AI News — June 6, 2026

Model releases

• [2026-06-03] OpenAI — shipped a substantial GPT-Rosalind update, deepening its domain-specialized life-sciences model across drug discovery, genomics, and wet-lab workflows, and opening the research preview to eligible organizations worldwide for the first time. It folds in GPT-5.5’s agentic coding and tool use atop stronger domain intelligence (medicinal chemistry, protein engineering, genomics), and OpenAI says it completes long-horizon quantitative-biology analyses using 31% fewer tokens than GPT-5.5 — a real cost lever for data-heavy research pipelines. Access stays gated behind a trusted-access program (legitimate research, biosecurity governance, enterprise-grade security). (official, source)

  This is narrow by design — most developers won’t touch GPT-Rosalind — but the pattern is worth watching: OpenAI is shipping vertical, gated frontier models (here, with a token-efficiency claim attached) rather than only general-purpose ones, which is how domain-specific agentic tooling tends to arrive ahead of broad availability.

• [2026-06-03] OpenAI — confirmed the retirement schedule for two older models: GPT-4.5 leaves ChatGPT on June 27, 2026 (30-day sunset) and o3 on August 26, 2026 (90-day sunset), as GPT-5.5 Instant settles in as the default. (source, official)

  The near-term one to act on is GPT-4.5’s June 27 cutoff — only three weeks out. Any prompt, eval, or automation still pinned to GPT-4.5 needs re-pointing and re-testing against GPT-5.5 before then; o3’s August date leaves more runway.

MCP

• [late May 2026] Security researchers — VIPER-MCP, a combined static-plus-dynamic auditing framework, scanned ~39,884 open-source MCP-server repositories and surfaced 106 zero-day vulnerabilities, each confirmed with an end-to-end exploit trace, with 67 CVE IDs assigned to date; the disclosures and CVE assignments have continued landing through early June. It’s the largest empirical look yet at MCP servers as an attack surface, on top of the 40+ MCP CVEs already disclosed earlier in 2026. (source, source) (Paper is late May; flagged here because the CVE wave is current and prior briefings only covered the spec, not the server-side security picture.)

  The practical read for anyone running MCP servers: taint-style flaws (untrusted input flowing into tool execution) are the dominant pattern, and the study confirms they’re exploitable via ordinary agent-mediated prompts, not just theoretical. Worth auditing any self-hosted or third-party MCP servers for unauthenticated exposure and input handling now, rather than waiting for the stable spec’s authorization hardening on July 28.

Watch list

• Gemini 3.5 Pro (this month): at I/O (May 19) Google said 3.5 Pro was already in internal use and limited Vertex preview, with GA “next month” — i.e. June, no committed date. The Flash tier already leads some agentic/coding benchmarks (Terminal-Bench 2.1, MCP Atlas) but regressed on hard reasoning — exactly the gap Pro needs to close, so watch the model card, pricing, and API name on day one. (source)

  The number that matters is API price as much as the benchmark grid: an aggressive price makes Pro a plausible default to swap a coding agent onto rather than just another leaderboard entry.

• xAI Grok V9-Medium (mid-June): xAI reportedly finished training a ~1.5-trillion-parameter model — about 3× its current production model and aimed at the coding lead — with a public release expected mid-June. No model card or independent benchmarks yet. (source) (unconfirmed)

  Worth watching given xAI already shipped the Grok Build coding agent in late May; a much larger model behind it would be the test of whether the “coding lead” framing is marketing or shows up on independent evals.

• Anthropic developer billing split (June 15): Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party agents reportedly move off subscription limits onto a separate monthly credit pool (~$20 Pro / $100 Max 5x / $200 Max 20x), metered at full API rates with no rollover; Anthropic still hasn’t detailed it on an official page. If it lands as described, a subscription seat no longer covers programmatic usage — worth pricing out ahead of the date. (source) (unconfirmed)

  The sharpest impact is on bursty automation — per-commit CI runs and GitHub Actions — where full-rate metering with no rollover turns a usage spike into a bill spike.

• MCP spec 2026-07-28 final: stateless core, Tasks, and MCP Apps move to stable, and clients must validate the iss parameter per RFC 9207. Server operators relying on sticky sessions will need to migrate — and, in light of the VIPER-MCP findings above, the authorization hardening in this release is worth getting ahead of. (source)