Callista AI Weekly (February 09 - February 15, 2026)
This week marked a decisive shift from AI experimentation to execution. From Google pushing agentic checkout straight into Search, to Chinese labs setting the global tempo on models and costs, to Switzerland hosting a landmark AI week at EPFL while charting its own governance path, these developments carry immediate implications for Swiss businesses in every sector. Below, we distill what mattered most and why it counts for your strategy now.
New AI Use Cases
Monetizing the chatbot era: OpenAI tests ads in ChatGPT
OpenAI began testing ads inside ChatGPT on February 9, a historic monetization turn for the world’s most-used AI chatbot. U.S. users on the Free and Go (8 USD per month) tiers now see contextual ads at the bottom of answers, clearly labeled and separated from responses. Pro, Business, and Enterprise subscribers remain ad-free, and users under 18 will not see ads. OpenAI said ads do not influence ChatGPT’s responses and will not appear near sensitive topics, including health, politics, or mental health. The timing, one day after Anthropic’s Super Bowl spot that appeared to mock ad-supported chatbots, underscored the mounting debate about how foundational AI interfaces will be funded.
AI in the warehouse: drones that “get curious”
Gather AI raised 40 million USD in Series B funding on February 9. Spun out of CMU, the startup blends classical Bayesian techniques with neural networks to make drones “curious,” autonomously discovering misplaced stock, low inventory, and safety hazards. By bringing AI into physical logistics environments, Gather AI exemplifies how autonomy and perception can compress audit cycles, elevate inventory accuracy, and harden safety compliance without adding headcount.
Healthcare at scale: UCSF rolls out ChatGPT Enterprise
UCSF announced a system-wide deployment of ChatGPT Enterprise on February 9, rolling out a HIPAA-compliant, data-protected version to its 9,000 plus existing AI users starting February 17, with full community access by February 23. It is among the largest academic medical center deployments to date, signaling a maturing path for secure, compliant AI in clinical and administrative workflows.
Agentic finance: deterministic outputs for regulators
Meridian AI raised 17 million USD in seed funding on February 11 at a 100 million USD post-money valuation. The startup builds a standalone IDE for agentic financial modeling, designed to produce deterministic, auditable outputs tailored for regulated institutions. By December 2025, Meridian reported 5 million USD in signed contracts. This moves beyond “Copilot-in-Excel” toward a purpose-built environment for compliant automation.
Healthcare devices and imaging: operational gains and new modalities
GE HealthCare launched ReadyFix remote fleet management for ECG devices and secured FDA clearance plus CE Marking for Allia Moveo, a compact AI-powered interventional imaging system, on February 14. Meanwhile, Samsung confirmed a February 25 launch event for its Galaxy S26 with AI features front and center, and a goal to double Gemini AI-equipped mobile devices to 800 million units by the end of 2026.
Commerce and customer experience: AI meets “where’s my order” and selfies
FedEx rolled out AI tools, Tracking plus and Returns plus, that plug into merchant sites to handle delivery and return queries, reportedly cutting order-status inquiries by 42 percent. Gucci became the first luxury brand to launch a Sponsored AI Lens on Snapchat, using generative AI to transform selfies into styled personas. And inference providers including Baseten, DeepInfra, and Together AI reported cutting AI costs by up to 10 times using open-source models on NVIDIA Blackwell GPUs.
Major Vendor Updates
China’s Lunar New Year AI war: GLM-5, ByteDance’s triple-shot, DeepSeek’s 1 million-token context
Timed to the Lunar New Year, Chinese labs unleashed a flurry of releases:
Zhipu AI’s GLM-5 on February 11 is an open-source Mixture-of-Experts model with approximately 745 billion parameters, 44 billion active per inference across 256 experts, trained entirely on domestic Chinese chips including Huawei Ascend, Moore Threads, Cambricon, and Kunlunxin. Zhipu claims GLM-5 approaches Claude Opus 4.5 on coding and surpasses Gemini 3 Pro on some tests. The company’s Hong Kong shares surged 70 percent that week. On February 12 Zhipu raised its GLM Coding Plan prices by 30 to 60 percent, the first significant LLM price hike in China in 2026.
ByteDance released Seedance 2.0 for video generation, producing 4 to 15 second clips at 2K resolution, followed by Seedream 5.0 Lite for images and Doubao 2.0 on February 14 to 15, upgrading China’s most popular AI chatbot with roughly 155 million weekly active users. ByteDance claimed Doubao 2.0 matches GPT-5.2 and Gemini 3 Pro on multi-step task execution and promoted Doubao during the state broadcaster’s New Year gala, including luxury car giveaways.
DeepSeek pushed V4 preview upgrades on February 11, expanding context from 128K to 1 million tokens and updating its knowledge cutoff, with full V4 expected around February 17.
Kuaishou released Kling 3.0, a video generation model with native multi-language audio generation.
MiniMax launched M2.5 and M2.5 Lightning, claiming near state-of-the-art performance at roughly one twentieth the cost of Claude Opus 4.6.
The signal is unmistakable. Chinese vendors are no longer simply catching up. On cost efficiency, model variety, and user acquisition tactics, they are setting the tempo.
OpenAI: faster coding and contentious retirements
OpenAI released GPT-5.3-Codex-Spark on February 12, a smaller model optimized for real-time coding on Cerebras’ Wafer Scale Engine 3, delivering 1,000 plus tokens per second. It is the first shipped milestone in OpenAI’s January 2026 Cerebras partnership, available to ChatGPT Pro users via the Codex app and VS Code extension.
In a more controversial move, OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on February 13, following earlier retirements of GPT-5 Instant and Thinking. The GPT-4o move, in particular, triggered backlash, prompting an invite-only subreddit created by thousands of users.
Amazon, Anthropic, xAI, and fundraising signals
Amazon Bedrock added six open-weights models on February 10 via Project Mantle, a new distributed inference engine, broadening AWS’s model marketplace with OpenAI API-compatible endpoints. Anthropic focused on Claude Code stability and performance updates during the week. xAI announced Grok 4.20 on February 15 as a significant improvement over Grok 4.1, with early checkpoints ranking near the top on ForecastBench, and the launch landed on February 17. Anthropic was reportedly closing in on a 20 billion USD round at a 350 billion USD valuation on February 9, with investor demand doubling the initial target.
AI Governance Developments
EU targets Meta over WhatsApp AI access
The European Commission issued a Statement of Objections to Meta on February 9 over excluding third-party AI assistants from WhatsApp after Meta’s October 2025 policy change. From January 15, 2026, only Meta AI remained on the platform. Interim measures are being considered quickly. Meta called the findings baseless. This marks messaging platforms as critical AI distribution channels in the eyes of regulators.
India enacts what may be the world’s strictest AI content rules
India’s Ministry of Electronics and IT notified the Information Technology Amendment Rules 2026 on February 10, effective February 20. Platforms must comply with government takedown orders within three hours, all synthetically generated content must be prominently labeled, and platforms must deploy automated tools to prevent illegal AI content including deepfakes and AI-generated CSAM. The timing dovetailed with India’s AI Impact Summit starting February 16, signaling a coordinated push to shape AI conduct at scale.
Munich puts cyber and AI at the core
At the 62nd Munich Security Conference, February 13 to 15, European leaders elevated cyber and AI to core national security concerns. Google’s Kent Walker released a whitepaper, “Staying Ahead of the Shadows: Digital Resilience in the Era of AI,” and Google Threat Intelligence reported adversaries using AI to automate reconnaissance and hyper-realistic phishing. The event reflected Europe’s broader pursuit of strategic autonomy in critical technologies.
AI safety exodus alarms policymakers
A wave of high-profile AI safety researcher resignations drew attention on February 15. Mrinank Sharma left Anthropic on February 9, posting “the world is in peril.” Zoe Hitzig departed OpenAI over ChatGPT ad testing, warning of manipulation risks. Two co-founders and five staff left xAI. Yoshua Bengio, chair of the 2026 International AI Safety Report, warned of completely unexpected problems, including psychological harm from emotional attachment to AI chatbots. Additional governance analyses noted that safety testing is getting harder as models distinguish test settings from real deployments, a RAND report cautioned that U.S. dominance in LLMs should not be taken for granted, and former FCC Chairman Tom Wheeler urged U.S. policymakers to watch the owners of AI technology rather than the technology itself.
Switzerland-specific governance and national capacity
EPFL hosts a landmark AI week
Switzerland’s Applied Machine Learning Days, AMLD, Intelligence Summit celebrated its 10th anniversary at EPFL from February 10 to 12. Highlights included keynotes on brain–spine interfaces enabling thought-controlled walking, the Swiss AI Initiative showcasing Apertus, Switzerland’s sovereign foundation model with over one million downloads since its September 2025 launch and backed by 20 million CHF through 2028, and sessions from Isomorphic Labs on AlphaFold and drug discovery and Roche on AI for de-risking clinical trials. The Open-Source LLM Builders Summit on February 9, organized by the Swiss AI Initiative, underscored Switzerland’s growing open-source leadership.
Switzerland charts its own AI governance path
On February 10, the Swiss Digital Advisory Board, Beirat Digitale Schweiz, convened in Bern under Federal Councillor Albert Rösti to discuss implementing the Council of Europe’s AI Convention. Switzerland is pursuing an alternative to the EU AI Act via a dual strategy, a formal legal framework with a consultation draft due by end of 2026 paired with industry self-regulation. The media and energy sectors presented self-regulatory proposals, and Federal Chancellor Viktor Rossi also attended. This approach aims to blend innovation flexibility with sector-led safeguards.
ETH Zurich gains UN AI representation
On February 12, two ETH Zurich computer scientists, Professors Menna El-Assady and Bernhard Schölkopf, were elected to the UN’s newly established Independent International Scientific Panel on Artificial Intelligence. Among 40 global experts, they will assess AI’s opportunities, risks, and societal impacts for international policymakers.
Auterion’s defense JV and Switzerland’s “drone dilemma”
Auterion, an ETH Zurich spin-off, announced a joint venture with Airlogix at the Munich Security Conference on February 13 to mass-produce AI-guided autonomous strike drones for Ukraine and NATO allies, backed by 40 million euros with production in Germany. Swiss arms export laws pushed the company’s HQ relocation to the U.S., though Zurich offices remain. Swiss media highlighted a “drone dilemma,” with Armasuisse testing Auterion systems for the Swiss Army even as defense work must occur abroad.
Breakthrough Research
A foundation model for neuroimaging: reading brain MRIs in seconds
The Prima model from the University of Michigan, published in Nature Biomedical Engineering and reported on February 10, is a foundation model trained on over 200,000 MRI studies and 5.6 million imaging sequences. It achieved diagnostic accuracy up to 97.5 percent across 52 neurological conditions and can assess case urgency in seconds. A one-year health system evaluation of around 30,000 MRI studies showed Prima outperformed other state-of-the-art AI systems. For hospitals and insurers, this suggests the potential for faster triage, earlier interventions, and more consistent reads, subject to clinical validation and robust governance.
Frontier AI still stumbles on research-grade math
The First Proof challenge results on February 13 showed top models struggling with genuine research-level mathematics. Eleven eminent mathematicians, including Fields Medalist Martin Hairer, posed 10 unpublished problems. GPT-5.2 Pro and Gemini 3.0 Deepthink solved only 2 of 10 in single-attempt mode, revealing a performance gap between olympiad-style problem solving and original research capability.
Neuromorphic computing tackles PDEs with near-ideal scaling
Sandia National Laboratories introduced NeuroFEM in Nature Machine Intelligence on February 14, an algorithm enabling neuromorphic hardware to efficiently solve partial differential equations, core to weather, fluid dynamics, and nuclear simulations. Tested on Intel’s Loihi 2, it showed close to ideal scaling, nearly halving compute time when doubling cores, at significantly lower energy costs than standard CPUs. For energy-intensive simulation workloads, this signals a future of greener, faster physics at the edge of new hardware paradigms.
Other notable findings: safety vulnerabilities and clinical disparities
Microsoft researchers described “GRP-Obliteration,” an attack that can break safety mechanisms across 15 major language models using a single malicious training prompt, raising GPT-OSS-20B’s attack success rate from 13 percent to 93 percent across 44 harmful categories. This underscores urgent needs for red-teaming and robust defense-in-depth.
A Nature study on AI in emergency medicine reported concerning performance disparities, with mid-tier models, for example Mistral Medium at 61.2 percent and DeepSeek R1 at 66.3 percent, lagging top-tier systems in critical tasks, evidence that just good enough models may pose operational risks in high-stakes care.
Axios reported on February 15 that AI advances are outpacing academic publishing cycles, leaving much research outdated by publication, complicating evidence-based regulation and procurement cycles.
Conclusion
Three patterns define this extraordinary week. First, the center of gravity in AI model development shifted visibly eastward. Zhipu’s GLM-5 trained on domestic Chinese chips, ByteDance’s triple model release, DeepSeek’s silent V4 preview, and MiniMax’s cost-competitive M2.5 collectively demonstrate that Chinese labs are not simply catching up, they are setting the tempo, particularly on cost efficiency and open-source availability. Second, agentic AI crossed the threshold from prototype to production. Google’s UCP-powered checkout on Etsy and Wayfair, GitHub’s Agentic Workflows in CI and CD, and the OpenClaw security crisis all show that autonomous AI agents are now operating at scale, with real commercial impact and real security consequences. Third, governance is fracturing along geographic lines. The EU is using antitrust to force AI platform openness, India is imposing the world’s strictest content rules, Switzerland is betting on self-regulation, and the Munich Security Conference elevated AI from a tech policy footnote to a core defense concern.
Ready to explore how Agentic AI can transform your organization? Visit us at https://www.callista.ch/agentic-ai to discover how we can guide your journey into this exciting new era of AI-powered productivity.
Sources
Google Blog — "New tech and tools for retailers to succeed in an agentic shopping era" (UCP/agentic commerce launch) — February 11, 2026
CNBC — "China's tech titans are giving away money and cars in 'The Lunar New Year AI War'" — February 13, 2026
The Register — "Brussels eyes crowbar for Meta's WhatsApp AI lockout" — February 9, 2026
WKZO/Reuters — "Chinese AI startup Zhipu releases new flagship model GLM-5" — February 11, 2026
The Register — "OpenAI introduces ads…for the people!" — February 10, 2026
TechCrunch — "Anthropic closes in on $20B round" — February 9, 2026
The New Stack — "GitHub's Agentic Workflows bring 'continuous AI' into the CI/CD loop" — February 14, 2026
ScienceDaily — "AI reads brain MRIs in seconds and flags emergencies" — February 10, 2026
Al Jazeera — "Why are experts sounding the alarm on AI risks?" — February 15, 2026
CNBC — "Alibaba's RynnBrain, ByteDance's Seedance 2.0 are among new China AIs" — February 14, 2026
AWS — "Amazon Bedrock adds support for six fully-managed open weights models" — February 10, 2026
ETH Zurich — "Professors Menna El-Assady and Bernhard Schölkopf elected for UN AI Scientific Panel" — February 12, 2026
TechCrunch — "Gather AI, maker of 'curious' warehouse drones, lands $40M" — February 9, 2026
SiliconANGLE — "Agentic financial modeling startup Meridian gets $17M in funding" — February 11, 2026
