Google's long-awaited Gemini 3.5 Pro lands with 2M context and Deep Think reasoning. SpaceX closes the world's largest startup deal. Meta launches its first paid AI model. Claude Fable 5 joins the frontier. Global VC hits a record $510B in H1 2026.
The week of July 17–24, 2026 will be remembered as the moment the AI frontier widened in every direction at once. Google finally shipped the Gemini 3.5 Pro it had delayed for an architectural rebuild — and then followed up with Gemini 3.6 Flash just four days later. SpaceX officially closed the $60 billion acquisition of Cursor, the world's largest startup deal on record, and unveiled Grok 4.5 as the first joint model. Meta entered the paid AI market with Muse Spark 1.1. Anthropic's Claude Fable 5 arrived as a new reasoning flagship. And the numbers from the first half of 2026 made the scale of the AI investment wave undeniable: global VC hit $510 billion, driven almost entirely by AI-adjacent deals.
TL;DR: Gemini 3.5 Pro launches July 17 (2M context, Deep Think reasoning layer); Gemini 3.6 Flash follows July 21. SpaceX closes the $60B Cursor acquisition and ships Grok 4.5 MoE. Meta debuts Muse Spark 1.1 — first paid Meta AI model — with 1M context and computer-use. Claude Fable 5 available at $10/$50 per 1M tokens. Global VC: $510B in H1 2026, a new record. Together AI raises $800M at $8.3B. Croatia's Hypefy raises $7.2M Series A. ICML 2026 introduces selective activation sparsity.
After a widely-noted delay driven by a ground-up architectural rebuild, Google DeepMind shipped Gemini 3.5 Pro on July 17 — a model that represents the most significant departure from Gemini's previous architecture since the original 1.0 launch. The rebuild was not a minor version bump: Google's engineering team rewrote core components of the transformer stack specifically to support what they call the Deep Think Reasoning Layer, an explicit reasoning mechanism that separates deliberate chain-of-thought computation from standard token prediction.
2M-token context, Deep Think Reasoning Layer, autonomous workflow capabilities. Rebuilt architecture for complex multi-step reasoning, scientific research, and enterprise agentic tasks.
Speed-optimised follow-up released just four days after 3.5 Pro. Targets high-volume workloads, real-time applications, and cost-sensitive deployments on Google Cloud.
Text-to-speech model supporting 70+ languages, now in preview. Designed for voice assistants, accessibility tools, and multilingual customer service applications.
Integrated into Boston Dynamics robots via Google Cloud partnership. Enables embodied reasoning for physical-world AI systems — manipulation, navigation, object recognition in unstructured environments.
The Deep Think Reasoning Layer is Google's architectural response to the same problem that motivated OpenAI's o-series and Anthropic's extended thinking mode: for complex problems, generating an answer immediately is less accurate than generating a structured reasoning trace first. Deep Think allocates a separate compute budget to deliberate reasoning — effectively allowing the model to "think before it speaks" — and then synthesises that reasoning into a final response. Google claims this improves performance on multi-step logical inference, scientific problem-solving, and long-horizon planning tasks by 18–34% compared to Gemini 3.0 Pro on matched benchmarks.
The 2-million-token context window is not just a benchmark headline. In practice, it means an enterprise deployment can ingest an entire legal contract portfolio, a full codebase, or a multi-year customer interaction history into a single API call and reason across all of it coherently. Until now, most enterprise AI architectures required complex retrieval-augmented generation (RAG) systems to compensate for shorter contexts. With 2M tokens native, the architecture for many enterprise use cases simplifies dramatically — reducing latency, eliminating retrieval errors, and lowering engineering complexity.
Competitive positioning: Gemini 3.5 Pro directly challenges Claude Fable 5 (1M-token context, $10/$50 per 1M tokens) and OpenAI Sol in the frontier reasoning segment. Early benchmarks suggest 3.5 Pro leads on scientific reasoning and long-document tasks; Fable 5 leads on instruction-following and agentic task completion; Sol leads on coding. The enterprise decision is no longer about which model is "best" but which model is best for your specific workload.
The acquisition of Cursor (developed by Anysphere, founded by MIT researchers) by SpaceX for $60 billion in stock officially closed this week, making it the largest startup acquisition in history by a significant margin. Cursor had crossed $1 billion ARR — a milestone reached faster than almost any B2B software company on record — and the deal values it at 60× revenue, a multiple that reflects not just current performance but SpaceX's strategic bet on AI-native software development as a platform.
Days after closing, SpaceX AI (xAI's new parent brand) and Cursor jointly released Grok 4.5 — a 1.5-trillion-parameter Mixture-of-Experts model trained on Cursor's interaction dataset, the largest corpus of real-world coding sessions ever assembled. The model scored 83.3% on Terminal-Bench 2.1 — a benchmark measuring performance on complex shell, scripting, and systems programming tasks. Pricing is $2 input / $6 output per million tokens, significantly below GPT-5.6 Sol and Claude Fable 5.
Why this matters beyond coding: The Cursor training data is not just code. It contains the full context of how engineers make decisions: the queries they run, the documentation they read, the errors they debug, the architectures they choose. A model trained on this data learns software engineering judgment, not just syntax. That judgment transfers to adjacent domains — data analysis, technical writing, system design — making Grok 4.5 competitive well outside the IDE.
SpaceX stock dropped approximately $600B in market capitalisation over the four days following the announcement — a reaction that reflects investor uncertainty about a rocket company betting heavily on software AI. But the strategic logic is clear: SpaceX operates some of the most complex engineering systems on earth. Owning the best AI coding assistant — and the data advantage it generates — is a direct input to keeping that engineering complexity manageable as the organisation scales. The deal also positions SpaceX AI as a credible competitor to Anthropic Claude Code and GitHub Copilot for developer tooling at the enterprise level.
Meta has spent the past two years releasing open-source models for free. That changed on July 9 with the launch of Muse Spark 1.1 — Meta's first commercial paid AI model, priced at $1.25 input / $4.25 output per million tokens. The move signals that Meta is ready to monetise its AI capabilities directly rather than relying solely on advertising and social commerce to capture AI value.
1M-token context window. Computer-use capabilities across desktop, browser, and mobile. Ranked first on JobBench and Finance Agent V2. Meta's first paid commercial AI model.
Muse Spark 1.1's top rankings on JobBench (a benchmark measuring AI performance across knowledge-worker tasks like email triage, research, scheduling, and document drafting) and Finance Agent V2 (financial analysis, report generation, and data interpretation) position it as a strong candidate for enterprise workflow automation. At $1.25/$4.25 per million tokens, it significantly undercuts both Sonnet 5 and Gemini 3.5 Pro for organisations that do not need frontier reasoning capability but do need high-quality knowledge work automation at scale.
The computer-use capability — allowing the model to control desktop, browser, and mobile interfaces — is notable. Until now, reliable computer-use has been a differentiator for Claude Sonnet 5. Meta entering this space at a substantially lower price point will compress margins on computer-use automation deployments and accelerate adoption among cost-sensitive SMBs.
Anthropic's latest flagship model arrives with a 1M-token context window and is positioned as the strongest reasoning model in Anthropic's lineup. API pricing is $10 input / $50 output per million tokens — premium positioning that reflects its target audience: enterprise research, complex multi-step agentic workflows, and tasks where accuracy and reasoning depth matter more than cost-per-token.
Fable 5 completes a full refresh of Anthropic's model lineup: Claude Sonnet 5 handles mid-tier enterprise workloads, Claude Haiku 4.5 handles cost-sensitive high-volume tasks, and Fable 5 handles the frontier. For organisations already integrated with Anthropic's API, the model is a drop-in upgrade that does not require architectural changes — the same tool definitions, system prompts, and agentic frameworks work without modification.
The most-discussed paper from this year's International Conference on Machine Learning introduces a training technique called selective activation sparsity. The core insight: instead of activating all of a model's parameters for every inference (standard dense computation), selective activation sparsity trains models to identify and use only the subset of parameters most relevant to each specific task — leaving the rest dormant.
On reasoning benchmarks, models trained with selective activation sparsity matched the performance of models three times their size in raw parameter count. The implications for the economics of AI deployment are significant:
Also from research this week: University of Osaka researchers used an AI simulation model to resolve long-standing open questions about water's anomalous physical properties — specifically the density anomaly and compressibility behaviour near 4°C. Separately, an AI-accelerated neutron star simulation dramatically shortened the time required to model heavy element (gold, platinum) production in stellar mergers, validating observations from gravitational wave detectors.
Crunchbase data confirmed that global venture capital investment reached $510 billion in the first half of 2026 — a new record driven almost entirely by AI-adjacent deals. North America accounted for the majority, but European AI investment also hit a multi-year high. The funding landscape is bifurcating: mega-rounds (>$100M) continue at a pace seen only in 2021, while seed and Series A activity is more selective, with investors demanding demonstrated revenue or clear monetisation paths.
Together AI, which provides cloud infrastructure for training and running open-source AI models, raised $800M in a Series C at an $8.3 billion valuation. The round signals continued investor conviction that open-source model infrastructure — allowing organisations to run DeepSeek, Qwen, and other open models on managed cloud — represents a durable market, not just a fallback for teams that cannot afford commercial APIs.
Stockholm-based Neko Health — founded by Spotify's Daniel Ek — raised $700M to expand its AI-powered preventive health diagnostics platform across Europe and the US. The platform uses full-body scanning and AI analysis to detect metabolic, cardiovascular, and oncological conditions years before symptoms appear. At $700M, it is one of the largest health AI rounds in European history.
Chai Discovery, which builds AI systems for protein structure prediction and drug candidate generation, raised $400M from Index Ventures, Kleiner Perkins, and Sequoia Capital. The company's open-source Chai-1 model has been widely cited in academic drug discovery research; the commercial platform monetises workflow automation on top of it for pharmaceutical and biotech clients.
With capable models now available at $1–$5 per million tokens and mature orchestration frameworks (Anthropic Claude Code, LangChain, CrewAI, OpenAI Agents SDK), deploying a production AI agent for a small business is no longer a six-month engineering project. Here is a practical framework for getting started in a week.
The highest-ROI first agents are narrow and repetitive — tasks that take a skilled employee 30–60 minutes, happen multiple times per day, and follow a consistent process. Good first candidates:
Not every task needs a frontier model. As a rough guide:
An agent without tool access is just a chatbot. The tools that drive the most business value:
For the first version, do not deploy a fully autonomous agent. Instead, build a workflow where the agent drafts an action and a human approves it before execution. This catches errors before they have external impact, builds team trust in the system, and generates high-quality training examples for future fine-tuning. After two to four weeks of supervised operation, you will have enough data to identify which action categories the agent handles reliably and can graduate those to full autonomy.
Croatian AI startup Hypefy closed a $7.2M Series A — a significant round for Croatia's startup ecosystem, where total startup funding reached only $125M through May 2026. Hypefy builds AI-powered tools for influencer marketing and brand content performance, helping brands identify high-ROI creator partnerships and optimise content distribution across social platforms. The round positions Hypefy to expand across the DACH and wider CEE markets.
Zagreb-based Arkensight continues development of its AI visual intelligence platform for CCTV, drone, and satellite analysis — focused on asset detection and damage assessment for insurance, infrastructure, and government clients. The platform uses computer vision to automate what currently requires manual review by trained analysts, reducing assessment time from days to hours for large-scale events like natural disasters or infrastructure incidents.
Croatia now hosts at least 36 active generative AI companies (ensun.io), with the broader ecosystem including established players like Mindsmiths (enterprise AI tooling), Gideon Brothers (AI-powered autonomous forklifts), and Memgraph (graph database for AI applications). Croatia's participation in the CEE tech growth story — where AI investment is growing faster than in Western Europe from a lower base — reflects both the talent density from Zagreb's technical universities and the cost competitiveness for early-stage product development.
Gemini 3.5 Pro, Muse Spark 1.1, Claude Fable 5 — knowing which model fits your specific workload is the difference between a failed pilot and a production system. AI Workshop helps Croatian and European businesses navigate the model landscape, design the right architecture, and build agents that deliver measurable results.
Book a Free Consultation