🧭 Anthropic Grants METR Employee-Like Access — The Deepest Independent AI Safety Evaluation Arrangement Yet
Anthropic has published a detailed update on how it is opening its operations to independent external evaluators. The headline arrangement: six staff from METR (a nonprofit frontier AI evaluator) now hold what Anthropic describes as "employee-like access" — company hardware, office space, and internal system access roughly equivalent to Anthropic's own risk-assessment teams. METR can publish risk assessments freely, with Anthropic only able to require redaction of narrowly defined sensitive categories. This goes significantly further than the standard NDA-gated model access that has characterised most third-party evaluations to date.
A three-tier framework for external accountability
Anthropic is formalising graduated access levels that other evaluators can progress through:
- Tier 1 — Narrow investigations: Anthropic-approved publication rights; evaluators work on specific, scoped questions
- Tier 2 — Broader information access: Independent publication rights; more of Anthropic's internal data is available
- Tier 3 — Full embedded evaluation: The METR arrangement — hardware, office space, internal access, nearly unrestricted publication
Other evaluators — SecureBio (biosecurity) and Accenture (enterprise risk) — currently operate at lower tiers and could advance over time. The arrangement is framed as a structural response to the "White House Accord on Super Intelligence," which Anthropic signed.
What METR cannot see
- Customer data
- Attorney-client communications
- Model architecture details
Why this matters for anyone building on Claude
The METR arrangement changes the information asymmetry between Anthropic and external parties. Until now, safety claims about frontier models were largely self-reported. Embedded independent evaluators with free publication rights create a structural check: if METR finds a risk that Anthropic hasn't publicly acknowledged, they can say so. For enterprises under regulatory scrutiny — financial services, healthcare, critical infrastructure — the existence of a credible third-party programme is increasingly a procurement prerequisite, not just a nice-to-have. Watch whether this becomes a template that other frontier labs are pressured to match.
AI safety
METR
independent evaluators
governance
accountability
third-party audits
White House Accord
🧭 Claude Code v2.1.300 — OpenTelemetry GenAI Attributes Land, /advisor Expands to Claude Platform on AWS
Claude Code v2.1.300 (released 14 October) ships two additions with meaningful impact for teams running Claude Code in production CI/CD pipelines: full OpenTelemetry GenAI semantic-convention attribute support on trace spans, and /advisor availability on Claude Platform on AWS. The release also includes 88 total changes across additions, modifications, improvements, and bug fixes.
OpenTelemetry GenAI attributes — what's new
Claude Code now emits standard OpenTelemetry attributes on every trace span, letting you correlate agent runs with your existing observability infrastructure without custom instrumentation:
gen_ai.provider.name — which model provider was usedgen_ai.conversation.id — stable identifier for the conversation threadgen_ai.tool.name — the specific tool invoked per spanerror.type — structured error classificationcache_creation_5m_tokens and cache_creation_1h_tokens — cache write volumes broken down by TTL, added to the api_request event
Practical setup: routing Claude Code spans to your OTEL collector
Set OTEL_EXPORTER_OTLP_ENDPOINT to your collector before launching Claude Code. The new gen_ai.* attributes map directly to the OpenTelemetry GenAI semantic conventions, so dashboards built for other GenAI tooling (LangChain, the OpenAI SDK) will display Claude Code spans without schema changes. The cache TTL breakdown lets you track whether your prompts are hitting the 5-minute or 1-hour cache tier — a useful signal for optimising long-running agent pipelines where cache misses have real cost implications.
# Enable OTEL export before running Claude Code
export OTEL_EXPORTER_OTLP_ENDPOINT="http://your-collector:4318"
export OTEL_SERVICE_NAME="claude-code-ci"
claude --print "run build and tests" 2>&1
/advisor on Claude Platform on AWS
The /advisor feature — which lets the primary model consult a more capable secondary model mid-task — is now available on Claude Platform on AWS. It is disabled by default and must be explicitly enabled via the /advisor slash command, the advisorModel setting, or the --advisor CLI flag. You can also hard-disable it with CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 if your security policy requires predictable model usage.
Claude Code
OpenTelemetry
observability
tracing
AWS
/advisor
CI/CD
v2.1.300
🧭 Early Anthropic Backer Predicts $2.4 Trillion IPO Pop — Enterprise Mix and Profitability Speed Cited as Differentiators
Fortune spoke with an early Anthropic investor (on condition of anonymity) who expects Anthropic to price its IPO shares around a $1.8 trillion valuation but anticipates trading to reach $2.3–$2.4 trillion once public markets open — a prediction that would surpass SpaceX's opening-day high of $2.1 trillion. The backer cited Anthropic's 80% enterprise revenue mix, two consecutive quarters of operating profitability, and what they described as the company running "a little less chaotically" than rivals as the foundation for that outlook. Note: these are investor predictions, not stated guidance from Anthropic. Treat all valuation figures as estimates.
Context: what's known versus what's estimated
- Known (from Anthropic's investor day, 14 Oct): Q2 2026 revenue exceeded $11.5B; two consecutive profitable quarters; November Nasdaq listing on track
- Estimated (anonymous VC prediction): $1.8T IPO price; $2.3–2.4T trading peak; "fastest-growing company in history" characterisation
- Anthropic has not confirmed any specific IPO valuation target in public statements
Why enterprise concentration matters for this valuation story
The 80% enterprise revenue figure is the underappreciated detail here. Consumer AI products (Claude.ai subscriptions) typically command lower revenue-multiples than B2B SaaS because of higher churn and margin pressure. A model with 80% enterprise revenue — long-term contracts, procurement budgets, lower churn — is more comparable to Salesforce or ServiceNow than to a consumer app. If that mix holds through the IPO, it's the structural argument for why the valuation multiples suggested by this investor might not be as absurd as critics contend. That said, Anthropic's S-1 will be the document to watch — particularly the concentration risk disclosures flagged at the October 14 investor day, where two clients were noted to represent roughly 25% of total revenue.
IPO
valuation
Nasdaq
enterprise
revenue
investor
funding
