AI news: February 2026

February 2026 revolved around agentic coding. OpenAI stacked three Codex releases in two weeks, starting with GPT-5.3-Codex, and Google answered at the high end with Gemini 3.1 Pro and in image generation with Nano Banana 2. In between, MiniMax lowered the price floor again and OpenAI unveiled Frontier, its enterprise agent platform.

GPT-5.3-Codex lands across the whole Codex surface

OpenAI released GPT-5.3-Codex on February 5 for paid ChatGPT plans in the app, CLI, IDE extension and web; API access was left for “the coming weeks”. Per OpenAI, it is 25% faster than GPT-5.2-Codex and uses fewer tokens for the same results. The big jumps are in terminal and computer use: 77.3% on Terminal-Bench 2.0 (13 points over its predecessor) and 64.7% on OSWorld-Verified (26 points up); on SWE-Bench Pro the gain is 0.4 points. Benchmarks are the vendor’s own.

For teams already running coding agents, the useful detail is where it improves: terminal work and full computer use, not classic issue-solving, which barely moves.

Source: OpenAI · also on Neowin and Digital Applied

GPT-5.3-Codex-Spark: real-time coding on Cerebras hardware

A week later, on February 12, OpenAI introduced GPT-5.3-Codex-Spark, a slimmed-down variant and its first real-time coding model: over 1,000 tokens per second, a 128k context window, text-only for now. It runs on Cerebras’ Wafer Scale Engine 3, the first output of that partnership. By default it makes minimal, targeted edits and does not run tests unless asked. It is a research preview for ChatGPT Pro users only, and usage does not count against standard Codex limits.

Near-zero latency changes how the tool is used: edits as you type instead of tasks you queue. In exchange the model is less thorough: OpenAI positions it for quick edits, with the larger Codex models still covering extensive refactors.

Source: OpenAI · also on Help Net Security and Cerebras

The Codex app arrives on macOS for every plan

OpenAI opened the month on February 2 with the Codex app for macOS, downloadable on any ChatGPT plan, free tier included for a limited time. It manages multiple agents in per-project threads, with background computer use (the agent sees, clicks and types with its own cursor while you keep working), a built-in browser, worktree support so several agents can touch the same repository, scheduled automations and SSH connections to remote devboxes. No Windows or Linux version was announced.

Orchestrating several agents in parallel no longer requires custom tooling; the app supplies the management layer itself.

Source: OpenAI · also on 9to5Mac and VentureBeat

Gemini 3.1 Pro doubles its predecessor on ARC-AGI-2

Google launched Gemini 3.1 Pro on February 19 in preview for developers (Gemini API, AI Studio, Vertex AI) and consumers (Gemini app, NotebookLM). The number that drove the conversation: 77.1% on ARC-AGI-2, externally verified, more than double Gemini 3 Pro’s 31.1%. Technical coverage places it ahead of Anthropic’s and OpenAI’s frontier models on that test, at a preview price under half of Anthropic’s equivalent, and praises the practical gains in SVG generation, UI work and code, with some skepticism about benchmark-chasing.

For anyone picking models per task, the relevant move is price per capability: high-end reasoning at mid-tier cost, still in preview and with no general-availability date.

Source: Google DeepMind · also on Latent.Space and Zvi Mowshowitz

Nano Banana 2: Google’s image model, faster and at half the price

On February 26 Google released Nano Banana 2 (technically Gemini 3.1 Flash Image): precise text rendering, consistency for up to 5 characters and 14 objects per image, resolution from 512px to 4K, and a simultaneous rollout across the Gemini app, Search, Flow, Google Ads and the APIs (in preview). It reached 141 new countries. Per the external evaluator Artificial Analysis, it tops text-to-image rankings and costs about $67 per thousand images, half of GPT Image 1.5 or Nano Banana Pro.

For high-volume creative production (product shots, ads, per-language variants), the halved price with legible in-image text is the argument; the cost figures come from the external evaluator, not Google.

Source: Google DeepMind · also on Deevid AI and Latent.Space

MiniMax M2.5 pushes agent pricing further down

MiniMax released M2.5 and M2.5-Lightning on February 12. The company cites 80.2% on SWE-Bench Verified and speed comparable to Western frontier models, priced at $0.30 per million input tokens and $2.40 per million output for Lightning (M2.5 costs half that), plus hourly pricing: $1 per hour at a sustained 100 tokens per second. Per MiniMax, that works out 10 to 20 times cheaper than the top models from Anthropic, Google and OpenAI. No independent verification of these benchmarks has surfaced yet: every figure is the vendor’s.

If those numbers hold up, long-running agentic workloads get noticeably cheaper. Since no external verification exists yet, they need validating in your own tests before migrating anything.

Source: MiniMax

Gemini makes music: Lyria 3 reaches the app

Google integrated Lyria 3 into the Gemini app on February 18: 30-second tracks with cover art from text or photos, auto-generated lyrics if you don’t provide them, and control over style, vocals and tempo. It ships in 8 languages, restricted to users 18 and over, with a SynthID watermark on every track, filters against existing material and a rights-reporting mechanism. Paid subscribers get higher limits, with no published figures.

For content teams this is instant scratch audio (jingles, backgrounds for short video) with watermark traceability handled out of the box. The 30-second cap per track keeps its use to mockups.

Source: Google DeepMind

Gemini 3 Deep Think, updated for science and engineering

On February 12 Google updated Deep Think, its long-reasoning mode: 84.6% on ARC-AGI-2 (verified by the ARC Prize Foundation), 48.4% on Humanity’s Last Exam without tools, a 3455 Codeforces Elo and gold-medal level on the 2025 international olympiads for math, physics and chemistry. Google shows concrete examples: reviewing mathematical papers at Rutgers, optimizing crystal growth at Duke, and turning sketches into 3D-printable files. Available in the app only for Google AI Ultra subscribers; API access remains in beta behind a request form.

The usage profile is specific: science and engineering problems that can wait minutes for a better answer. Beyond the verified ARC-AGI-2 score, the benchmarks are Google’s.

Source: Google DeepMind

Hedra Omnia unifies character video in a single model

Hedra launched Omnia on February 5: one model that processes vision, text and audio together to generate clips of up to 8 seconds at 1080p, instead of chaining separate systems for image, motion and sound. Audio does more than sync lips: it drives the pacing and expressiveness of the whole clip, with careful details like natural blinking, stable hands and logos that don’t warp. Generally available from day one on Hedra’s platform with free signup, aimed at UGC, video podcasts and branded content.

For presenter or avatar video it is a jump in perceived quality with no workflow change: the same prompt produces a clip with fewer visible artifacts.

Source: Hedra · also on Evolution AI Hub

OpenAI Frontier: an enterprise platform for managing agents

Also on February 5, OpenAI introduced Frontier, its enterprise platform for building, deploying and managing agents: agent onboarding, a feedback loop styled after performance reviews, and support for agents built outside OpenAI, connected to external data and applications. Early customers cited: HP, Oracle, State Farm and Uber. No pricing was disclosed and access is limited at launch; broader rollout was announced for the following months.

The proposal of treating agents as managed operating capacity, with a lifecycle and evaluation, is the part of the announcement with the longest reach. Restricted access and missing pricing make it impossible to evaluate yet.

Source: OpenAI · also on TechCrunch and InfoQ