JadidJust-in AI Developments, Incidents & Deadlines
May 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

AI news, Thursday, May 28, 2026

12 stories

TLDR

  • Cognition raises $1B at $26B post-money, ARR now $492M, enterprise usage up 50% MoM for 6 months running (May 27) Trending — Lux Capital, General Catalyst, and 8VC led; Devin maker more than doubled its valuation since September's $10.2B round, with Mercedes-Benz, NASA, Goldman Sachs, and Santander cited as customers. (TechCrunch | Bloomberg)
  • Gemini Interactions API: new steps schema is the default as of May 26; legacy outputs schema is removed June 8 (May 26) Pricing — Hard two-week migration window. New features ship only to the new schema. (Google AI — migration guide)
  • Claude Code v2.1.149–152 ships a PowerShell permission-bypass fix and a per-category /usage breakdown (May 25–27) Security — Built-in PowerShell cd.., cd\, cd~, X: could change the working directory outside the workspace without triggering the permission system. Update Windows installs immediately. (Claude Code changelog)
  • Microsoft Copilot Studio: computer-using agents now generally available across commercial Power Platform geographies (May 26 post; GA from May 13) Agents — Vision-driven UI navigation across any application, plus Claude Sonnet and OpenAI computer-using model both supported; sub-500ms real-time voice. (Microsoft Copilot blog)
  • Tensormesh raises $20M (AMD Ventures, CoreWeave, NVentures) for a KV-cache-native inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ platform; team behind LMCache (May 27) Infra — Claims up to 10× reductions in latency and GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ spend by reusing cached attention state. (BusinessWire | HPCwire)

Action Items

Things your engineering team should look at or do this week.

  • Migrate Gemini Interactions API to the new steps/response_format schema before June 8 — Default switched May 26; legacy outputs schema is permanently removed June 8 (two weeks). You can pin the legacy with header Api-Revision: 2026-05-07 only until that date, and any new Interactions API features land only on the new schema. Concrete steps: (1) read response content from steps[] instead of outputs[], handling both user_input and model_output step types; (2) move response_mime_type settings into the new polymorphic response_format; (3) update function-call parsing to find function_call step types in steps[]; (4) audit any prod gateways still on the May 7 revision header and clear them before June 8. (Google AI — Interactions breaking-changes guide | Gemini API release notes)

  • Update every Windows Claude Code install to v2.1.149+ today; v2.1.152 is current — The PowerShell permission bypass let cd.., cd\, cd~, and drive-letter switches (X:) change the working directory outside the workspace without firing the permission prompt, and a related parsing gap trusted stale PWD/OLDPWD/DIRSTACK values across cd/pushd/popd. The v2.1.149 release also tightened the git worktree sandbox write allowlist. Roll-out: claude update or winget upgrade Anthropic.ClaudeCode; for managed enterprise installs, push v2.1.152 through your endpoint management and confirm the new /usage per-category breakdown is visible (skills, subagents, plugins, per-MCP server) so engineers can self-diagnose limit hits without opening tickets. (Claude Code changelog)

Models & Tools

  • Microsoft Copilot Studio — computer-using agents GA + redesigned workflows + real-time voice (May 26 update post; GA effective May 13) Agents Production — Vision-and-reasoning agents navigate live UIs across browser, desktop, and any app that renders a screen — no API or selector required. The May 2026 release also ships a rebuilt workflow canvas with conditional branching and parallel execution, Work IQ extensibility for workforce-analytics signals, and real-time voice with sub-500ms latency. Notable model-platform move: Copilot Studio now supports OpenAI's computer-using model and Anthropic's Claude Sonnet side by side. Live in US, EU, APAC, UAE; government cloud rollout staged for H2 2026. (Microsoft Copilot blog)

  • Claude Code v2.1.149 → v2.1.152 — /usage per-category breakdown, keyboard diff scroll, GFM task lists, enterprise MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ setting (May 25–27) Security — Headline dev-facing changes: /usage now decomposes limit consumption by skills, subagents, plugins, and per-MCP-server costs; /diff is keyboard-navigable (arrows, j/k, PgUp/PgDn, Space, Home/End); Markdown output renders GFM - [ ] / - [x] task lists as actual checkboxes; new allowAllClaudeAiMcps managed setting lets enterprise IT load Claude.ai cloud MCP connectors alongside managed-mcp.json. (Claude Code changelog)

  • Tensormesh launches KV-cache-native inference platform; $20M extension from AMD Ventures, CoreWeave, NVentures (May 27) Infra Open Source — Built by the team behind LMCache, the leading open-source KV-cache project. Claim: up to 10× reductions in latency and GPU spend by storing and reusing computed attention state across requests. Total raised now $24.5M. Hardware integration roadmap explicitly names AMD, CoreWeave, and NVIDIA — i.e., not just a CUDA story. (BusinessWire announcement | HPCwire | SiliconANGLE)

  • LangChain langchain-perplexity 1.3.0 (May 27) Open Source — New use_responses_api flag for ChatPerplexity switches to Perplexity's Responses-style endpoint, plus routine dependency and infra updates. Worth a look if you're routing through Perplexity for grounded retrieval — the Responses API matters for structured tool use. (LangChain releases)

APIs & Pricing

  • Gemini Interactions API — new steps schema becomes default May 26; legacy outputs removed June 8 (May 26) Trending — The change restructures the API shape to support mid-flight steering and async tool calls. Two relevant headers: Api-Revision: 2026-05-20 opted you in early; Api-Revision: 2026-05-07 is the only escape hatch and stops working June 8. Any feature shipped after May 7 is steps-only. Treat this as a hard deadline — Google did not give months of runway. (Google AI — migration guide | Gemini API release notes)

Security

  • Claude Code v2.1.149 — PowerShell permission-bypass fix (Windows) (May 25) Production — Built-in PowerShell cd variants (cd.., cd\, cd~, X:) could change the working directory outside the workspace without going through the permission prompt, and a related fix corrected stale-environment trust of PWD/OLDPWD/DIRSTACK across cd/pushd/popd. The release also tightened the git worktree sandbox write allowlist (previously covering the entire main repo root rather than just the shared .git directory). No public CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ assigned at time of writing; treat as a high-priority Windows update regardless. (Claude Code changelog)

Industry & Business

  • Cognition raises $1B+ at $26B post-money; $492M ARR; 50% MoM enterprise growth (May 27) Trending — Lux Capital, General Catalyst, and 8VC co-led; Founders Fund, Ribbit Capital, Atreides, and Layer Global joined. That's a 2.5× jump from the September $10.2B post. Customer list cited in the round: Mercedes-Benz, NASA, Goldman Sachs, Santander. The signal is the ARR figure paired with the 50% MoM enterprise-usage growth over six months — autonomous coding agents are now landing real production budgets, not pilots. (TechCrunch | Bloomberg)

  • Anthropic $30B+ round at $900B pre-money officially closed during week of May 26 (May 26–27) Trending — Update to the earlier May 22 reporting: Sequoia, Dragoneer, Altimeter, Greenoaks each ~$2B; Microsoft, NVIDIA, Founders Fund, General Catalyst also in. First time a private AI company tops OpenAI's $852B mark. Expected to be the last private round before an October 2026 IPO window. (Bloomberg | Yahoo Finance)

  • China widens AI travel restrictions — government agencies imposing their own curbs on top of private-firm directives (May 26) — Initial Bloomberg report on DeepSeek/Alibaba researcher curbs now appears to be only one layer; Chinese government agencies are separately restricting researchers they employ or fund. The talent-flight playbook used to acquire frontier expertise from other countries doesn't apply here. (Bloomberg)

  • Samsung consumer-electronics union files court motion to block AI-chip pay-deal vote (May 26) Hiring — South Korea. The proposed package disproportionately benefits the semiconductor division, whose AI-chip output for NVIDIA and Google has become the profit engine of the company. Watch as a leading indicator for the compensation pressure forming inside every company where a specific technical division has become the global AI bottleneck (TSMC, ASML, others). (Reuters via Yahoo Finance)

Policy & Safety

  • NYT: surge in AI-generated pro se filings is straining federal and state courts (this week) Trending — Pro se civil filings up 20–40% in several district courts as self-represented litigants use ChatGPT, Claude, and others to draft complaints, motions, and briefs. Mixed picture: technically competent pleadings that previously got dismissed on form are now surviving initial screening, but a meaningful fraction contain hallucinated citations or legally specious claims that still consume judicial time. Net-positive vs. net-negative depends on whether court funding scales with volume — and it isn't. (New York Times)

Benchmarks & Evals

  • Peer-reviewed study (n>100,000): generative AI outperforms the average human on standardized creativity tests (May 27) — Headline measures: Alternative Uses Task, Remote Associates Test, divergent-thinking scoring. The authors are explicit that this measures structured creative output — the kind these tests are validated to predict for hiring — not embodied, culturally-embedded creativity. The hiring implication is real: AI now beats the average candidate on the exact tests that screen for creative potential, at a sample size large enough that the result isn't a small-benchmark artifact. (ScienceDaily writeup)

Type to search · ↑↓ to move · enter to open