Which new AI models from the first half of March 2026 should developers know about?

In the first half of March 2026, OpenAI released GPT-5.4 (5 March) with a context window of about a million tokens, tool search and native computer use. On 13 March Anthropic made the 1-million-token context window of Claude Opus 4.6 and Sonnet 4.6 generally available at standard prices. And the Covenant-72B paper (9 March) showed that a 72-billion-parameter model can be pre-trained over the internet by participants who needed no permission to join, with scores close to Meta's LLaMA-2-70B from 2023.

Illustration: a robot reading a tablet in a library, with mice playing on a shelf — GPT-5.4 looks up a tool's definition only when it needs that tool.

What does GPT-5.4 bring?

OpenAI released GPT-5.4 on 5 March 2026 in ChatGPT (as GPT-5.4 Thinking), in the API and in Codex, together with GPT-5.4 Pro for the hardest tasks. According to The Batch, the models accept up to 1,050,000 tokens of input and return up to 128,000, with reasoning effort that you can set up to xhigh. API prices in March 2026 were $2.50 per million input tokens and $15 per million output tokens for GPT-5.4, and $30 and $180 for GPT-5.4 Pro. The Batch places that at the top of the market.

For developers, two features stand out. With tool search, the model no longer gets every tool definition in the prompt upfront: it receives a short list of available tools and looks up a definition at the moment it needs that tool. For systems with many tools, OpenAI says, this saves thousands of tokens per request and keeps the prompt cache intact. The second feature is computer use. According to OpenAI, GPT-5.4 is its first general-purpose model with native computer use: it operates software by writing code, for example with Playwright, or with mouse and keyboard commands based on screenshots. On OSWorld-Verified, a test of desktop tasks, OpenAI reports 75.0%, against 47.3% for GPT-5.2 and a human baseline of 72.4%.

How does GPT-5.4 compare in independent tests?

In tests by Artificial Analysis, reported by The Batch, GPT-5.4 Pro at xhigh reasoning had the top score on GDPval-AA, BrowseComp, Terminal-Bench Hard, SWE-Bench Pro and MCP Atlas. On the Artificial Analysis Intelligence Index, a weighted average of ten benchmarks, it scored 57 points, just behind Gemini 3.1 Pro Preview at 57.2. It led the coding and agentic parts of that index with 57 and 69 points. Running the full index cost $2,950 with GPT-5.4 Pro, against $892 with Gemini 3.1 Pro Preview.

What changed for long context in Claude?

On 13 March 2026 Anthropic made the 1-million-token context window of Claude Opus 4.6 and Sonnet 4.6 generally available. Standard prices now apply across the whole window: $5 per million input tokens and $25 per million output tokens for Opus 4.6, and $3 and $15 for Sonnet 4.6. A request of 900,000 tokens costs the same per token as one of 9,000. During the beta, Opus 4.6 cost $10 per million input tokens and $37.50 per million output tokens for prompts above 200,000 tokens, according to Anthropic's launch post of 5 February 2026. Long requests no longer need a beta header, the normal rate limits apply at every length, and one request can hold up to 600 images or PDF pages, up from 100.

Illustration: four people around a large open book covered in tiles, with gears and a clock — Since 13 March 2026, a long Claude request costs the same per token as a short one.

Simon Willison pointed to the difference with competitors: Google charges more above 200,000 tokens for Gemini 3.1 Pro, and OpenAI above 272,000 tokens for GPT-5.4. Below those thresholds the list prices are close: $2.50 per million input tokens for GPT-5.4 against $3 for Sonnet 4.6, and $15 for output for both.

What does Covenant-72B show about decentralized training?

Covenant-72B (Lidin et al., 9 March 2026) is a language model of 72 billion parameters, pre-trained over ordinary internet connections by participants who needed no permission to join. Coordination software called Gauntlet, running on the Bittensor blockchain, scored every contribution and rewarded the participants. Each peer ran at least eight NVIDIA B200 GPUs. Per round at most 20 peers contributed (16.9 on average), and at least 70 different peers took part over the run, which covered about 1.1 trillion tokens. The weights are on Hugging Face under the Apache 2.0 license.

In the authors' comparison, the base model scores 67.1 on MMLU, against 65.6 for Meta's LLaMA-2-70B from 2023, which was trained on 2 trillion tokens in a data centre. The chat version scores 67.4 on MMLU and 26.3 on MATH. The authors conclude that open, permissionless pre-training is feasible at this scale. Jack Clark of Import AI adds a caveat: by his estimate the run used about 160 chips, while frontier models are trained on tens to hundreds of thousands.