🧭 Anthropic IPO Roadshow Begins — Ticker ANT, Nasdaq Debut Targeted Week of November 16, Valuation at $1.2–1.3T
Anthropic's IPO roadshow is set to begin the week of November 9, with the Nasdaq listing debut targeted for the week of November 16 — just before Thanksgiving. The company will trade under the ticker ANT on the Nasdaq Global Select Market. As of this morning no official price range, share count, or listing date has been confirmed in an S-1/A filing, but investor presentations are scheduled to begin Monday. Bloomberg pegs the current valuation discussion at $1.2–1.3 trillion — a step back from the $2T figure floated in early October but still the largest US tech IPO since at least 2021.
Key metrics driving the roadshow narrative
- Q3 2026 revenue: $16.8B — approximately 45% quarter-over-quarter growth, the second consecutive quarter of adjusted operating profit. Q2 came in at $11.6B (the first quarter of GAAP operating profit: $974M).
- Book coverage: When the S-1 price range was confidentially submitted, demand from institutional investors was reported to be four times oversubscribed — consistent with the November 2 amended S-1/A, which added Q3 results and named retail broker-dealers (Fidelity, Schwab) for the retail tranche.
- Valuation slide: From ~$2T in early October to $1.2–1.3T now, partly driven by broader market rotation out of AI-infrastructure names over the past three weeks. The $965B private valuation from the May 2026 Series H remains the last officially confirmed figure.
- Use of proceeds: The S-1 lists compute infrastructure, safety research expansion, and international market development. No acquisitions are indicated.
What to watch during roadshow week
The amended S-1/A with an official price range will be filed with the SEC before the first institutional presentation — likely Monday or Tuesday. That filing will also confirm the total share count and the percentage of secondary shares (existing investor sell-down vs. primary issuance). Watch the spread between primary and secondary: a heavy secondary offering signals early investors cashing out aggressively; a primary-heavy split signals Anthropic retaining more cash on balance sheet. Given the stated compute-infrastructure spend, a primary-heavy split is widely expected.
IPO
Nasdaq
ANT
valuation
roadshow
S-1
Q3 2026
🧭 Claude Dreaming Engineering Deep-Dive — Wiki Construction Jumps 13%→48%, 20,000 Tokens Saved Per Query
Anthropic has published a detailed engineering post on Claude Dreaming, the async memory-consolidation system previewed at the Code with Claude conference in May 2026. The post provides the first comprehensive performance benchmarks and design rationale since the research preview launch. The headline numbers: on a 350,000-file corpus, the fraction of wiki-construction criteria met rose from 13% to 48% over 128 dreaming sessions; on a legal multi-document Q&A benchmark, accuracy hit 87% using wiki + search, versus 73% with search alone — with approximately 20,000 tokens saved per question.
How dreaming works (engineering summary)
- Input → Output: A dream is a single async job. It takes one memory store (a directory of plain-text note files) plus 1–100 recent session transcripts, and produces an updated memory store. Duplicates are merged, stale entries are replaced with evidence from newer sessions, and newly surfaced patterns are captured as new notes.
- Not weight updates: The memory store is plain text — not model fine-tuning or LoRA. Claude reads and rewrites its own notes. This means dreams are auditable: you can open the files and read what the agent "learned."
- Correction discipline: The dreaming agent only corrects a note when session transcripts provide explicit counter-evidence. It does not generalise speculatively. Hallucinated notes are therefore reduced over time (as sessions provide corrections) but never fully eliminated.
- Token economy: On long-running agent tasks, combining the wiki (retrieved relevant notes) with search saves roughly 20k tokens per query turn. At current Opus 4.7 pricing that's ~$0.12 saved per turn — material for agents running hundreds of turns per day.
- Availability: Currently a research preview via the Claude Managed Agents API. The post links to an opt-in form for production access. No GA date announced.
Implementation pattern for long-running agents
The recommended architecture from the post: schedule dreams between agent sessions, not during. A dream triggered after each session adds ~30–90 seconds of latency (depending on session count and corpus size) but does not block user-facing responses. For agentic pipelines that run overnight batch work, trigger a dream at the end of each nightly run so the morning session starts with a consolidated, up-to-date memory store.
# Pseudocode — triggering a dream via Managed Agents API
import anthropic
client = anthropic.Anthropic()
dream_job = client.managed_agents.dreams.create(
agent_id="my-agent-id",
memory_store_id="my-memory-store",
session_ids=recent_session_ids, # last N sessions
max_sessions=50, # cap to control cost
)
# Poll for completion (async — typically 30–90s)
result = client.managed_agents.dreams.wait(dream_job.id)
print(f"Dream complete. Notes updated: {result.notes_updated}")
Claude Dreaming
Managed Agents
memory consolidation
agent memory
token efficiency
research preview
wiki construction