🗞️ GenAI Catch-up Report 2026-09-27

5 themes from 2026-09-12 to 09-27, the first run's two weeks (selected from 221 candidates). Details: research note.

1️⃣ 🚀 Opus 5.5 and GPT-6 Sol/Luna ship on the same day, starting a price war

Opus 5.5 at $4/$20 is 20% below Opus 5; Sol/Luna drop 50%. Reactions: "Anthropic led on capability surprise, OpenAI on price".

Impact highPotential highBoth tracks

2️⃣ 🎯 Jev: a decision model that returns "typed probabilistic decisions", not text

Released by TypeSafe on 09-15: 70–500 ms, $0.042/M input. Vercel and LangChain integrated it, six clones in two days, sign-ups paused.

Impact highPotential highGenAI engineering

3️⃣ 📄 Claude Code reads AGENTS.md, and instruction files get audited

Reads AGENTS.md when there is no CLAUDE.md (2.1.277), 740 points on HN. A telemetry-dependent bug, fixed in 2.1.281, and /doctor prompt-audit.

Impact highPotential highCoding AI

4️⃣ 🛡️ Copilot, Kiro, Claude Code and Cursor ship controls for unattended runs at once

Sandboxes, OTel, .env blocking, refusing risky rm, model pinning, deploy monitoring. Boundaries and observation move into settings.

Impact highPotential highCoding AI

5️⃣ 🏗️ Azure, Google and AWS ship platform controls for always-on agents

Foundry Routines GA and egress control, Google's anomaly detection, AgentCore's multi-account setup and a new runtime (Tokyo included).

Impact highPotential highGenAI engineering

GenAI Catch-up Report · 2026-09-27 · 🤖 Generated by Claude Opus 5.5

1️⃣ 🚀 Opus 5.5 and GPT-6 Sol/Luna ship on the same day, starting a price war

On 2026-09-22 OpenAI announced about an hour after Anthropic, moving default models, prices and availability on three clouds at once.

Key points

  • Opus 5.5: $4 / $20, cache reads $0.20 (Opus 5: $5 / $25 / $0.50), 1M context, 128k max output.
  • Claimed: 40% cheaper and 30% faster than Opus 5, Terminal-Bench 4.0 at 66.4% (52.3%), OSWorld 2.0 at 81.8%.
  • GPT-6 Sol costs $2 / $10 and Luna $0.10 / $0.50, 50% below GPT-5.6's promotional prices.
  • Claude Code 2.1.280 makes Opus 5.5 the default, and Pro / Team Standard move from Sonnet to Opus.
  • Same-day availability on Bedrock (with a JP profile), Google Cloud's Model Garden, and Codex / ChatGPT Work.

Perspectives and debates

  • OpenAI: on AutomationBench, Sol (xhigh) beats Opus 5 (max) at one eleventh of the cost.
  • Willison / AINews: Luna is "one tenth" of Haiku 4.5; max "over-thinks"; "40% cheaper" is at default (medium) effort.
  • Addy Osmani / uhyo: turns drive cost; high is fine day to day, and max is an option for extra points.

"Opus 5" "Opus 5.5" "GPT-6 Sol" "Haiku 4.5" "GPT-6 Luna" 0 1 2 3 4 5 6 USD Input price (USD per million tokens)

Technical details

  • IDs: claude-opus-5-5, gpt-6-sol, gpt-6-luna; on Bedrock global.anthropic.claude-opus-5-5.
  • Opus 5.5 keeps adaptive thinking always on, Fast mode costs $8 / $40, and a classifier checks actions before they run.
  • OpenAI keeps the cache across effort and tool changes, with explicit breakpoints and a dashboard.

Sources: Anthropic · OpenAI · Simon Willison · Addy Osmani

GenAI Catch-up Report · 2026-09-27 · 🤖 Generated by Claude Opus 5.5

2️⃣ 🎯 Jev: a decision model that returns "typed probabilistic decisions", not text

TypeSafe AI released it on 09-15, and within two weeks Vercel and LangChain integrated it, six clones appeared and field tests began in Japan.

Key points

  • Takes a state and returns Choice / Score / Noul (yes or no) judgements with calibrated probabilities via parallel sampling.
  • Trained with RLCD (Reinforcement Learning for Calibrated Decisions) so that "higher confidence means higher accuracy".
  • Claimed: 70–500 ms, 40–200× faster than LLMs, $0.042 / M input, free output (about 1/48 of GPT-5.6 Terra).
  • Vercel AI Gateway (09-16, typesafe-ai/jev) and langchain-typesafe (09-17) added support; sign-ups paused by 09-22.
  • AINews: six clones within two days of launch (Laya, OpenJev, Kev-0.5B and more), and 36 million video views in two days.

Perspectives and debates

  • LangChain: "use an LLM for open-ended reasoning and generation, and Jev for fast, structured decisions along the way".
  • Copes: read headline multiples "as a ceiling"; 0% type errors cover "the shape of an answer, not whether it's correct".
  • In Japan (YESOD): 8 cases for $0.0004, no parse failures, Japanese trails English, and 49% confidence needs a plan.

yes no State + typed question Jev: judgement + calibrated probability Confidence ≥ 0.7? Handle in code Hand to an LLM or a person

Technical details

  • API: POST /v1/systemone (TypeSafe); Vercel offers experimental_evaluate in AI SDK 7 and /v1/evaluate.
  • SDKs: TypeSafeClassifier (LangChain) and llm-typesafe (Simon Willison's CLI plugin).
  • Not suited to numeric and date arithmetic, generation or summaries, images or audio, or explaining its decisions.

Sources: TypeSafe AI · LangChain · Simon Willison · YESOD

GenAI Catch-up Report · 2026-09-27 · 🤖 Generated by Claude Opus 5.5

3️⃣ 📄 Claude Code reads AGENTS.md, and instruction files get audited

2.1.277 (09-18) reads AGENTS.md when there is no CLAUDE.md: 740 points on HN, plus a 484-point bug report on telemetry.

Key points

  • Projects without CLAUDE.md now get AGENTS.md read, switchable under "Project instructions" in /config.
  • Codex, Gemini CLI, Devin, Cursor and Copilot already read AGENTS.md; Claude Code is the last to follow.
  • 2.1.281 (09-23) extends it to Bedrock / Vertex / Foundry and LLM gateways; .agents/skills is still unsupported.
  • With telemetry off (DISABLE_TELEMETRY=1 etc.) it was not read, since loading depended on a remote flag; fixed in 2.1.281.
  • /doctor prompt-audit in 2.1.283 (09-25) flags instructions written for older models and drift from reality.

Perspectives and debates

  • HN: calls for full parity over a CLAUDE.md-first rule, ln -s AGENTS.md CLAUDE.md long predates it, lock-in criticism.
  • AAIF (Linux Foundation) stewards MCP, AGENTS.md and A2A, and Agent Router 1.1 unifies APIs and MCP (Publickey).
  • Classmethod / Zenn: prompt-audit finds many gaps between the text and reality; compare behaviour before deleting.

yes no Claude Code 2.1.277 CLAUDE.md present? Read CLAUDE.md Read AGENTS.md Codex / Gemini CLI / Cursor / Copilot read the same AGENTS.md

Technical details

  • 2.1.277 also adds CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1.
  • Bug cause: agents-md loading depended on the feature flag tengu_agents_md_mod (default false), issue #95690.
  • 2.1.283: /doctor prompt-audit, availableModelsMatch, deniedModels.

Sources: Claude Code changelog · szypowi.cz · Publickey · Classmethod

GenAI Catch-up Report · 2026-09-27 · 🤖 Generated by Claude Opus 5.5

4️⃣ 🛡️ Copilot, Kiro, Claude Code and Cursor ship controls for unattended runs at once

From 09-19 to 09-25, four independent vendors shipped controls that assume nobody is watching, in four layers.

Key points

  • Copilot app: local sandboxing (files, network and credentials; preview) and OTel traces.
  • Kiro CLI 2.24: bulk approval with /tools trust-all, and .env is no longer passed to MCP servers and tools automatically.
  • Claude Code 2.1.278–283: server-side classifier as auto-mode default, risky rm denied after 2 min, managed model pinning.
  • Cursor: Rollouts (deploy monitoring that proposes revert PRs) and Security Review (vulnerability reports on every PR).

Perspectives and debates

  • All four moved boundaries and observation out of code and into settings.
  • The product answer to "Putting coding agents on a leash", a theme of Thoughtworks Radar volume 34 (April).
  • Classmethod: layered defence with hooks, permissions and sandboxes; dynamic shell expansion is a limit of static checks.

Boundary: Copilot sandbox, Kiro blocks .env Approval: Kiro trust-all, Claude Code rm denial and model pinning Observation: Copilot OTel, Claude Code prompt-id header Post-release: Cursor Rollouts and Security Review

Technical details

  • Copilot: the "Sandbox" setting, /sandbox on, and the export endpoint under telemetry in managed-settings.json.
  • Claude Code: availableModelsMatch, deniedModels, CLAUDE_CODE_AUTO_MODE_SERVER, gateway assume_role and guardrail.
  • Cursor: no auto-merge or auto-rollback; Teams / Enterprise only.

Sources: GitHub · Kiro · Claude Code · Cursor

GenAI Catch-up Report · 2026-09-27 · 🤖 Generated by Claude Opus 5.5

5️⃣ 🏗️ Azure, Google and AWS ship platform controls for always-on agents

Three vendors shipped, at different layers, features that treat an agent as an always-on actor rather than a chat partner.

Key points

  • Foundry Routines GA (09-24): agents run on timer, CRON or event triggers (GitHub issue, Teams), with an audit trail.
  • Foundry egress control (09-24, preview): FQDN allow rules applied in Audit / Enforced mode, Deny by default.
  • Google (09-15 to 09-16): Model Armor, Semantic Governance Policies, Agent Anomaly Detection (off-path OTel trace analysis).
  • AWS (09-24): an AgentCore Gateway in a central account fronts business-unit MCP servers, default-deny with Cedar.
  • New AgentCore Runtime (09-18): microVMs, usage-based billing, cold start P75 1.9–2.0 s, available in Tokyo.

Perspectives and debates

  • Foundry: "The first generation of agents waited for a person to send a message"; now they respond to business events.
  • Egress rules live outside code, tool access is default-deny at the gateway, and anomaly detection reads traces.
  • No independent analysis covering all three vendors was found in the window (unverified).

User JWT Agent: central account AgentCore Gateway: default-deny with Cedar AgentCore Identity: OAuth 2.0 M2M MCP server: business unit A MCP server: business unit B

Technical details

  • Foundry: azure-ai-projects; egress policies are ARM JSON (mode, defaultAction: Deny) attached through RaiConfig.
  • Google: ADK 1.2+, Security Command Center, business rules in plain language, closed-loop remediation.
  • AWS: Runtime / Gateway / Identity, Strands, Okta OIDC, platformVersion: V2.

Sources: Foundry · Google · AWS · AgentCore Runtime

GenAI Catch-up Report · 2026-09-27 · 🤖 Generated by Claude Opus 5.5