AI News Digest, April 25, 2026: Open-Source Frontier AI Catches Up. Same Week, HR Hiring Slows.
Three numbers carry this week’s story. DeepSeek shipped an open-source model that scores within striking distance of Opus 4.6 and Gemini 3.1 Pro at roughly one-sixth the inference cost. The World Economic Forum reported US entry-level job postings are down 35% in 18 months. And the EU AI Act’s main enforcement window opens in 99 days. The race for open-source frontier AI just got real, and it is colliding with two things HR leaders cannot ignore: a shrinking entry-level market and a regulatory deadline that quietly classifies most hiring AI as “high-risk.”
DeepSeek V4 Brings Open-Source Frontier AI Within Striking Distance of GPT-5.5
On April 24, DeepSeek released a preview of V4-Pro (1.6 trillion total parameters, 49B activated per token) and V4-Flash (284B total, 13B activated) under an open-source license. Both models ship with a native 1-million-token context window and a new “Hybrid Attention” stack that pairs Compressed Sparse Attention with Heavily Compressed Attention. The result: KV cache shrinks to roughly 10% of DeepSeek V3.2 at 1M tokens, and single-token inference uses about 27% of the prior FLOPs (Source: MarkTechPost; Simon Willison’s analysis).
On benchmarks, V4-Pro-Max scores 83.5 on OpenAI MRCR-1M and 62.0 on CorpusQA-1M. It beats Gemini 3.1 Pro High and trails Claude Opus 4.6 Max on both. VentureBeat estimates API pricing at roughly one-sixth of Opus 4.7 and GPT-5.5 (VentureBeat). Fortune flagged the close integration with Huawei silicon as a sign that China-side compute is now competitive enough to train and serve open-source frontier AI without US chips (Fortune).
Why this matters if you run HR or run a startup: the cheap-good-fast triangle just shifted again. If you have been telling your CFO that frontier-grade AI is “too expensive to use across HR ops,” that argument is on borrowed time. A 1M-token window means an employee’s entire tenure file. offer letter, every performance review, comp history, exit interview transcript. fits inside one prompt. At V4-Flash pricing, running summarization or anomaly detection across thousands of employee files becomes a question of will, not budget.
The harder question is procurement. Open-source frontier AI also means open weights you can self-host on infrastructure you control. For EU and India-based companies, that gets attractive again under the EU AI Act and DPDP-style data residency rules. For US startups on lean stacks, the hosted API at fractional cost is the easier on-ramp.
What to do this week: ask your data lead which HR data set you could not justify running through Opus or GPT-5.5 because of cost, then re-price it against V4-Flash. The answer changes whether you can run review-cycle analytics, exit-interview theming, or inbound resume triage at scale. Teams already plugged into the top AI tools for HR will move first.
Open-Source AI’s Other Side: The Entry-Level Job Market Just Flattened
Speaking on BBC Newsnight on April 22 and to BBC News on April 24, former UK Prime Minister Rishi Sunak said graduates’ concerns that AI is “flattening” entry-level hiring are “justified.” He cited private conversations with CEOs in law, accountancy, and creative industries (Personnel Today; NPR coverage of the wider tech labor squeeze). Sunak referenced World Economic Forum data showing US entry-level postings have fallen 35% in 18 months, with the British Chambers of Commerce flagging a parallel UK decline.
This is the demand-side mirror of the open-source frontier AI race. As model cost per task collapses, the lowest-rung work goes first: triage, junior research, first-draft documents. Sunak’s framing, “flat is the new up”, growth without headcount, is the planning assumption you should be testing inside your own org. For an HR leader at a 200-person company, the question is not whether to cut entry-level roles. It is what to do with the early-career talent you already have. Reskill them onto AI-supervision and customer-facing work, or watch the AI skills gap in HR harden into an attrition problem.
Frontier AI Coding Agents Hit a $25B Valuation
Cognition AI, the maker of autonomous software engineer Devin, is in early talks to raise hundreds of millions at a $25 billion valuation. That is more than double the $10.2B mark set in September 2025, and roughly six times the $4B valuation from March 2025 (Bloomberg; SiliconANGLE). Devin’s annual run-rate revenue grew from about $1M in September 2024 to $73M by June 2025, per Sacra and Crunchbase data.
For founders, this is the cleanest signal yet that investors are funding agents that close tickets, not copilots that suggest. The same pattern is appearing in HR and finance tools: vendors are now pitching agents that submit, approve, and reconcile, not just summarize. If you are evaluating new HR or finance software this quarter, force the demo to show end-to-end actions, not just chat. The market is pricing autonomy at a premium, and stacks that still need a human in the loop for every keystroke will get repriced. The same logic applies inside your own company: lean on AI agents in HR workflows to handle high-volume, low-judgment work, and free your HRBPs for the moves that actually shift the dial.
Frontier AI Compliance: EU AI Act Enforcement Hits T-100 Days
As of April 25, the EU AI Act’s main enforcement window opens in 99 days, on August 2, 2026. Article 50 transparency rules, the Annex III high-risk regime (which covers most AI used in employment, recruitment, and worker management), the regulatory sandbox framework, and the full penalty regime (fines up to 7% of global turnover) all take effect on that day (EU AI Act Implementation Timeline; Article 50; Annex III).
If you use any AI in hiring, performance management, or workforce monitoring inside the EU, August 2 is your runway. Most high-risk AI systems can self-certify under Annex VI, but the burden is real: a quality management system, technical documentation, conformity records, and human-oversight design have to be in place before the deadline. For US or UK founders selling AI tools into EU customers, the burden falls on you as the provider. Ask your AI vendor for their conformity assessment plan. If they cannot show one, that is your liability story for August.
This is also where open-source frontier AI ties back into compliance. A self-hosted V4-Flash inside your own VPC reduces some Article 10 (data governance) and Article 14 (human oversight) headaches because you control the data path. Hosted black-box APIs make the audit harder. Asanify’s global HRMS already runs with documentation and audit-trail features that map to the Annex III evidence requirements.
Quick Hits
- Google to invest up to $40B in Anthropic. $10B in cash now at roughly a $350B valuation, plus up to $30B more tied to performance milestones. Google Cloud will deliver 5 GW of compute to Anthropic over five years, starting in 2027 (CNBC; TechCrunch).
- Meta confirms 8,000 layoffs (10% of staff) plus 6,000 unfilled roles. Cuts begin May 20. Zuckerberg is roughly doubling AI capex to $115-135B in 2026 (vs $72B in 2025). Microsoft, on the same week, opened a “Rule of 70” voluntary buyout to about 8,750 US employees (CNN Business; CNBC).
- Arinox AI and KOGO begin shipping CommandCore, India’s first sovereign agentic “AI-in-a-box”. Air-gapped NVIDIA-Qualcomm hardware, KOGO OS, 80+ ready-to-deploy agents across HR, finance, and supply chain. Pricing starts at ₹10 lakh per unit; first units shipped April 19 to defense, BFSI, and government customers (Digit; NewsBytes).
The week’s open-source frontier AI moves do not change Asanify’s mandate, they accelerate it: payroll, contracts, and people data must stay private, auditable, and ready for high-risk AI scrutiny. If you are mapping your stack against Annex III evidence requirements before August, take a closer look at how Asanify’s global HRMS handles audit trails and HR data governance.
FAQ
Q: What is open-source frontier AI?
It refers to models released with open weights that match or come close to the performance of leading closed-source systems like GPT-5.5 and Claude Opus 4.6. DeepSeek V4, released April 24, 2026, is the latest example, with native 1-million-token context and inference cost roughly one-sixth of the leading closed models.
Q: When does the EU AI Act start applying to AI used in HR?
The main enforcement window opens on August 2, 2026, just under 100 days from April 25. Most AI used in recruitment, performance management, and workforce monitoring is classified as “high-risk” under Annex III, and triggers requirements for a quality management system, technical documentation, and conformity records before that date.
Q: How is AI affecting entry-level hiring in 2026?
World Economic Forum data shared by former UK Prime Minister Rishi Sunak in April 2026 shows US entry-level job postings have fallen 35% in 18 months, with parallel declines in the UK. Roles in law, accountancy, and creative industries are most affected, as employers automate junior tasks rather than backfill them.
Not to be considered as tax, legal, financial or HR advice. Regulations change over time so please consult a lawyer, accountant or Labour Law expert for specific guidance.
