AI news: April 2026

In April 2026 the release cycle got shorter: GPT-5.5 arrived barely six weeks after GPT-5.4, at twice the price and aimed at agents. Meanwhile, open weights had their best month in a long while: Gemma 4 in five sizes, DeepSeek-V4 with a million-token context, Kimi K2.6 and NVIDIA’s multimodal Nemotron 3 Nano Omni.

GPT-5.5: six weeks later, twice the price

OpenAI launched GPT-5.5 on April 23, available in ChatGPT and the API from day one, pitched as a model “for real work and agents”. The company’s figures: 73.1% on its internal Expert-SWE eval (versus GPT-5.4’s 68.5%) and 78.7% on OSWorld-Verified; technical press adds 82.7% on Terminal-Bench 2.0. It costs $5 per million input tokens and $30 per million output, double GPT-5.4, which remains available; the Pro variant, for long-horizon research and whole-codebase refactors, goes up to $30 and $180.

This is now a pricing decision rather than a straight upgrade: GPT-5.4 remains available at half the cost, so it is worth measuring whether the agentic gains justify double the spend per use case.

Source: OpenAI · also on the system card and Tech Insider

Gemma 4: five open sizes under Apache 2.0

Google released Gemma 4 on April 2: five variants (multimodal E2B and E4B with image, audio and video; a unified 12B; a 26B MoE with 4B active; and a dense 31B), all Apache 2.0, with context up to 256K and support for over 140 languages. Per Google, the 31B ranks third on LMArena’s text leaderboard (Elo 1452) and the 26B MoE ranks sixth while using only 4B active parameters. Downloadable from day one on Hugging Face, Ollama, vLLM and llama.cpp.

For self-hosted infrastructure, the 26B MoE is the interesting piece: high-tier performance at small-model inference cost. The leaderboard positions are Google’s citation, though LMArena is publicly verifiable.

Source: Google · also on Hugging Face

ChatGPT Images 2.0: the first image model that reasons

OpenAI launched gpt-image-2 on April 21, its third flagship image model and the first with an explicit thinking mode: it plans the composition, can search the web for references and reviews its own output before generating. It also improves multilingual text with character-level accuracy in non-Latin scripts (Japanese, Korean, Chinese, Hindi, Bengali). In ChatGPT, Thinking mode is reserved for paid plans; in the API, a 1024x1024 image runs from $0.006 at low quality to $0.211 at high.

For creatives with in-image text across languages, character-level accuracy removes the manual touch-up that used to force a round trip through design tools; the reasoning mode is slower and pricier, so it is not the default for everything.

Source: OpenAI · also on MindWired and WaveSpeed

DeepSeek-V4: a million-token context at lower inference cost

DeepSeek released two MoE models in preview on April 24: V4-Pro (1.6 trillion parameters, 49B active) and V4-Flash (284B, 13B active), both with a one-million-token window. The architectural novelty is compressed attention in alternating layers (4x and 128x): per DeepSeek, at a million tokens V4-Pro needs 27% of the inference FLOPs and 10% of the KV cache of V3.2. The Register reports pricing of $0.14 per million input tokens on Flash and $1.74 on Pro, and that the models were validated on both Nvidia GPUs and Huawei Ascend accelerators. DeepSeek’s own agent figures: 80.6 on SWE-Bench Verified. No multimodality at launch.

The headline is the cost of long context: a million tokens is no longer a premium-tier luxury. For workloads that analyze whole documentation sets or repositories, the open checkpoints on Hugging Face let you test without going through the Chinese API.

Source: DeepSeek on Hugging Face · also on The Register

xAI opens its voice stack: STT/TTS APIs and Grok Voice Think Fast

xAI shipped its voice stack via API in two steps. On April 17 it released two standalone audio APIs, speech-to-text and text-to-speech, built on the same stack that already powers Grok Voice, Tesla vehicles and Starlink support, with expressive voices and multilingual coverage. On April 23 it added grok-voice-think-fast-1.0, its flagship voice model, aimed at multi-step customer-support and sales workflows; xAI describes it as handling precise data entry and real-time reasoning at no added latency, and says it already serves Starlink’s customers.

For teams building phone or voice agents, xAI enters as a direct competitor to Google’s and ElevenLabs’ voice APIs, with a model built for the tool-calling back-and-forth of support work.

Source: xAI STT/TTS · also on Grok Voice Think Fast 1.0

Nemotron 3 Nano Omni: text, image, video and audio in one open model

NVIDIA published Nemotron 3 Nano Omni on April 28, an open 30B model (3B active) with a hybrid Mamba-Transformer-MoE architecture that handles text, image, video and audio with a 262K-token context, enough for over 5 hours of audio. NVIDIA places it first in document understanding (MMLongBench-Doc, OCRBenchV2) and in video/audio comprehension among open omni models, with comparison figures from its own benchmark. Weights, datasets and training recipes are published; it is free on OpenRouter and supported in Ollama, vLLM and llama.cpp.

For multimodal pipelines on your own infrastructure (long meeting transcription and understanding, video, scanned documents) it is the month’s most complete open option, with the caveat that the comparisons are the vendor’s.

Source: NVIDIA · also on Hugging Face

Kimi K2.6: open weights for long-horizon coding

Moonshot AI launched Kimi K2.6 on April 20: a 1-trillion-parameter MoE (32B active), 256K context, focused on long-horizon coding and multi-agent orchestration. Moonshot cites 80.2% on SWE-Bench Verified and 89.6% on LiveCodeBench v6, and describes a 12-hour task with over 4,000 tool calls. The license is modified MIT with a branding clause: products with over 100 million monthly users or $20 million in monthly revenue must display “Kimi K2.6” visibly. On Hacker News, some assessments place it below Anthropic’s models in general capability.

At $0.60 per million input tokens, few models match its cost-to-capability ratio for long, mechanical coding tasks, though the community counterpoint and the branding clause are worth weighing before adopting it.

Source: Moonshot AI · also on Hugging Face

GPT-Rosalind: OpenAI’s model for life sciences

OpenAI introduced GPT-Rosalind on April 16, a reasoning model specialized in drug discovery, genomics and protein work, named after Rosalind Franklin. It queries specialized databases, analyzes literature and proposes experimental paths; it ships with a free Codex plugin connecting over 50 scientific tools and data sources. Collaborators cited: Amgen, Moderna, the Allen Institute and Thermo Fisher. It is a research preview, restricted to OpenAI’s trusted-access program.

It signals a shift toward discipline-specific vertical models, but access is gated: for biotech teams, the practical step today is applying to the access program and evaluating the Codex plugin.

Source: OpenAI · also on Euronews and Pharmaphorum

Gemini 3.1 Flash TTS: expressive speech with control tags

Google launched Gemini 3.1 Flash TTS on April 15, in preview: speech synthesis with audio tags controlling style, pacing and delivery, native multi-speaker dialogue, over 70 languages and a SynthID watermark. Voice parameters can be exported as Gemini API code to keep the same voice across projects. Artificial Analysis, an external evaluator, gives it a 1,211 Elo on its TTS benchmark, in the best quality-to-cost quadrant. No public pricing yet.

Voice-as-code export solves an operational problem: keeping a brand voice consistent across products. It remains in preview without published pricing, so for now it can only be tested.

Source: Google

Muse Spark: Meta’s multimodal reasoning model

Meta Superintelligence Labs announced Muse Spark on April 8: a natively multimodal reasoning model with tool use and multi-agent orchestration. Meta claims capabilities equivalent to Llama 4 Maverick with more than an order of magnitude less compute, and cites 58% on Humanity’s Last Exam with the Contemplating mode on; all figures are its own. It already runs on meta.ai and the Meta AI app, with the API in private preview and a rollout announced for WhatsApp, Instagram and Messenger in the following weeks.

Users will get this through Meta’s mass distribution, with the API still in private preview. For product teams, the immediate effect lies in what their users will be able to do inside WhatsApp and Instagram.

Source: Meta AI · also on Meta Newsroom