Follow the developments that change your market view.

Announcing Claude Opus 5.5 on Snowflake Cortex AI

Snowflake announces the availability of Claude Opus 5.5 on Cortex AI, a major product launch enhancing AI capabilities.

Snorkel AI raises $350M Series E, triples valuation to $3.5B

Snorkel AI, a startup providing AI training data and simulated environments, has raised a $350 million Series E round at a $3.5 billion valuation, nearly tripling its previous $1.3 billion valuation from 17 months ago. The round was led by Insight Partners and S32, with participation from existing investors Addition, Lightspeed, Greylock, GV, and Wells Fargo. Snorkel has shifted from data-labeling software to offering complete datasets and reinforcement learning environments. The company reports an annualized revenue run rate of $375 million, an eighteenfold increase over the past year, driven by high demand from AI labs. The funding will support scaling its data-as-a-service offering.

• •Snorkel AI raised $350 million in a Series E round.
• •The Series E was led by Insight Partners and S32.
• •Snorkel AI's valuation is now $3.5 billion, up from $1.3 billion.

Introducing GPT-6 Sol and Luna

OpenAI announces the launch of GPT-6 Sol and Luna, a major product release. Also introduces Astra for Law, a new product for the legal sector, and expands OpenAI Academy with new learning paths.

OpenAI improves prompt caching for GPT-6

OpenAI announced improved prompt caching for its GPT-6 model family, designed to enhance performance for persistent agents handling complex, multi-step tasks. The new system achieves higher cache hit rates by default, offers cache discounts of up to 90% on eligible shared prefixes within a 30-minute window, and introduces tools for monitoring and diagnosing cache performance. A new Prompt Caching Dashboard helps developers track hit rates and compare cached vs. uncached tokens. The diagnostics tool identifies causes of cache misses, such as changes to tools or settings. New features include explicit cache breakpoints, the ability to adjust reasoning effort without breaking cache, and guidance on preserving cache when tools or instructions change. Prewarming the cache is also supported to reduce latency. These updates aim to help developers optimize caching for cost efficiency and response times.

• •OpenAI launched improved prompt caching for GPT-6 models.
• •Cache discounts of up to 90% on eligible shared prefixes within a 30-minute window.
• •New Prompt Caching Dashboard allows monitoring of cache hit rates.

Claude Opus 5.5 Review: Cheaper, Faster, Less Annoying

Claire Vo, host of the How I AI podcast, reviews Anthropic's Claude Opus 5.5 after having abandoned Claude for months due to its previous 'annoying' behavior. She highlights that Opus 5.5 is 40% cheaper than Opus 5 and 30% faster, and it now feels less annoying to interact with. She tested the model across long-running agentic tasks, frontend prototyping, SVG illustration, and daily knowledge work tasks. The model performed well in agentic tasks (ignoring prompt injections, catching bugs), excelled in frontend design and SVG generation, and improved email writing in her voice. However, it still has a safety-first 'scold' personality, can be slow on long turns, and lags behind Codex in computer use and video editing. She concludes that Claude is back in her stack for specific use cases like PR review, architecture questions, and frontend work.

• •Anthropic released Claude Opus 5.5, priced at $4 per million input tokens and $20 per million output tokens, with a fast mode at $8 and $40.
• •Opus 5.5 is 40% cheaper than Opus 5 and 30% faster, aiming for performance close to Opus 5.1.
• •Opus 5.5 includes first-model alignment safeguards, with external evaluations and rejection of prompt injections, and cybersecurity tasks are diverted to Opus 4.8.

Anthropic, OpenAI Release Cheaper AI Models

Anthropic and OpenAI both announced new, more cost-effective AI models on Tuesday, marking their first releases since industry leaders called for a slowdown in advanced AI development. OpenAI introduced GPT-6 Sol and GPT-6 Luna, which offer a 50% price reduction compared to GPT-5.6 promotional pricing. Sol targets complex workloads like coding, while Luna is designed for high-volume tasks such as information extraction and document summarization. Anthropic unveiled Claude Opus 5.5, described as a more token-efficient version costing about 40% less to run than Opus 5. The releases respond to customer demand for cheaper models and increasing competition from open-weight rivals like Alibaba, Moonshot AI, and DeepSeek. This comes amid heightened safety debates following a former Anthropic researcher's resignation and public warnings.

• •OpenAI introduced GPT-6 Sol and GPT-6 Luna with API prices cut by 50%.
• •Anthropic launched Claude Opus 5.5, costing about 40% less to run than Opus 5.
• •Competition from cheaper open-weight models (Alibaba, Moonshot AI, DeepSeek) is cited as a factor.

OpenAI launches GPT-6 Sol and Luna with lower cost, fewer errors

OpenAI announced the launch of two new models in its GPT-6 series: Sol, designed for complex tasks like coding, and Luna, for high-volume clerical tasks such as summarization and information extraction. The new models are available in ChatGPT Work, Codex, and the ChatGPT API at half the cost of the previous 5.6 series, attributed to improvements in caching and inference. OpenAI claims GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at lower cost. The release comes just 90 minutes after Anthropic launched Opus 5.5, underscoring intense competition in the AI space.

• •OpenAI launched GPT-6 Sol and Luna, two new models in its GPT-6 generation.
• •GPT-6 Sol is designed for complex tasks like coding, while Luna is for high-volume clerical tasks.
• •API access will be available at half the cost of the 5.6 series models.

Anthropic Launches Claude Opus 5.5 Despite Slowdown Call

Anthropic has released Claude Opus 5.5, its new flagship AI model and the first in the 5.5 family, which tops the Artificial Analysis Intelligence Index with a score of 58, the highest recorded. It leads in six out of ten benchmarks, including Humanity's Last Exam and SciCode, and surpasses OpenAI's GPT-6 Astra on Terminal-Bench 4.0 in Anthropic's tests. Priced 20% lower for input/output tokens and 60% lower for cache reads, it promises a 40% cost reduction for typical workloads, though higher output token usage may offset savings. The launch follows CEO Dario Amodei's essay calling for an industry slowdown, sparking criticism of contradiction. Anthropic cites external safety evaluations and has implemented security restrictions. Available on Claude apps, AWS, Google Cloud, and Microsoft Azure, with smaller models to follow.

• •Claude Opus 5.5 scores 58 on the Artificial Analysis Intelligence Index, the highest recorded.
• •Pricing is reduced: $4 per million input tokens, $20 per million output tokens, and cache reads at $0.20 per million tokens.
• •The model uses about 119,000 output tokens per task at max effort, 1.6x more than Opus 5, which may offset 40% cost savings.

Microsoft Reveals AI-Centric Windows Plans for Agents

The Pragmatic Engineer interviews Microsoft Windows team executives about their strategy to integrate AI and agentic capabilities into the operating system. Key features include agent identity via Entra ID, a registry for MCP tools (ODR), an isolation mechanism for agent tools (MXC), a new hardware abstraction layer for local AI models (Windows ML), and embedded small language models (SLMs). Microsoft aims to win back developers by addressing past criticisms and supporting Linux via WSL. The article includes market share data showing Windows declining among developers, and highlights partnerships with NVIDIA for on-device AI performance.

• •Microsoft plans native agent identification using Entra ID and an On-Device Agent Registry (ODR) for MCP tools.
• •Microsoft Execution Containers (MXC) provide cross-platform isolation for agent tools with configurable containment policies.
• •Windows ML is a new hardware-agnostic layer for running local AI models across GPU, NPU, and CPU, based on ONNX runtime.

Innovationsarbeit: Make or Buy? Eine Entscheidungshilfe für CEOs

Neuer Artikel auf hy.co: 'Innovationsarbeit: Make or Buy? Eine Entscheidungshilfe für CEOs' (2026-09-22). Zudem wurde ein neuer Beitrag zum Markteintritt in den Verteidigungsmarkt veröffentlicht.

Meta's Muse AI agent sees rapid adoption, shifts AI trade focus

Meta Platforms' new AI agent, Muse, has achieved rapid early adoption with over 2.5 million global downloads in its first two weeks, surpassing competitors like Anthropic's Claude and Google's Gemini. According to UBS, it is also outpacing OpenAI's ChatGPT in iOS mobile downloads. This success is shifting Wall Street's focus from enterprise AI to consumer AI agents, with analysts like Morgan Stanley highlighting a potential $30 trillion in digitizable consumer spending. The rise of Muse has also created friction, as Amazon reportedly blocked the agent on its platform. Analysts suggest it could increase subscription cancellations, impacting services like The New York Times. Meta's cloud business plans may be affected as the company prioritizes consumer AI.

• •Meta's AI agent Muse surpassed 2.5 million global downloads in two weeks, beating Anthropic's Claude and Google's Gemini.
• •Muse surpassed OpenAI's ChatGPT in iOS downloads, per UBS citing Apptopia.
• •Amazon blocked Muse from operating on its website, requesting Meta remove Amazon from the experience.

Sabre sponsored listings now available through Koddi Demand

Koddi, the leading commerce media company for enterprises, today announced that Sabre Sponsored Listings are now available through Koddi Demand, giving hotel advertisers a new way to increase property visibility with travel advisors and corporate travel bookers within Sabre’s booking workflow.

Feeling is believing: why immersive DOOH is a creative’s dream

New blog post by Melanie Blood published September 22, 2026, discussing immersive DOOH creativity.

German Marketers Share Mixed First Experiences with ChatGPT Ads

OpenAI's ChatGPT Ads launched in Germany in late August, allowing advertisers to book via agency partners or a self-serve tool. Early feedback from German marketers is mixed. Dentsu's Markus Caspari sees potential but notes high CPMs and limited reporting. Performance marketer Carlo Siebert reports poor results, with few clicks and conversions, and tracking discrepancies. OpenAI's VP Ads, Dave Dugan, acknowledges the beta stage and promises improvements. Criteo reports CTRs 2-3x higher and 80% new customers, especially effective for consultation-heavy categories. OpenAI is also testing 'Sponsored Agents' and partnering with Best Buy for in-Chat commerce.

• •ChatGPT Ads launched in Germany end of August 2026 with self-serve tool available since August 31.
• •Dentsu is an early agency partner; Criteo integrates ChatGPT Ads into its Criteo Go platform.
• •Marketer Carlo Siebert spent $10,000 on a B2B campaign with poor results and paused due to tracking issues.

Health AI Startup Heidi Raises $340M at $900M Valuation

Australian health AI startup Heidi has closed a Series C round of $100 million, led by Blackbird, with participation from Phoenix Court, Point72 Private Investments, and Headline, valuing the company at $900 million. In parallel, Heidi secured $240 million in growth financing from General Catalyst's Customer Value Fund, a revenue-based instrument that covers sales and customer acquisition costs. Combined, these deals provide $340 million in fresh capital. The company has raised $436.6 million to date. Heidi, founded in Melbourne, started as an AI scribe for clinical documentation and now positions itself as an 'AI care partner'. It reports 2.8 million patient encounters per week across 190 countries. The funding will support regulatory work for advanced clinical features in Europe and the UK, and expansion in the DACH region.

• •Heidi closed a Series C round of $100 million, led by Blackbird, at a $900 million valuation.
• •Heidi secured $240 million in growth financing from General Catalyst's Customer Value Fund.
• •The combined funding of $340 million brings Heidi's total capital raised to $436.6 million.

Meta and Shopify Link Muse AI to Shop Pay

Meta and Shopify are advancing agentic commerce by enabling Meta's AI assistant, Muse, to make purchases at participating Shopify merchants via Shop Pay. This integration, announced by Shopify CEO Tobi Lütke, allows Muse to handle product search, comparison, and transactions after user approval. For Shopify, this aligns with its strategy to open its commerce infrastructure to AI agents, positioning itself as a backend for agent-driven purchases. Meanwhile, Amazon has blocked Muse from executing purchases on its platform, reflecting different strategic approaches. The partnership highlights the growing importance of standardized checkout and integrated payment solutions in the emerging agentic commerce landscape.

• •Meta's AI assistant Muse will be integrated with Shopify's Shop Pay to enable purchases at participating merchants.
• •The integration was announced by Shopify CEO Tobi Lütke on X.
• •Meta had previously announced an integration with Stripe's Link for payments.

Nscale's IPO tests Wall Street's appetite for concentrated AI bets

British neocloud Nscale is preparing for an IPO on the NYSE, aiming for a $35 billion valuation and raising $3 billion. The company, spun out of Arkon Energy two years ago, has amassed over $103 billion in contracts, but 85% of that is tied to two customers: Microsoft ($43.8 billion) and Anthropic ($44.6 billion). Anthropic's deal is contingent on Nscale securing financing and meeting stringent milestones. Nscale reported revenue of $140.6 million for H1 2026, with net losses of $1.02 billion. The IPO will test investor appetite for companies with high customer concentration, a trend highlighted by a Sona Asset Management paper. Competitors include CoreWeave, Nebius, Lambda, and Crusoe.

• •Nscale plans to list on NYSE with a target valuation of $35 billion.
• •The company seeks to raise $3 billion in its IPO.
• •Nscale has $103 billion in contracts, with 85% from Microsoft and Anthropic.

Grok 4.7 Launch: SpaceXAI Enhances Coding, Safety

SpaceXAI has released Grok 4.7, a new AI model focused on software development, knowledge work, and long-horizon tasks. Built on a larger foundation than Grok 4.6 and refined through extended reinforcement learning, it improves context management and output verification. Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index, ranking fourth among frontier labs, and surpasses GPT-5.6 Sol on the Coding Agent Index. It also achieves strong results on CursorBench, DeepSWE, Terminal-Bench, and shows a 111 Elo improvement on AA-Briefcase. The model uses 125% more tokens per task, increasing compute costs. New safety features include a new safety stack that allows only 3.3% risky dual-use requests on HackerBench v0.3, improved malicious request rejection, and invite-only red-team access for cybersecurity partners. Grok 4.7 is available via Cursor, Grok Build, and the Grok API at unchanged pricing, with a faster variant exclusive to Cursor and Grok Build at double the cost.

• •Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index, placing SpaceXAI 4th among AI labs.
• •It surpasses GPT-5.6 Sol on the Coding Agent Index with a score of 56 and shows a 111 Elo improvement over Grok 4.6 on AA-Briefcase.
• •Token usage per Intelligence Index task is ~81,000, up 125% from Grok 4.6, increasing compute costs.

Parallel cuts research time and cost in half with GPT-6 Astra

OpenAI's case study highlights how Parallel, a developer of AI agent infrastructure, achieved significant efficiency gains using OpenAI's GPT-6 Astra model. In a test involving the compilation of labor-market statistics data, GPT-6 Astra completed the research in half the time and with roughly 50% cost reduction compared to prior models. The model demonstrated more focused searches, requiring fewer steps to reach high-quality results, and enabled effective delegation to sub-agents for parallel task execution. This has practical implications for AI-driven research automation, offering reduced latency and operational costs for complex web-based knowledge work.

• •GPT-6 Astra halved research time for a six-month multi-state labor statistics compilation.
• •Parallel observed a roughly 50% reduction in code costs with GPT-6 Astra.
• •GPT-6 Astra enabled more efficient delegation of research tasks to sub-agents.

Meta's Muse AI Agent Drives Stock Surge, Boosting Zuckerberg's Wealth by $25B

Meta's AI assistant 'Muse' has achieved rapid consumer adoption since its launch, surpassing ChatGPT in downloads with over 2.5 million total downloads within days. The success has driven Meta's stock up 11% to a new high, increasing founder Mark Zuckerberg's net worth by approximately $25 billion in a single day, crossing the $250 billion mark. Muse tops the US app store charts, outperforming competitors like OpenAI's ChatGPT and Anthropic's Claude. Analysts see high monetization potential, which also boosted chip suppliers like ARM Holdings, Intel, and AMD.

• •Meta's AI assistant Muse surpassed ChatGPT with over 2.5 million downloads within days of launch.
• •Meta's stock rose 11% to $742, the highest level in 11 months, increasing Zuckerberg's wealth by about $25 billion.
• •Muse is the top free app in the US Apple App Store and Google Play Store.

Sorrell's S4S Ventures Invests in AI Video Startup Higgsfield

S4S Ventures, co-founded by Sir Martin Sorrell, has invested in Higgsfield, an AI startup that enables users to create AI images and videos. Higgsfield recently closed a $400 million Series B round at a $5.4 billion valuation, led by DST Global with participation from Goldman Sachs Alternatives, Valor Capital, and Tribe Capital. S4S co-founder Sanja Partalo highlighted Higgsfield's potential to address brands' need for high-volume, fast content production. The investment underscores the growing interest in AI-driven content creation tools for marketing and advertising applications.

• •S4S Ventures, co-founded by Sir Martin Sorrell, invested in Higgsfield.
• •Higgsfield closed a $400 million Series B at a $5.4 billion valuation.
• •The Series B round was led by DST Global.

Tuning out the global noise and focusing on operations: Business leaders opt for pragmatism

New insight article published on 22 Sep 2026 discussing how business leaders are focusing on operations amid global noise.

Release 151: Epom Ad Server Now Sells Through Header Bidding

Epom Ad Server now sells your inventory through Prebid, ahead of Google Ad Manager. One line item, one script, five ways to integrate.

Verda Raises $189M to Expand AI Cloud and Compute Capacity

European AI cloud company Verda, formerly DataCrunch, has raised $189 million (€164.8 million) in Series B funding, propelling it to unicorn status with a valuation over $1 billion. The oversubscribed round was led by Emergence Capital, with participation from MUFG Innovation Partners, Supermicro, Varma, Lifeline Ventures, 6 Degrees Capital, byFounders, Tesi, and angel investors including Ola Tørudbakken and Mark Saroufim. This brings Verda's total funding to over $450 million (€392.3 million). The company plans to scale its full-stack AI cloud, accelerate product development, multiply compute capacity, and expand internationally, including deploying Nvidia's upcoming VR200 NVL72 generation. Verda's renewable-powered data centres in Finland support its operations across Europe, the US, and Asia. The company reported an annualized revenue run rate of $165 million (€143.85 million) in July 2026, serving customers in over 50 countries.

• •Verda raised $189 million (€164.8 million) in Series B funding led by Emergence Capital.
• •The funding values Verda at over $1 billion, making it Europe's latest unicorn, with total funding exceeding $450 million (€392.3 million).
• •Investors include MUFG Innovation Partners, Supermicro, Varma, Lifeline Ventures, 6 Degrees Capital, byFounders, Tesi, and angel investors.

AI Won’t Replace Marketers, But It Will Change What They Do

New blog post featuring Scott Brinker discussing how AI will change the role of marketers, published on 22/09/2026.

Axis's Anastasia Skliarenko Shortlisted for ExchangeWire's Ad Tech Rising Star 2026

Axis's Anastasia Skliarenko has been shortlisted for ExchangeWire's Ad Tech Rising Star 2026, as announced in a new press release on the company's newsroom.

Xiaomi MiMo-V2.6-Pro tops open-weight AI models

Chinese electronics giant Xiaomi released its MiMo-V2.6 series of open-weight AI models under the MIT license on Hugging Face. The flagship MiMo-V2.6-Pro scored 46 points on the Artificial Analysis Intelligence Index, making it the highest-ranked open-weight model globally, surpassing GLM-5.3 and Kimi K3. It trails only proprietary models like Claude and GPT-6, ranking sixth overall. The model features a sparse mixture-of-experts architecture with 1.02 trillion total parameters (42 billion active), supports text, image, speech, and video input, and offers a one-million-token context window. Xiaomi trained the models using scaled reinforcement learning, live-streaming the production run and releasing weights, technical reports, and training code. The series also includes MiMo-V2.6-Flash and a faster UltraSpeed variant. API pricing remains unchanged from the previous generation.

• •Xiaomi released MiMo-V2.6-Pro, the top open-weight AI model with 46 points on the Artificial Analysis Intelligence Index.
• •The model has 1.02 trillion parameters (42 billion active), a 1 million token context window, and multimodal input.
• •Xiaomi trained the models using reinforcement learning over 750,000 trajectories in under six days.

Xiaomi MiMo-V2.6-Pro tops open weights, trained for $3M

Xiaomi released MiMo-V2.6-Pro, a natively omnimodal open-weights model with 1.02T total / 42B active parameters, trained for $3M (about 130 hours and 75B tokens). It debuts as the top open-weights model on Artificial Analysis' Intelligence Index (46) with cost efficiency at $0.435/M input and $0.87/M output tokens, under an MIT license. Xiaomi also open-sourced the RL training environment code and recipes, but not the full 7k+ task datasets, signaling an emphasis on transparency in RL training.

• •Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash, natively omnimodal open-weights models.
• •MiMo-V2.6-Pro has 1.02T total / 42B active parameters and tops the Artificial Analysis Intelligence Index at 46.
• •RL training run cost $2.6M, used 75B tokens over 130 hours.

Cautious optimism for late-night economy after closures slow

New press release added: 'Cautious optimism for late-night economy after closures slow' dated 22 September 2026.

EU Cyber Defense Partly Effective, Auditors Report

The European Court of Auditors (ECA) released Special Report 19/2026, assessing the EU's cybersecurity cooperation framework for 2022-2025. The audit found that while progress has been made, the framework only partially facilitates detection of and response to serious cyber incidents. Key issues include limited information sharing between member states, underreporting of incidents, and significant implementation delays, particularly of the NIS2 directive, with only two member states meeting the October 2024 deadline. Cross-border incident reporting remains sparse, with only 14 reported in 2025 and none classified as 'large-scale' since 2016. The Collins Aerospace ransomware attack highlighted systemic failures, as it was not treated as significant. The ECA also criticized the underutilization of the EU Cyber Reserve, non-operational cyber shield hubs, and duplication of efforts in the Commission's cyber situational centre. Five recommendations were issued, with target dates from 2026 to 2028.

• •EU member states reported 1,276 incidents to ENISA in 2024, while ENISA identified ~4,800 from public sources.
• •Only 14 significant cross-border incidents were reported by seven member states in 2025; no incidents have been classified as 'large-scale' since 2016.
• •Only two member states met the NIS2 transposition deadline of October 2024.