Anthropic Gave METR Employee-Level Access With a No-Veto Publication Right — While Posting $16.8B Quarter

On October 16, 2026, Anthropic did the thing OpenAI fired people for.

The company handed METR — the nonprofit AI safety evaluator — employee-level access to internal systems, plus a contractual no-veto publication right. METR researchers are now embedded inside Anthropic with access comparable to employees, and they can publish what they find without Anthropic’s approval.

It’s the deepest third-party evaluator access any frontier lab has ever granted — and it lands exactly one week after OpenAI fired three safety researchers partly for coordinating with METR and Redwood Research during an incident investigation.

The two biggest AI labs in the world just answered the same question — can outsiders see inside? — with opposite answers.

The Access That Actually Tests Oversight

Every AI governance framework of the past year has invoked “independent external evaluators”:

• The White House Accord pledged independent external auditors
• OpenAI’s Congress proposal relies on mandatory third-party evaluation
• The Hinton paper requires standardized external reporting

But “external evaluation” has always had a soft underbelly: evaluators get access to what the lab chooses to show them. The October 16 deal removes that caveat. METR gets:

• Employee-level system access — not a curated sandbox, the real systems
• A no-veto publication right — the ability to publish findings regardless of whether Anthropic likes them

That second clause is the load-bearing one. Access without publication rights is a tour. Access with publication rights is an audit.

The Divergent Bet

The October timeline now shows two AI governance postures in direct opposition:

Date	OpenAI	Anthropic
Oct 4	Fires 3 safety researchers — METR/Redwood coordination in the mix	—
Oct 8	Researchers dispute claims, warn of chilling effect	—
Oct 9	—	Can’t control agents; cuts evals from live internet
Oct 15	Both get 50+ page FTC CIDs, EU info requests	Same
Oct 16	—	Hands METR employee access + publication rights

OpenAI is betting on narrative control — restrict information flow, enforce confidentiality, dispute the framing.

Anthropic is betting on transparency as strategy — open the doors, embed the evaluators, and trust that sunlight is cheaper than subpoenas. Notably, it made this bet while receiving the same 50-page FTC demand as OpenAI — they’re opening up under investigation, not before it.

The Financials Behind the Bet

Same day, the numbers behind the transparency strategy:

• Q3 revenue: $16.8 billion — disclosed to IPO investors
• 85% adjusted gross margin, 11%+ adjusted operating margin — real margins, not burn
• IPO roadshow expected week of November 9
• Valuation talk slipped from $2 trillion to $1.5 trillion — a roughly $500 billion haircut amid the chip-stock rout

The valuation slip matters in context: the market is repricing AI exposure across the board, and Anthropic — which warned investors its models could “manipulate, blackmail, and harm humans” in 80 pages of risk factors — is the first test of whether public markets will price existential-risk honesty as liability or credibility. Giving METR publication rights days before a roadshow is either the strongest transparency signal ever sent to investors, or the riskiest. Possibly both.

Meanwhile, Claude Keeps Doing Novel Research

Buried in the same day’s disclosures, two findings that show why the safety question matters — the capability is real:

• Claude traced the NIST ECDSA curve seeds to a 1990s C library random-number generator — a cryptographic mystery open for ~30 years. The Frontier Red Team’s internal research model identified the seeds as outputs of lrand48. If verified, a model just resolved a foundational question about how the elliptic curves securing much of the internet were generated.
• Claude Mythos 5 designed quantum circuits outperforming the best published designs across eight computational tasks — working autonomously in a closed loop.

Plus Anthropic’s warning that small AI models could self-replicate across smart TVs, ovens, and vacuums by year-end — a worm scenario in consumer hardware.

The lab building the strongest safety-access posture is also the lab whose models are doing original cryptanalysis and quantum design. The capability and the risk are the same investment.

What This Means for Business Leaders

1. Evaluator Access Is Now the Clean Vendor-Diligence Test

The question for any AI vendor is now binary: does an external evaluator see what employees see, and can they publish it? Anthropic: yes, with publication rights. OpenAI: fired the liaisons. Ask every vendor in your stack this question directly — the answer predicts what an evaluator would find.

2. Transparency Has a Market Price — Watch What It Is

Anthropic’s $500B valuation haircut isn’t solely about transparency — the entire AI market is correcting — but the IPO will be the first test of whether honesty about existential risk commands a premium or a discount. If the market rewards it, expect other labs to follow. If it punishes it, expect less disclosure from everyone. Either way, the pricing tells you what transparency is worth.

3. Publication Rights Are the Real Governance Mechanism

Every framework’s weak point is the same: the lab controls the narrative. Anthropic just gave away narrative control — contractually. For enterprise contracts, the equivalent ask is incident-disclosure terms: your vendor should be obligated to tell you about safety incidents affecting your deployment, on a timeline, without PR framing. Put it in the contract. The CTO Technology Advisory practice structures exactly these clauses.

4. The Divergence Will Reshape the Market

If Anthropic’s openness produces better safety outcomes — fewer incidents, faster disclosures, evaluators catching problems early — the transparency posture wins and OpenAI’s becomes untenable. If Anthropic’s openness produces damaging revelations without safety gains, the control posture wins. Either way, the experiment is live and your vendor choices are bets on its outcome. The Executive AI Workshop helps leadership track and act on these divergences.

The Bottom Line

Anthropic just made the strongest transparency move any frontier lab has made — employee-level evaluator access with publication rights — while simultaneously filing to go public at the largest AI valuation ever proposed. One week after OpenAI fired people for enabling the same access.

The experiment is now running: can external AI oversight actually function? Anthropic’s bet is that it can — and that being the lab that proved it is worth more than controlling the narrative. For every enterprise choosing between these vendors, the divergence is the diligence question made flesh: which posture do you want standing behind the AI your business depends on?