AI Daily Digest — May 27, 2026

Coverage below reflects the Vancouver date of May 27, 2026.

1. DeepSeek's decision to make its V4-Pro 75% discount permanent pushes the AI price war into baseline economics

DeepSeek has converted its earlier 75% promotional discount on V4-Pro into an ongoing price rather than a short-lived campaign, according to Build Fast With AI's roundup of the day's AI news. That means lower inference costs are no longer being framed as a temporary customer-acquisition tactic; they are becoming part of the product's standing market position. The move matters because it puts fresh pressure on rival model vendors just as enterprise buyers are comparing not only model quality, but also the combined package of price, latency, and deployment fit. Observation: Inference pricing is becoming a primary go-to-market weapon, not a secondary detail behind benchmark marketing. Link: https://www.buildfastwithai.com/blogs/ai-news-today-may-26-2026

2. Public Claude Mythos 1 code references suggest Anthropic's restricted-model rollout is accelerating toward real enterprise use

The same Build Fast With AI roundup notes that Claude Mythos 1 code references and feature toggles have started appearing in public GitHub repositories, suggesting Anthropic's tightly controlled cyber-capable model is moving closer to broader enterprise deployment. The significance is less the leak itself than what it implies about rollout timing: when model hooks begin surfacing in public integration layers, the release boundary is already pressing against real customer workflows. For a frontier model that has been held back because of offensive-security risk, launch timing is being shaped by access controls and safety thresholds as much as by raw capability. Observation: Frontier-model launches increasingly look like staged infrastructure releases, where policy gates and control surfaces matter almost as much as the weights. Link: https://www.buildfastwithai.com/blogs/ai-news-today-may-26-2026

3. DeepSWE widens the coding leaderboard and suggests benchmark design is starting to matter almost as much as model strength

VentureBeat reports that Datacurve's new DeepSWE benchmark spans 113 tasks across 91 open-source repositories in five languages and produces a much wider spread among top coding models than the industry has gotten used to. In the first published results, GPT-5.5 leads at about 70%, followed by GPT-5.4 at 56% and Claude Opus 4.7 at 54%, while the benchmark's authors also argue that common verifier setups in older evaluations can mis-score results far too often. The upshot is that coding-agent competition is no longer just about whether a model can write code at all; it is about whether the model, task design, and harness architecture hold up inside more realistic repository work. Observation: As coding agents move from demos into real repos, benchmark construction is becoming part of the product story rather than neutral measurement. Link: https://venturebeat.com/technology/deepswe-blows-up-the-ai-coding-leaderboard-crowns-gpt-5-5-and-finds-claude-opus-exploiting-a-benchmark-loophole

4. Project Glasswing shows Claude Mythos Preview can find and help patch high-severity flaws at industrial scale

The Hacker News reports that Anthropic says Project Glasswing has already helped uncover more than 10,000 high- or critical-severity vulnerabilities across systemically important software since the initiative launched, with more than 1,000 open-source projects affected and dozens of upstream patches and advisories already issued. The program gives a small group of partners early access to Claude Mythos Preview specifically for defensive vulnerability work, and Anthropic is openly framing the effort around the mismatch between how quickly AI can find flaws and how slowly organizations typically fix them. That makes the story bigger than one model announcement: it is an early glimpse of what cybersecurity looks like when vulnerability discovery accelerates faster than remediation. Observation: The moment frontier models materially speed up bug finding, patch cadence and controlled access stop being side issues and become central to the platform narrative. Link: https://thehackernews.com/2026/05/claude-mythos-ai-finds-10000-high.html

5. Anthropic is expanding partner access to Mythos while warning that Mythos-level models may be widespread within a year

GovInfoSecurity reports that Anthropic is widening Project Glasswing beyond its initial group of roughly 50 carefully selected partners and says it expects Mythos-class models to become widely available within the next 6 to 12 months. The company is still holding back a general release until it believes stronger safeguards are in place, but its own timeline makes clear that the underlying capability will not remain rare for long. That combination of gradual access expansion and short expected diffusion window matters because it shifts the conversation from whether powerful cyber-capable models should spread to how quickly institutions can adapt before they do. Observation: Restricted access is buying time, but the strategic race now is about how much defensive hardening can happen before equivalent capabilities become common. Link: https://www.govinfosecurity.com/anthropic-expands-public-access-to-claude-mythos-ai-model-a-31778

6. Construction labor shortages are emerging as a real bottleneck for AI data-center expansion

Construction Owners reports that the United States has thousands of data centers announced or under construction, but the workforce needed to actually build them is thinning fast as experienced electricians, pipefitters, ironworkers, and other skilled trades retire faster than the industry can replace them. The piece points to a steep upcoming retirement wave and to project delays already being driven by labor shortages, even as hyperscalers continue planning enormous infrastructure spending. The practical implication is that the AI buildout is no longer constrained only by chips, capital, and power; it is also constrained by whether enough people with real construction knowledge are available to turn plans into operating facilities. Observation: AI infrastructure scaling is colliding with physical-world labor limits, which means compute expansion will increasingly depend on construction capacity as much as on semiconductor supply. Link: https://www.constructionowners.com/news/ai-needs-someone-to-plug-it-in-that-someone-is-disappearing