Exceptional day in AI: Anthropic announces a 380 billion, OpenAI deploys a real-time coding model at 1000+ tokens/second on Cerebras, and Google pushes Gemini 3 Deep Think to new records in sciences. On the ecosystem side, xAI restructures its teams around 4 products, Mistral invests €1.2 billion in Sweden, and GitHub temporarily suspends GPT-5.3-Codex for reliability reasons.
Anthropic raises $30B in Series G
February 12 — Anthropic announces a 380 billion. The round is co-led by GIC and Coatue, with participation from D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, and MGX. The list of additional investors includes Sequoia Capital, BlackRock, Goldman Sachs, Fidelity, Microsoft, and NVIDIA among others.
| Metric | Value |
|---|---|
| Amount raised | $30 billion |
| Post-money valuation | $380 billion |
| Annualized revenue (run-rate) | $14 billion |
| Annual growth | 10x per year (last 3 years) |
| Clients >$100k/year | 7x growth in one year |
| Enterprise Clients >$1M/year | Over 500 (vs 12 two years ago) |
| Fortune 10 Clients | 8 out of 10 |
| Claude Code run-rate | $2.5 billion (doubled since January 2026) |
| Claude Code weekly users | 2x since January 1st, 2026 |
| GitHub commits via Claude Code | 4% (doubled in 1 month) |
| Claude Code business subscriptions | 4x since start of 2026 |
| Claude Code enterprise share | >50% of Claude Code revenue |
The funds will be used for fundamental research, product development, and infrastructure expansion. Claude remains the only frontier AI model available on the three major cloud platforms: AWS (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry). Anthropic trains and runs Claude on a diverse range of AI hardware (AWS Trainium, Google TPUs, NVIDIA GPUs).
This fundraising reflects the incredible demand we are seeing from these customers, and we will use this investment to continue building the enterprise-grade products and models they have come to depend on. — Krishna Rao, CFO of Anthropic (@AnthropicAI on X)
GPT-5.3-Codex-Spark: real-time model on Cerebras
February 12 — OpenAI launches GPT-5.3-Codex-Spark in research preview, a real-time coding model optimized for ultra-fast inference on the Cerebras Wafer-Scale Engine 3. This is the first OpenAI model deployed on non-NVIDIA hardware.
Codex-Spark is a compact version of GPT-5.3-Codex, designed for interactive work where latency matters as much as intelligence: targeted edits, logic refactoring, interface adjustments with immediate feedback.
| Specification | Detail |
|---|---|
| Tokens/second | > 1,000 (vs 50-100 for standard models) |
| Context window | 128k tokens |
| Format | Text only (multimodal planned later) |
| Client/server overhead reduction | 80% (Persistent WebSocket) |
| Reduction overhead per token | 30% |
| Reduction time-to-first-token | 50% |
The model runs on the Cerebras Wafer-Scale Engine 3, a monolithic processor that maintains tightly coupled compute for high-throughput inference. This launch marks the first step of the OpenAI-Cerebras partnership announced in January 2026.
Codex-Spark is available in research preview for ChatGPT Pro subscribers, in the latest versions of the Codex app, CLI, and VS Code extension. API access is limited to selected partners, with broader deployment planned.
On SWE-Bench Pro and Terminal-Bench 2.0, the model shows solid performance by completing tasks in a fraction of the time compared to GPT-5.3-Codex. Separate rate limits apply during the research preview phase.
Eventually, Codex will offer two complementary modes: long-horizon reasoning (autonomous tasks of several hours/days) and real-time collaboration (rapid iteration). Both modes will be able to combine, with delegation of long tasks to sub-agents in the background.
Gemini 3 Deep Think: record scores in sciences
February 12 — Google announces a major update to Gemini 3 Deep Think, its specialized reasoning mode. The model is now designed to solve concrete challenges in science, research, and engineering, beyond abstract theory.
Deep Think was updated in close partnership with scientists and researchers. The goal is to tackle complex research problems, where data is often messy or incomplete and where no single solution exists. The model combines deep scientific knowledge and utility in engineering.
Record Benchmarks
| Benchmark | Score | Note |
|---|---|---|
| Humanity’s Last Exam | 48.4% (without tools) | New standard for frontier models |
| ARC-AGI-2 | 84.6% | Verified by ARC Prize Foundation |
| Codeforces | Elo 3455 | Competitive programming |
| IMO 2025 | Gold Medal | International Mathematical Olympiad |
| IPhO 2025 (written) | Gold Medal | International Physics Olympiad |
| IChO 2025 (written) | Gold Medal | International Chemistry Olympiad |
| CMT-Benchmark | 50.5% | Advanced Theoretical Physics |
Real-world Use Cases
- Mathematics (Lisa Carbone, Rutgers University): Deep Think identified a subtle logical flaw in a highly technical mathematics paper, which had escaped human peer review. She works on mathematical structures necessary for high-energy physics.
- Materials Science (Wang Lab, Duke University): Optimization of manufacturing methods for the growth of complex crystals, candidates for semiconductor materials. Deep Think designed a recipe to grow thin films of over 100 micrometers.
- Mechanical Engineering (Anupam Pathak, Google Platforms & Devices): Iteration on physical prototypes at the speed of software. Ability to transform a sketch into a 3D printable object (STL file generation).
Gemini 3 Deep Think is available starting today for Google AI Ultra subscribers in the Gemini app, and for the first time via the API in early access for researchers, engineers, and enterprises.
xAI restructuring: 4 teams and Macrohard
February 11 — xAI publishes a 45-minute all-hands on X, revealing a major restructuring of the company following the departure of half of its co-founders (6 out of 12). The latest departures are those of Tony Wu and Jimmy Ba.
Elon Musk announces the reorganization of xAI into 4 main teams:
| Team | Focus | Leader |
|---|---|---|
| Grok | Chatbot + voice | - |
| Coding | Application code system | - |
| Imagine | Video generation | - |
| Macrohard | Computer simulation → enterprises | Toby Pohlen |
Macrohard is the most ambitious project: according to Toby Pohlen, the system “is capable of doing everything a computer can do on a computer”. The ambition extends to the design of rocket engines by AI.
Key Metrics Announced
- X has exceeded $1 billion USD in annual recurring revenue (subscriptions)
- Imagine generates 50 million videos per day
- Over 6 billion images in the last 30 days
Musk’s interplanetary vision extends from orbital data centers to lunar bases. These announcements come after the acquisition of xAI by SpaceX (February 2) and the departure of 6 co-founders out of 12 over the last year (Kyle Kosic to OpenAI, Christian Szegedy, Igor Babuschkin, Greg Yang, Tony Wu, Jimmy Ba).
Mistral invests €1.2 billion in Sweden
February 11 — Mistral AI announces an investment of €1.2 billion ($1.4 billion USD) to build AI infrastructure in Sweden, in partnership with EcoDataCenter. This is Mistral’s first AI infrastructure investment outside of France.
The data center will be located in Borlänge (Sweden) on the EcoDataCenter site, with an opening planned for 2027. The facility will provide 23 megawatts of computing power and will host the latest generation NVIDIA Vera Rubin GPUs.
| Detail | Value |
|---|---|
| Amount | €1.2bn / $1.4bn |
| Partner | EcoDataCenter |
| Location | Borlänge, Sweden |
| Opening | 2027 |
| Capacity | 23 MW |
| GPU | NVIDIA Vera Rubin |
| Projected Revenue | €2bn over 5 years |
Arthur Mensch, CEO of Mistral AI, stated that this investment is “a concrete step towards building independent capabilities in Europe, dedicated to AI”. The choice of Sweden is explained by its access to clean and relatively cheap energy. The goal is to deliver a fully European AI stack — designed, built, and operated across the entire AI value chain, with data processed and stored locally in Europe.
Qwen-Image-2.0: unified image generation and editing
February 10 — Alibaba launches Qwen-Image-2.0, a next-generation foundation model for image generation that unifies text-to-image generation and image editing in a single architecture.
Major changes compared to v1:
- Unified generation + editing architecture (previously separate models)
- More compact model: 7B parameters (vs 20B for v1)
- Improved text rendering with prompt support up to 1K tokens
- Native 2K resolution (2048x2048)
| Benchmark | Qwen-Image-2.0 | FLUX.1 |
|---|---|---|
| DPG-Bench | 88.32 | 83.84 |
| AI Arena (blind eval) | #1 | - |
The model excels in rendering text in Chinese and English across different formats: infographics, posters, calligraphy, signage. Qwen-Image-2.0 is available for API testing on the Alibaba Cloud BaiLian platform and for free via Qwen Chat (chat.qwen.ai).
Deep Research upgrades to GPT-5.2
February 10 — OpenAI updates ChatGPT Deep Research with the GPT-5.2 model, replacing the previous o3 and o4-mini versions. The update brings app connections, search on specific sites, and real-time controls.
| Feature | Detail |
|---|---|
| Model | GPT-5.2 (replaces o3 / o4-mini) |
| App connections | Gmail, Google Calendar, Google Contacts + any MCP |
| Targeted search | Restriction to specific sites and trusted sources |
| Real-time controls | Progress tracking, interruption, and mid-run redirection |
| Research plan | Createable and editable before launch |
| Interface | Redesigned sidebar + full-screen report view |
| Export | PDF and DOCX |
Users can now connect Deep Research to any MCP server or application, and restrict web searches to trusted sites for more accurate and credible reports. Apps work in read-only mode as reliable sources.
Deployment began on February 10 for Plus and Pro users, with Free and Go users to follow in the coming days.
ChatGPT joins Pentagon’s GenAI.mil
February 10 — OpenAI announces the deployment of a custom version of ChatGPT on GenAI.mil, the US Department of Defense’s enterprise AI platform. ChatGPT joins Google Gemini (launched in December 2025) and xAI Grok (added in late December 2025) on the platform.
| Detail | Value |
|---|---|
| Platform | GenAI.mil (DoD) |
| Potential users | ~3 million (military, civilians, contractors) |
| Active platform users | 1.1 million unique since launch |
| Classification | Unclassified data only |
| Infrastructure | Authorized government cloud |
| Data isolation | Data isolated, not used to train OpenAI models |
Use cases cover policy document synthesis and analysis, procurement material drafting, internal report and compliance checklist generation, and research and planning assistance.
ElevenLabs for Government
February 11 — ElevenLabs launches ElevenLabs for Government, an initiative dedicated to the public sector announced at the ElevenLabs Summit in London. This platform offers voice and chat AI agents designed for government organizations, available 24/7, in multiple languages, via phone, chat, email, and WhatsApp.
The platform aims for three objectives:
- Accessibility and inclusion: multilingual omnichannel agents capable of resolving citizen requests instantly
- Efficiency and productivity: automation of high-volume contacts and reduction of transfers between services
- Trust and engagement: improving access to communications aligned with public policies
| Deployment | Details |
|---|---|
| Czech Republic | National employment and benefits hotlines, ~5,000 calls/day, 85% autonomous resolution |
| Ukraine | Digital twin of a senior official + public services, in partnership with the Ministry of Digital Transformation |
| Midland, Texas (USA) | Civic voice concierge “Jacky”, estimated reduction of 7,000 missed calls/month |
The platform is SOC 2 Type 2, GDPR, CPRA, and HIPAA certified to meet public sector security requirements.
Anthropic covers electricity price hikes
February 11 — Anthropic announces that it will cover electricity price increases for consumers caused by its data centers. The company makes four concrete commitments:
| Commitment | Detail |
|---|---|
| Grid infrastructure costs | Covering 100% of grid connection costs |
| New energy production | Commissioning new generation capacities adapted to needs |
| Grid pressure reduction | Investment in peak consumption reduction systems |
| Community investment | Creation of hundreds of permanent jobs and thousands of construction jobs |
Anthropic donates $20 million for AI regulation
February 12 — Anthropic announces a $20 million donation to Public First Action, a new bipartisan 501(c)(4) organization in the US. The goal is to mobilize citizens and policymakers on AI issues in the public interest.
Anthropic’s tweet highlights that AI is being adopted faster than any technology in history, and that the window to get public policies right is closing.
Claude Code Desktop: local plugins and marketplace
February 11 — Boris Cherny (Claude Code lead at Anthropic) publishes a thread detailing what engineers love about Claude Code: its customizability. He lists hooks, plugins, LSPs, MCPs, skills, effort, custom agents, status lines, output styles as customization mechanisms.
In parallel, Lydia Hallie announces that Claude Code on desktop now supports local plugins. Users can install custom slash commands, skills, and MCP servers from the marketplace, with automatic synchronization between desktop and CLI.
GitHub Mobile: Model Picker for Copilot
February 11 — GitHub adds a model selector to the Copilot Coding Agent in the GitHub Mobile app, on iOS and Android.
Copilot Pro and Pro+ users can now choose the AI model powering their coding agent sessions directly from their phone, without having to switch to desktop.
| Available Model | Publisher |
|---|---|
| Auto (default — optimizes speed and performance) | GitHub |
| Claude Opus 4.5 | Anthropic |
| Claude Opus 4.6 | Anthropic |
| Claude Sonnet 4.5 | Anthropic |
| GPT-5.1-Codex-Max | OpenAI |
| GPT-5.2-Codex | OpenAI |
Auto mode automatically selects a model to optimize speed and performance based on availability. The feature is available now for Copilot Pro and Pro+ subscribers. Business and Enterprise support is coming soon. Note: GPT-5.3-Codex, made GA on February 9, is not yet in the mobile list.
GitHub suspends GPT-5.3-Codex (reliability)
February 10 — GitHub announces the temporary suspension of the GPT-5.3-Codex deployment in GitHub Copilot for platform reliability reasons, just one day after announcing its general availability.
The tweet, which quotes the February 9 announcement on the GPT-5.3-Codex GA, generated 357,500 views and 1,000 likes, witnessing the impact on the developer community. Several users reported service availability issues in the replies.
The fact that the model is not yet in the Mobile Model Picker list (announced February 11) could be related to this suspension.
Manus launches Project Skills
February 12 — Manus (now under Meta) launches Project Skills, a feature that allows creating dedicated skill libraries per project.
Each project can assemble a set of skills selected from “Team Skills” pools or personal collections. The idea: transform a standard project folder into an intelligent and autonomous workspace.
| Feature | Description |
|---|---|
| Libraries per project | Assembly of custom toolkits from team or personal pools |
| Contained workflows | Only skills explicitly added to the project can be triggered |
| Locked workflows | Admins can lock the skill set to standardize processes |
Benefits highlighted: faster onboarding, scaling of individual expertise to the organization, and building “institutional knowledge” that improves over time. Available immediately for all Manus users.
Gemini CLI v0.28.0 + Extension Settings
February 10-11 — Google releases version 0.28.0 of Gemini CLI and announces Extension Settings, a new feature to simplify extension configuration.
Gemini CLI v0.28.0 (February 10)
| New Feature | Detail |
|---|---|
/prompt-suggest command | New slash command for prompt suggestions |
| Automatic theme | Adaptation based on terminal background |
| IDE compatibility | Support added for Positron IDE |
| Improved OAuth | Interactive and non-interactive OAuth consent |
| Checklist component | Component for structured task management in plan mode |
| Sub-agents | Dynamic policy registration for sub-agents |
| Background execution | Background shell commands |
Extension Settings (February 11)
Gemini CLI extensions can now define configuration parameters that users provide upon installation. Automatic setup with interactive prompts, built-in security (sensitive data like API keys are stored in the system keychain), and centralized management via gemini extensions config. Data Cloud extensions (BigQuery, Cloud SQL, Firestore, Looker, Spanner) already implement this feature.
AI Shopping in Google Search and Gemini
February 11 — Google announces the integration of shopping features directly into its AI products — Search AI Mode and the Gemini chatbot.
Users can now buy Etsy and Wayfair items directly in the Gemini chatbot. Direct Offers, a new feature in AI Mode, allows brands to offer discounts to potential buyers. New ad formats are being tested in Search AI Mode for retailers.
Google is working with Walmart, Target, and Shopify on a checkout protocol allowing consumers to finalize their purchases directly within Google’s AI products.
GTIG Report: Distillation attacks on Gemini
February 12 — The Google Threat Intelligence Group (GTIG) publishes a report revealing that Gemini has been targeted by distillation attacks — massive cloning attempts via repeated prompts.
One campaign submitted over 100,000 prompts to Gemini before being detected. The attacks specifically targeted Gemini’s reasoning algorithms. Attackers are primarily private companies and researchers seeking a competitive advantage.
Google’s systems detected the activity in real-time and adjusted protections. Google considers distillation as intellectual property theft.
Codex CLI v0.99.0: concurrent shell execution
February 11 — OpenAI releases Codex CLI v0.99.0 with concurrent shell execution, a configurable statusline, and dedicated app-server APIs.
| Feature | Detail |
|---|---|
| Concurrent shell execution | Direct shell commands no longer block a current turn |
| /statusline | Interactive metadata configuration in the TUI footer |
| Adaptive resume picker | New selector for resuming sessions |
| App-server APIs | Dedicated APIs for application integration |
| Admin controls | Admin controls for web search and networking |
| Images | GIF/WebP support |
| Snapshotting | Environment snapshotting |
| Steer mode | Stable and active by default (Enter sends, Tab queues) |
The npm packaging has been reworked: platform-specific binaries are distributed via @openai/codex dist-tags, reducing package size.
Qwen Chat AI Slides
February 12 — Qwen announces the upcoming arrival of AI Slides in Qwen Chat. The feature allows converting different types of documents into presentations: project docs → meeting slides, paper → oral deck, product info → sales deck, topic → teaching deck.
The feature is built with Qwen Agent + Qwen-Image 2.0 (the image model launched a few days prior). Chen Cheng demonstrated the ability to transform academic papers into oral presentations in a few minutes, with an integrated search agent.
What this means
Anthropic’s 380 billion valuation confirms the dominance of frontier models in the AI economy. The 2.5 billion share generated by Claude Code show that AI-assisted coding has become critical infrastructure for developers and enterprises.
The arrival of GPT-5.3-Codex-Spark on Cerebras marks a turning point in AI hardware diversification beyond NVIDIA. Inference at 1000+ tokens/second changes the nature of interaction with code models, making fluid real-time collaboration possible.
Gemini 3 Deep Think with its gold medals at science olympiads and its 84.6% score on ARC-AGI-2 demonstrates that frontier models are beginning to reach a level of scientific reasoning comparable to the best humans in specialized fields.
xAI’s restructuring into 4 teams with Macrohard — aiming to do “everything a computer can do” — and Mistral’s massive investment in Europe show that the race for AI infrastructure is accelerating globally. GitHub’s suspension of GPT-5.3-Codex recalls, however, that reliability remains a major challenge for large-scale deployment.
Sources
- Anthropic — Series G $30B
- Anthropic — Electricity
- Anthropic — Public First Action
- Anthropic — Claude Code Plugins
- OpenAI — GPT-5.3-Codex-Spark
- OpenAI — Deep Research
- OpenAI — GenAI.mil
- OpenAI — Codex CLI v0.99.0
- Google — Gemini 3 Deep Think
- Google — AI Shopping
- Google — Gemini CLI Extensions
- Google — GTIG Distillation
- xAI — Restructuring
- Mistral — Sweden Investment
- Qwen — Image 2.0
- Qwen — AI Slides
- ElevenLabs — For Government
- GitHub — Mobile Model Picker
- GitHub — Pause GPT-5.3-Codex
- Manus — Project Skills