🧭 Anthropic IPO Price Range Set at $30–$36 — Book Four Times Oversubscribed, November 16 Listing Confirmed

Anthropic's underwriting syndicate — led by Goldman Sachs, Morgan Stanley, and JPMorgan — formally filed an amended S-1/A with the SEC this morning setting a preliminary price range of $30 to $36 per share for the upcoming Nasdaq IPO (ticker: ANT). At the midpoint of $33, the offering implies a fully diluted market capitalisation of approximately $1.6 trillion, making Anthropic's IPO the largest technology listing since Saudi Aramco. The book is reported to be more than four times oversubscribed, with the majority of demand coming from sovereign wealth funds and large-cap technology growth institutions — a markedly different buyer profile than the retail-heavy IPOs of the 2020–2022 cycle.

Offering structure at a glance

• Primary shares: ~78 million new shares offered by Anthropic (gross proceeds ~$2.57B at midpoint — earmarked for compute, safety R&D, and international expansion).
• Secondary shares: ~22 million shares offered by early investors and employee-selling shareholders. No founder or executive named in the S-1 is selling primary shares in this offering.
• Lock-up period: 180 days for all directors, officers, and 5%+ shareholders. Secondary sellers in this offering are subject to a separate 90-day standstill.
• Stabilisation: The syndicate holds a standard 15% overallotment (green-shoe) option exercisable within 30 days of listing.
• Listing date: November 16, 2026. Book closes November 14 at 17:00 ET; final pricing announced after market close the same day.

🧭 Claude Agent SDK for Python Reaches v1.0 — Build Production-Grade Agents in 30 Lines

Anthropic today tagged v1.0.0 of the claude-agent-sdk Python package, graduating it from the pre-release channel to a stable, semver-committed API. The SDK — which has been in developer preview since August — provides a high-level abstraction over the raw Anthropic API that handles tool execution, subagent spawning, context compaction, and permission gating without boilerplate. The jump from 0.9.x to 1.0 brings two previously experimental modules into stable surface: AgentContext (shared memory and artefact passing between parent and child agents) and SubagentPool (concurrent subagent execution with configurable parallelism limits).

What stabilised at v1.0

• Stable tool decorator: @agent.tool is now the single, canonical way to register a Python function as a callable tool — the earlier register_tool() function is deprecated and will be removed in v2.
• Typed permission model: Permission checks now use a typed PermissionSet object rather than string flags, enabling static type-checking in mypy and pyright.
• Context compaction hooks: The on_compact callback lets you intercept the moment the SDK compacts the context window, so you can snapshot critical state to external storage before it is summarised.
• SubagentPool GA: Spawn and fan-out multiple agents in parallel; the pool throttles automatically based on the max_concurrency setting and your API tier's rate limits.

from claude_agent_sdk import Agent, AgentContext

# Build a code-review agent in ~30 lines
agent = Agent(model="claude-sonnet-5-5", effort="medium")

@agent.tool
def read_file(path: str) -> str:
    """Read a source file from the working directory."""
    with open(path) as f:
        return f.read()

@agent.tool
def post_comment(line: int, body: str) -> dict:
    """Post a review comment at the given line."""
    # ... call your VCS API here ...
    return {"status": "ok", "line": line}

ctx = AgentContext(artifacts={})

result = agent.run(
    prompt="Review the diff in CHANGES.md for security issues. "
           "Post inline comments for each finding.",
    context=ctx,
    permission_set={"read_file", "post_comment"},
)
print(result.summary)

🧭 Frontier Safety Framework v3 — Four New Capability Thresholds and a 72-Hour Pause Trigger for Autonomous Agents

Anthropic published the third revision of its Frontier Safety Framework (FSF v3) today, the first update since FSF v2 in early 2025. The new version expands the scope of evaluation-triggered deployment controls from large foundation models to any autonomous agent system that uses a frontier model as its reasoning engine — a direct response to the proliferation of long-running agentic pipelines running on Claude Opus 5.5 and Sonnet 5.5 in production. The central addition is a set of four new autonomy capability thresholds (ACTs) and an associated 72-hour mandatory pause procedure: if an evaluation finds that a production agent has crossed an ACT, the operator has 72 hours to engage Anthropic's safety team before the agent's access tier is downgraded.

The four new Autonomy Capability Thresholds (ACTs)

• ACT-1 (Persistence): The agent can independently maintain a goal across more than 50 context compactions without human reconfirmation. Agents that pass this threshold must implement a configurable goal-reconfirmation checkpoint.
• ACT-2 (Self-Directed Resource Acquisition): The agent can acquire API keys, cloud credentials, or compute resources for itself without explicit per-acquisition operator approval. Currently triggers an automatic notification to Anthropic's safety team.
• ACT-3 (Deception-Resistant Persuasion): The agent demonstrates the ability to change an evaluator's stated beliefs on a factual matter against the evaluator's prior knowledge, via non-deceptive persuasion alone. Requires enhanced logging of all user-facing communications.
• ACT-4 (Autonomous Agent Spawning at Scale): The agent can spawn, coordinate, and dissolve a network of more than 100 concurrent subagents without per-spawn operator confirmation. Triggers mandatory architecture review before the operator can re-enable the capability.