AI news: March 2026

March 2026 was GPT-5.4’s month: OpenAI shipped its first general-purpose model with computer use built in and, two weeks later, the mini and nano variants to make support agents cheaper. Google pushed at the other end with Flash-Lite and its live voice model, and Mistral released two pieces for teams that want AI in-house: Voxtral TTS with open weights and Forge for training custom models.

GPT-5.4 arrives with native computer use and 1 million tokens

OpenAI launched GPT-5.4 on March 5 in Thinking and Pro variants, available in ChatGPT and the API from day one. It is the company’s first general-purpose model with native computer use, aimed at agents that operate desktop applications, and it debuts a context window of up to 1 million tokens, OpenAI’s largest to date. The company claims 33% fewer factual errors than GPT-5.2 and 83% on its internal GDPval test. API pricing: $2.50 per million input tokens and $15 per million output (above 272,000 input tokens the rate goes up); the Pro tier costs $30 and $180.

For agent builders, the two changes that touch architecture are computer use without bolt-on tools and the 1-million context: cases that used to require chunking or RAG now fit in a single call, at the long-context surcharge.

Source: OpenAI · also on TechCrunch and the system card

GPT-5.4 mini and nano: the budget tier for subagents

On March 17 OpenAI completed the family with GPT-5.4 mini (400,000-token context, text and image, more than twice as fast as GPT-5 mini and close to the full model on SWE-Bench Pro and OSWorld per OpenAI) at $0.75 per million input and $4.50 per million output, and GPT-5.4 nano, API-only, for classification, extraction and coding subagents, at $0.20 and $1.25: the cheapest model in OpenAI’s catalog. Both generally available from the announcement; mini also reaches ChatGPT’s free tier.

The design pattern this makes cheaper is the orchestrator with subagents: the big model decides and nano workers grind through the mechanical tasks at a fraction of the cost.

Source: OpenAI · also on Microsoft Tech Community and The New Stack

GPT-5.3 Instant: fewer disclaimers, fewer hallucinations per OpenAI

OpenAI began rolling out GPT-5.3 Instant on March 3 as ChatGPT’s default model and in the API as gpt-5.3-chat-latest. The stated goal is more direct conversation, with fewer warnings and filler sentences. In its internal high-risk evaluations, OpenAI reports 26.8% fewer hallucinations with web access and 19.7% fewer without it. No independent benchmark appears in the coverage we reviewed; all figures are the vendor’s.

It affects any product built on the chat-latest alias: response tone changes without touching code, so it is worth re-running your prompt regression tests.

Source: OpenAI · also on the system card and EdTech Innovation Hub

Gemini 3.1 Flash-Lite: Google’s volume tier

Google announced Gemini 3.1 Flash-Lite on March 3, in preview on AI Studio and Vertex AI. It is the fastest and cheapest model in the Gemini 3 series: 2.5x lower time to first token than Gemini 2.5 Flash and 45% higher output speed, with configurable thinking levels built in. Google cites 86.9% on GPQA Diamond and 76.8% on MMMU Pro. Pricing: $0.25 per million input tokens and $1.50 per million output.

It competes in the same bracket as GPT-5.4 nano: high-volume tasks (classify, extract, summarize) where cost per token beats peak capability. Still in preview, with no general-availability date.

Source: Google DeepMind

Codex Security: the repository-auditing agent, in research preview

OpenAI launched Codex Security on March 6 for ChatGPT Enterprise, Business and Education customers, first month free. An evolution of the internal “Aardvark” project, the agent analyzes a repository, writes a natural-language description of how the application works, tests candidate vulnerabilities in a sandbox to weed out false positives, ranks them by impact and proposes patches. OpenAI says a 30-day run against 1.2 million commits surfaced about 800 critical vulnerabilities and over 10,000 high-severity ones, with the false-positive rate halved since the initial release. The company’s own figures, with no external audit.

For teams with a security backlog it is a cheap second reviewer that arrives before the annual pentest; the verification sandbox is what separates it from a plain static scanner.

Source: OpenAI · also on AIBusiness and SecurityWeek

MiniMax M2.7, open and trained by optimizing itself

MiniMax published M2.7 with open weights on March 18. The distinctive part is the process: during training, the model autonomously ran over 100 rounds of optimization on its own scaffold, with a 30% gain on internal evaluations per MiniMax. The Artificial Analysis Intelligence Index, an external evaluator, ranks it first of 136 models with a score of 50 against a field average of 19; independent coverage also notes a slight step back from M2.5 on SWE-Bench Verified (around 78% versus 80.2%).

It is the open model to test this month for agentic workloads, with a level-headed reading of the numbers: it leads the external aggregate index and gives up a little on classic issue-solving.

Source: MiniMax · also on MarkTechPost and Build Fast with AI

Gemini 3.1 Flash Live: Google’s real-time voice, GA for consumers

Google introduced Gemini 3.1 Flash Live on March 26, its best voice model for real-time interaction: Google cites 90.8% on ComplexFuncBench Audio (multi-step function calling by voice) and double the conversation context of the previous version, with a SynthID watermark on all generated audio. It is generally available to the public in Gemini Live and in over 200 countries through Search Live; for developers it remains in preview via the Gemini Live API. Early adopters include Verizon, The Home Depot and LiveKit.

The voice function-calling number is the one that matters for business: phone and support agents that execute multi-step actions without losing the thread. Developer access remains in preview, so production integrations have to wait.

Source: Google DeepMind

Holo3 sets a desktop-agent record

H Company introduced Holo3 on March 31 in two variants: Holo3-122B-A10B (10B active parameters out of 122B) and Holo3-35B-A3B, the latter with open weights under Apache 2.0. Per H Company, the large one reaches 78.85% on OSWorld-Verified, the reference benchmark for desktop agents, a new state of the art; the open one scores 77.8%. The company claims to match or beat GPT-5.4 and Opus 4.6 on this task at a much lower inference cost, a claim without third-party verification. The API includes a free tier.

An open variant with 3B active parameters brushing the state of the art in computer use brings desktop agents closer to self-hosted deployments, without depending on the frontier APIs.

Source: H Company · also on Hugging Face and AIHola

Voxtral TTS: open text-to-speech with 5-second cloning

Mistral released Voxtral TTS on March 23: a 4B-parameter text-to-speech model covering 9 languages, with 70 milliseconds of model latency and voice cloning from 3-5 seconds of reference audio. In Mistral’s own human evaluation it beats ElevenLabs Flash v2.5 on naturalness with 68.4% preference; independent analyses point out that it is an internal evaluation, that ElevenLabs covers over 70 languages against 9, and that early users report unstable cloning. Open weights under a non-commercial license (CC BY-NC 4.0) and a paid API at $0.016 per thousand characters.

For voice prototypes with data that cannot leave the building it is the first serious open option; the non-commercial weights license pushes you to the API as soon as there is a product.

Source: Mistral AI · also on MarkTechPost and SiliconANGLE

Mistral Forge: custom model training as a service

At Nvidia’s GTC on March 17, Mistral introduced Forge: a system for enterprises to train models on their internal data, covering pretraining, post-training and reinforcement learning, with support for dense and MoE architectures. It is built with coding agents as the first user: Mistral Vibe can use Forge to tune models, search hyperparameters and generate synthetic data. Early-access partners include ASML, Ericsson, the European Space Agency and Singapore’s DSO. For now there is only an interest form, with no pricing or timeline.

It is the opposite bet to fine-tuning generic models: a model of your own built from company data. With no public pricing, the only actionable step today is joining the list if the case justifies it.

Source: Mistral AI · also on VentureBeat and TechCrunch