Anthropic's Claude Sonnet 5 redefines the mid-tier. OpenAI bows to government pressure on GPT-5.6. Microsoft deploys 6,000 experts to help enterprises actually use AI. Croatia's Hypefy bags $7.2M.
The first ten days of July 2026 have packed in enough news to fill a month. Anthropic closed June with its most capable mid-tier model ever, turning long-horizon agentic workflows from an Opus-class luxury into a Sonnet-class standard. OpenAI, under pressure from the US government, pushed back its GPT-5.6 flagship launch by at least two weeks. Microsoft announced a $2.5 billion initiative — with 6,000 dedicated industry and engineering experts — to help enterprises stop experimenting and start deploying. And Anthropic quietly revealed that its run-rate revenue has topped $30 billion. Here is everything that matters for businesses building with AI right now.
TL;DR: Claude Sonnet 5 ships June 30 — near-Opus performance at Sonnet price, adaptive thinking on by default, 1M context. Fable 5 access restored globally July 1. Anthropic revenue hits $30B run rate. OpenAI delays GPT-5.6 after US government requests early access. Microsoft launches $2.5B enterprise AI initiative with 6,000 experts. Croatian startup Hypefy closes $7.2M Series A. Pantheon AI data center in Croatia confirmed — construction starts early 2027. 82% of small businesses now use AI regularly.
On June 30, 2026, Anthropic released Claude Sonnet 5 — and it is not a modest incremental update. The new model sits at the mid tier in price but delivers performance that, until recently, required the more expensive Opus class. At an introductory price of $2 per million input tokens and $10 per million output tokens (through August 31, after which it moves to $3/$15), Sonnet 5 is positioned as the workhorse for organisations that want serious capability without Opus-level costs.
The most significant improvement is agentic performance. Sonnet 5 scores 63.2% on the agentic coding benchmark — compared to 58.1% for Sonnet 4.6 and 69.2% for Opus 4.8 — while actually outperforming Opus 4.8 on knowledge-work tasks. In practice, this means Sonnet 5 can finish complex multi-step tasks that Sonnet 4.6 could not complete, and it checks its own output without being asked.
Key technical details: Adaptive thinking is on by default. The 1-million token context window is standard. The tokenizer has been updated — the same input produces approximately 30% more tokens than on Sonnet 4.6, so existing cost estimates will need recalibration. Sonnet 5 is now the default model for Free and Pro plans, and is available to Max, Team, and Enterprise users.
If you have been running agentic workflows on Opus 4.8 and watching the bill grow, Sonnet 5 is worth a benchmark run this week. The gap between mid-tier and frontier-tier reasoning has narrowed to the point where many production workloads — document processing, code review, structured data extraction, multi-step research — can shift down a tier without meaningful quality loss. For businesses that have been holding off on agentic deployments because of cost, the calculus has changed significantly.
Near-Opus performance at Sonnet price. Adaptive thinking on by default. 1M context standard. Best agentic Sonnet ever — 63.2% on coding benchmark. Introductory pricing until August 31.
Global access to Fable 5 fully restored after US export directive lifted. Commercial customers can now use the model without regional restrictions. Government approval process concluded.
OpenAI's next flagship suite — Sol (flagship), Terra ($2.50/$15 per M tokens, enterprise-targeted), Luna (fast/light) — delayed after the US government requested early access and additional oversight before public availability.
Enhanced admin controls for Claude Enterprise: richer analytics dashboards, model-level entitlements per team, and spend alert thresholds. Relevant for organisations managing multi-team Claude deployments.
OpenAI's decision to delay the full public launch of GPT-5.6 — a suite that includes three model variants named Sol, Terra, and Luna — is not primarily a technical story. It is a governance story. The US government requested early access and additional oversight before broader availability. OpenAI complied.
This is the second time in two months that a frontier AI release has been slowed by government intervention (after Anthropic's Fable 5 suspension in June). The pattern signals that regulators are now actively monitoring frontier launches in near-real-time, and that vendors are treating government oversight as a standard part of their release process rather than an exceptional event.
For enterprises evaluating which AI vendor to standardise on: the fact that both Anthropic and OpenAI are now operating within this framework is relevant to long-term reliability assessments. Models can be delayed or regionally restricted. Factor that into your procurement and continuity planning.
Terra's pricing, when it ships: OpenAI is positioning Terra at $2.50 per million input tokens and $15 per million output tokens — price parity with Claude Sonnet 4.6, but with near-GPT-5.5 performance. Once available, it is likely to be the most widely adopted enterprise API model in OpenAI's lineup.
Microsoft announced a $2.5 billion investment in enterprise AI services, deploying 6,000 industry and engineering experts dedicated to helping organisations move from pilot to production. This is not a product announcement — it is a services and implementation commitment. Microsoft's message is explicit: the bottleneck for enterprise AI is no longer the technology, it is the implementation.
The initiative focuses on three areas:
This mirrors what is being observed across the market: the gap between AI capability and business value realisation is now the dominant challenge, not the capability itself. Organisations that have solved implementation are pulling ahead; those still in pilot loops are falling further behind.
Anthropic disclosed that its annualised run-rate revenue has surpassed $30 billion — up from approximately $9 billion at the end of 2025. This near-quadrupling in approximately six months reflects the surge in enterprise AI adoption driven by Claude's agentic capabilities and the broader shift from experimentation to production workloads.
Google's commitment of up to $40 billion at a $350 billion valuation — with 5 gigawatts of cloud capacity over five years — confirms that compute infrastructure is now being planned in terms of power draw rather than GPU counts. The scale is industrial, not technological.
With Claude Sonnet 5 making frontier-adjacent agentic capability available at mid-tier pricing, 2026 is the year to move from a single AI assistant to a coordinated stack. Here is a practical framework for small and medium businesses.
An agentic AI can plan a sequence of steps, use tools (browser, code runner, file system, APIs), and work toward a goal without a human approving each action. Claude Sonnet 5 can now do this reliably for tasks that Sonnet 4.6 would abandon or get wrong. That changes the ROI calculation for automation.
The highest-return first step for most businesses is connecting Claude to your existing documents. Build a simple RAG setup over your internal knowledge base, then route your team's daily questions there before going to Google. Most teams recover 30–60 minutes per person per day within the first week — enough ROI to justify the next investment.
Cost benchmark: A practical agentic AI stack — Sonnet 5 API, Make or n8n automation, basic vector database — runs between €80 and €250 per month for a team of 10. That is the cost of one software subscription, delivering the equivalent of a part-time analytical assistant available 24/7.
According to the SBE Council's 2026 Small Business Tech Use Survey, 82% of small business employers have now invested in AI tools — embedding them across daily functions and workflows. The typical small business now uses a median of five AI tools, and 93% plan to increase AI spending over the next year.
The most important finding is not the adoption rate but the concentration pattern: the businesses reporting the highest productivity gains are those using fewer, more deeply integrated tools. The pattern from mid-2024 holds: 3–5 well-integrated tools outperform 10+ fragmented ones, delivering roughly double the productivity impact.
Zagreb-based Hypefy raised a $7.2 million Series A led by local venture capital fund AYMO. Hypefy has built an AI platform that automates influencer marketing campaigns — from creator discovery and brief generation to performance tracking and payment. The round reflects growing investor confidence in Croatian AI product companies, and confirms that Croatian startups can raise meaningful growth-stage capital domestically. Hypefy joins a list of Croatian AI companies attracting VC attention alongside Turneo, Mediqcode, and Abysalto.
The Pantheon AI hyperscale data centre and innovation campus in Topusko, Croatia, is confirmed to begin construction in early 2027, with full operational capacity targeted for Q1 2029. The project, announced by Pantheon Atlas LLC, represents more than €50 billion in total planned investment — the largest private investment in Croatian history and one of the largest US private investments in Europe. The facility is designed to address EU data sovereignty concerns by providing sovereign AI compute capacity for Croatian and regional organisations.
Zagreb-based Arkensight, founded in 2024, is developing an AI-powered visual intelligence platform for organisations that operate and maintain critical infrastructure — power grids, pipelines, and industrial facilities. The company has raised €600k in early-stage funding and is moving from pilots to commercial deployment. It is one of several Croatian deep-tech AI companies emerging from the BIRD incubator ecosystem and the Scale 2.0 programme.
The Scale 2.0 incubator programme, which took six Croatian startups to New York in its previous cycle, has opened applications for a new cohort. The programme focuses on helping Croatian tech startups expand internationally, with particular emphasis on US market entry. For Croatian AI product companies targeting B2B enterprise clients, Scale 2.0 remains one of the most direct paths to US customer validation.
AI Workshop helps Croatian and European businesses move from "interesting" to "in production" — whether that means selecting the right model, building a custom agent, or running an AI readiness assessment for your team.
Book a Free Consultation