OpenAI's three-model GPT-5.6 suite finally lands. Anthropic co-founds a $1.5B deployment firm with Blackstone. Microsoft brings MCP agents to every Office app. DeepMind cracks the Math Olympiad. Croatia's Farseer raises $7.2M.
Mid-July 2026 delivered a week that will be remembered as the moment AI moved decisively from models to deployment. OpenAI shipped the GPT-5.6 family it had delayed under government oversight. Anthropic stopped waiting for enterprises to come to it and co-founded a $1.5 billion implementation company with Blackstone. Microsoft embedded MCP agents directly into Word, Excel, PowerPoint, and Outlook — making agentic AI a standard Office feature for hundreds of millions of users. And from the research world, DeepMind published a system that placed in the top 1% on International Mathematical Olympiad problems. Here is everything that matters for businesses building with AI this week.
TL;DR: GPT-5.6 family (Sol, Terra, Luna) launches July 9. Anthropic + Blackstone found Ode — a $1.5B AI deployment firm. Microsoft MCP agents go live in all Office apps July 15. xAI releases Grok 4.5. DeepMind tops Math Olympiad top 1%. ICML 2026 introduces selective activation sparsity — smaller models matching 3× larger ones. AI startup formations up 24% YoY. Croatia's Farseer closes $7.2M Series A; 48% of Croatian businesses now use AI.
The Big Story: OpenAI GPT-5.6 — Three Models, One Big Launch
After a two-week delay driven by US government oversight requests, OpenAI publicly launched the GPT-5.6 family on July 9 — three distinct models serving different segments of the market.
The frontier model — advanced agentic capabilities optimised for complex coding, biology, and cybersecurity workflows. Targets research labs, enterprise AI teams, and high-stakes professional applications.
Near-GPT-5.5 performance at a price point competitive with Claude Sonnet 5. Positioned as the primary enterprise API workhorse — the model most businesses will standardise on.
Speed and cost-efficiency over maximum capability. The natural choice for high-throughput applications — customer-facing bots, real-time classification, and latency-sensitive pipelines.
Rate-limit fallback model for ChatGPT users, now with improved tone calibration, better personalisation, and reduced factual errors. Quietly one of the most widely used models in the world.
What Terra means for the enterprise market
Terra is the model to watch closely. At pricing competitive with Claude Sonnet 5 and near-frontier performance, it directly targets the mid-tier enterprise segment that has spent the past six months choosing between Anthropic and OpenAI as a primary vendor. The business case for maintaining a single-provider stack gets harder as both providers converge on similar capability at similar price — making integration quality, data residency, compliance support, and ecosystem fit the tie-breakers.
Practical comparison: Claude Sonnet 5 benchmarks higher on knowledge work and multi-step agentic tasks; Terra is expected to benchmark better on coding and structured JSON output. If your workload is split, benchmarking both on your actual tasks remains the most reliable guide — and at current price parity, the cost of running parallel evaluations is negligible.
Anthropic Bets on Implementation: The Ode Joint Venture
On July 15, Anthropic and private equity giants Blackstone, Hellman & Friedman, and Goldman Sachs announced the founding of Ode with Anthropic — a $1.5 billion AI implementation company that deploys elite engineers directly into enterprise clients to integrate AI into core business processes.
This is not a consulting firm in the traditional sense. The model is closer to an "embedded engineering" operation: Ode teams work inside client organisations, own the implementation, and are accountable for measurable business outcomes. CEO Chris Taylor described the ambition plainly: "It's pretty easy to imagine this as a trillion-dollar company someday if we execute well."
Why this matters more than another model launch
Ode is a bet on where the next wave of AI value will be captured. The model landscape is converging — Sol, Sonnet 5, and Gemini 3.5 are all capable of handling the majority of enterprise use cases. The differentiator is no longer raw capability; it is the ability to connect AI to existing enterprise systems, data, and processes at scale. Anthropic is effectively saying: selling API credits is not enough. We need to own the deployment.
Microsoft made the same move two weeks earlier with its $2.5B Frontier Company initiative. Amazon has committed $1B to similar forward-deployed engineering. The pattern is clear: the three largest AI vendors are all betting that the next $10B in enterprise AI revenue will come from implementation, not model access.
What this means for SMBs: The enterprise implementation market is being captured by large firms with significant upfront investment. For smaller organisations, the practical implication is that AI Workshop-style regional partners — who can deliver the same depth of integration at appropriate scale and price — are more valuable, not less. The Ode model confirms that deep integration is where the value lives; it does not make that value exclusive to large enterprises.
Microsoft MCP: Agents in Every Office App
On July 15, Microsoft shipped what may be its most consequential enterprise AI update of the year: Model Context Protocol (MCP) agents now work natively inside Word, Excel, PowerPoint, Outlook, and Catalyst. For the first time, enterprise organisations can deploy custom AI agents — connected to their internal data sources, APIs, and business systems — directly within the Office apps their employees already use.
The update also includes:
- Governed agent publishing — IT admins must approve agents before deployment across the organisation, with full audit trails
- Tenant-wide custom prompt galleries — standardised, compliance-checked prompts distributed to all users from a central admin panel
- Federated MCP connector management — central governance for which external systems agents can connect to
- Sensitivity label inheritance — documents processed by AI agents automatically inherit the sensitivity classification of their source data
The practical shift: from chatbot to embedded agent
Until now, AI in Office was largely a sidebar — a chat panel you could open alongside your document. MCP agents are different: they can read the document, query your CRM, check your ERP, and write results back — without leaving the app. A contract lawyer can ask an agent to cross-reference a draft against the company's approved clause library and flag deviations. A finance analyst can ask an agent to pull live actuals from the ERP and update the Excel model automatically. This is the shift from AI as an assistant to AI as an operator.
Other Model Launches: Grok 4.5 & the Open-Source Wave
Grok 4.5 released — xAI tightens the competitive gap
xAI released Grok 4.5 on July 8, with performance improvements across coding, reasoning, and real-time information access via X/Twitter integration. Grok's competitive advantage remains its live data pipeline — the only frontier model with real-time social and news data baked in rather than bolted on.
DeepSeek V4, Qwen 3.6, and GLM-5.2 raise the open-source floor
Three strong open-source releases — DeepSeek V4, Qwen 3.6 (Alibaba), and GLM-5.2 (Tsinghua/Zhipu) — launched in the same window, each approaching frontier-tier performance on coding and reasoning benchmarks. For organisations with data residency requirements or tight cost constraints, the open-source tier is now genuinely competitive with commercial APIs for many workloads.
Google ships Omni Flash, Nano Banana 2 Lite, and Gemini Live Translate
Google's July model releases include Omni Flash (natively multimodal for dynamic video workflows), Nano Banana 2 Lite (faster image generation), and Gemini 3.5 Live Translate (real-time speech-to-speech translation across 40+ languages). Veo 3.1 retains its position as the leading AI video generation model.
Research: DeepMind Tops Math Olympiad; ICML Delivers Sparsity Breakthrough
DeepMind Mathematical Reasoning System — top 1% on IMO
Google DeepMind published results for a mathematical reasoning system that scores in the top 1% on International Mathematical Olympiad problems — a benchmark considered beyond reach for AI systems as recently as 2024. The system uses structured search guided by learned heuristics for proof strategy generation, rather than brute-force token prediction. This is relevant beyond mathematics: the same structured-search approach is being applied to protein design, materials science, and software verification.
ICML 2026: Selective Activation Sparsity
The most-discussed paper from this year's International Conference on Machine Learning introduces selective activation sparsity — a training method that teaches models to activate only the most relevant parameters for each task, rather than firing the full network for every inference. On reasoning benchmarks, models trained this way matched models three times their size in parameter count. The practical implication: significantly lower inference costs for the same output quality, which accelerates the economics of edge deployment and on-device AI.
Also from research this week: The Allen Institute found that AI hallucinations correlate with underrepresented (not absent) training data — a nuance that changes how fine-tuning datasets should be curated. MIT and Stanford researchers found that self-correction training contributes more to reasoning quality than model size alone, suggesting that smaller, well-trained models may outperform larger ones on reasoning tasks where self-checking matters.
Tutorial: Evaluating AI Models for Your Business Workload
With GPT-5.6 (Terra), Claude Sonnet 5, Grok 4.5, and strong open-source options all available at competitive prices, the model selection question has become harder — not because the models are bad, but because they are all good. Here is a practical framework for choosing.
Step 1: Classify your workload type
- Knowledge work (summarisation, drafting, research synthesis, Q&A over documents) — Claude Sonnet 5 and Terra perform similarly; Sonnet 5 leads slightly on nuanced writing tasks
- Agentic / multi-step (autonomous research, multi-tool pipelines, long-horizon tasks) — Claude Sonnet 5 leads in this category among mid-tier models
- Coding (generation, review, debugging) — Terra and Sol show strong performance; Claude Code remains the best-integrated coding environment
- Real-time / live data — Grok 4.5 is the only frontier option with native live data integration
- Structured output / JSON — Terra and Claude Sonnet 5 both perform well; test on your specific schema
- High-volume / cost-sensitive — open-source (DeepSeek V4, Qwen 3.6) or Luna for throughput workloads
Step 2: Run a benchmark on your actual data
No public benchmark reflects your specific documents, your customers' questions, or your internal data. Spend one day running 50–100 representative inputs through your top two candidate models and score the outputs manually. The result will be more reliable than any leaderboard number.
Step 3: Factor in non-model considerations
At current price parity between Sonnet 5 and Terra, the model itself may not be the deciding factor. Consider: data residency and compliance requirements, integration quality with your existing stack, enterprise support SLAs, and vendor stability. For most Croatian and EU businesses, EU data residency support (available on Azure OpenAI Service and Anthropic's EU endpoints) is a non-negotiable filter.
AI for Business: Startup Formation Up 24%, Implementation Is the New Frontier
Bloomberg reports that AI is fuelling a record number of new business formations — projected new firm count is 24% higher year-on-year in AI-adjacent sectors. The July 2026 funding landscape favours embedded AI in procurement, customer support, cybersecurity, legal services, health diagnostics, and government operations. Founders outside elite US capital networks are advised to sharpen category positioning and demonstrate paying customers early — investors are moving from "this is interesting" to "show me the ARR".
Croatia & the Region: Farseer, Abysalto & AI Adoption Milestones
Farseer closes $7.2M Series A — AI financial planning for B2B
Croatia-based Farseer, which builds AI-powered financial planning and analysis (FP&A) software for mid-market enterprises, closed a $7.2M Series A — the largest Croatian AI SaaS funding round of 2026. The platform integrates with ERP systems to automate budgeting, forecasting, and variance analysis, reducing CFO team reporting cycles from days to hours. Farseer joins Memgraph, Gideon Brothers, and Mindsmiths as one of Croatia's most internationally visible AI product companies.
Abysalto AI robodog now operational with Zagreb Fire Brigade
Croatian AI startup Abysalto has deployed its AI-powered robotic dog (Unitree hardware + Abysalto AI software) operationally with the Zagreb Trnje Fire Department. The system — 50 cm long, 15 kg, 18 km/h, equipped with LiDAR and thermal sensors — performs structural assessment in burning buildings and earthquake rubble before human firefighters enter. Abysalto is targeting fully autonomous search missions within the next 12 months and is in early discussions with fire services in two other EU countries.
48% of Croatian businesses now use AI; AI revenue reaches €2.4B
Croatia's AI adoption has crossed the halfway threshold for the business community: 48% of Croatian businesses have integrated AI into operations (up from approximately 30% in 2024). AI-driven revenue in Croatia is estimated at €2.4 billion, with total AI investment reaching €120M in 2026. The most active adopters are in financial services, manufacturing, and retail. The fastest-growing segment is SMBs implementing AI-powered customer support and workflow automation.
What to Watch Over the Coming Weeks
- Ode's first enterprise engagements — which industries Anthropic targets first with the Ode model will signal where AI implementation ROI is clearest and most defensible
- MCP adoption in European enterprises — the Microsoft MCP rollout in Office will be the largest enterprise AI deployment in history measured by user count; watch for early enterprise case studies
- Claude Sonnet 5 pricing change — introductory pricing expires August 31; if you are evaluating Sonnet 5 for production workloads, benchmark at the lower price point now
- EU AI Act enforcement — Q3 2026 — compliance obligations for high-risk AI systems begin in earnest this quarter; the first formal enforcement actions will define expectations for the year
- Open-source model quality gap — DeepSeek V4 and Qwen 3.6 are approaching commercial quality; the window for commercial providers to justify premium pricing on capability alone is narrowing
- Anthropic and OpenAI IPO signals — Ode's founding and Anthropic's $30B run rate both support a 2026 IPO filing; watch for registration activity in Q3
Ready to move from models to results?
AI Workshop helps Croatian and European businesses design, build, and deploy AI systems that deliver measurable outcomes — not just interesting demos. Whether you need model selection, a custom agent, or a full AI readiness assessment, we work at the scale that fits your organisation.
Book a Free Consultation