surge: $19 Billion ARR

Behind the model releases, Anthropic's business momentum is staggering. The company's annualised run-rate revenue hit $19 billion in early March 2026 — more than doubling from $9 billion just three months prior. The primary driver? Claude Code, Anthropic's agentic coding tool, which alone reached a $2.5 billion run-rate.

Enterprise adoption is the bedrock of this growth. Anthropic serves over 300,000 business customers, with the number of companies spending over $100,000 annually growing sevenfold year-on-year. The biggest deployment: Deloitte rolling out Claude to approximately 470,000 employees. In February 2026, the company raised a $30 billion Series G at a $380 billion valuation — cementing its position as the most credible rival to OpenAI in the enterprise market.

OpenAI: GPT-5.4, the Pentagon Deal & the QuitGPT Exodus

March 2026 has not been kind to OpenAI. The company launched GPT-5.4 on 5th March — just two days after GPT-5.3 — in what observers described as a desperate attempt to contain a growing PR crisis. The trigger: a Pentagon contract announcement on 28th February 2026 offering the US Department of Defense access to OpenAI's models for "any lawful purpose." Anthropic, notably, had refused an identical deal hours earlier.

The backlash was immediate and severe. ChatGPT uninstalls spiked 295% in the days that followed. The hashtag #QuitGPT trended globally. Approximately 1.5 to 2.5 million users cancelled subscriptions or joined the boycott. Within days, Anthropic's Claude became the #1 free app on Apple's App Store.

ChatGPT's market dominance has been eroding for months. From a peak of roughly 90% market share in early 2025, OpenAI has dropped to around 70% by March 2026. Among enterprise AI spend tracked by Ramp's corporate data, OpenAI's adoption slipped from 36.8% to 35.9% in February 2026, while Anthropic jumped from 16.7% to 19.5% in the same month — one of the largest single-month gains recorded. At current growth rates, analysts project revenue parity between the two companies by mid-2026.

Perplexity: "Computer" — The Multi-Model AI Agent

Perplexity made arguably the most architecturally interesting product announcement of the month with the launch of "Computer" on 2nd March 2026. Rather than building a better single AI, Perplexity built an orchestration layer — a system that breaks complex user goals into tasks and subtasks, then assigns each to the best-suited AI model for that specific job.

Computer can run workflows autonomously for hours — or months — without user intervention, handling web research, document generation, data processing, and API calls to connected services including Gmail, Slack, GitHub, Notion, and Salesforce. Each task runs in an isolated compute environment with a real filesystem and browser access.

Two weeks after launch, Perplexity followed up with "Personal Computer" — software that turns a Mac Mini into a persistent, always-on AI assistant. The model processing happens on Perplexity's servers, but the agent has full access to local files and applications, enabling genuinely autonomous task execution in the background. Also launched: Model Council, which runs queries across multiple AI models simultaneously and synthesises agreement and disagreement — useful for professional research and high-stakes decision-making.

Computer is available now to Perplexity Max subscribers ($200/month), with Enterprise Max access coming soon.

NVIDIA: Nemotron 3 Super — Open-Source, 120B, Fully Free

While the AI labs battled it out on consumer mindshare, NVIDIA quietly dropped what may be the most significant open-source model release of 2026: Nemotron 3 Super. At 120 billion parameters (with only 12 billion active at inference time via Mixture-of-Experts), it achieves the throughput of a much smaller model while delivering frontier-class reasoning.

The model's native 1 million token context window makes it particularly well-suited for agentic applications requiring long-term memory. NVIDIA released it under a fully permissive open licence including weights, datasets, and training recipes — meaning any developer or business can run, fine-tune, or build on it without restriction.

For companies that can't or won't send data to closed API providers, Nemotron 3 Super represents a genuinely competitive self-hosted alternative. It's optimised for NVIDIA's Blackwell hardware but available in multiple formats including BF16, FP8, and NVFP4 for broader compatibility.

What This All Means for Businesses

March 2026 is a snapshot of an industry in rapid transition. A few themes stand out:

• The gap between "best" and "good enough" is closing fast. Sonnet 4.6 doing Opus-level work at Sonnet pricing means the cost of high-quality AI output is collapsing.
• Multi-model orchestration is the next frontier. Perplexity's Computer shows the direction of travel — not one model to rule them all, but specialised models deployed intelligently for each task.
• OpenAI's dominance is no longer guaranteed. The Pentagon controversy accelerated a shift that was already underway. Brand trust is now a competitive variable in AI.
• Open-source is closing the gap with closed models. NVIDIA's Nemotron 3 Super is the most capable open model yet — and freely deployable on your own infrastructure.
• Enterprise AI spend is concentrating. Anthropic's 7x growth in $100k+ customers isn't random — it reflects a flight to quality, safety reputation, and better coding tools.

Stay Ahead With Reactively

AI is moving faster than most teams can track — and the strategic implications for search, content, and digital marketing are significant. At Reactively, we're embedding these tools into real workflows for our clients every month. If you want to talk about what AI means for your marketing and SEO strategy in 2026, get in touch.