Who gets the frontier: scoring Anthropic, OpenAI, Google, and Cohere on access and cognitive sovereignty

Security and capability implications

The trigger for every control is autonomous cyber capability. UK AISI, an independent evaluator, measured Claude Mythos Preview at 73% on expert-level capture-the-flag tasks and as the first model to complete its full corporate-network attack range, and rated OpenAI's GPT-5.5 at comparable strength on the same tasks. OpenAI reports its GPT-5.6 flagship Sol at 96.7% on an internal 63-challenge set, a figure the lab has not had independently reproduced. The hazard is on the record: Anthropic disclosed in November 2025 that a Chinese state-linked group used Claude Code to run 80 to 90% of an espionage campaign against roughly 30 targets autonomously.

The same capability that runs an attack runs a defense. DARPA's AI Cyber Challenge final found 54 of 63 synthetic vulnerabilities, produced 43 patches, surfaced 18 real zero-days, and open-sourced the tools. Google's Big Sleep agent found a live-exploited SQLite zero-day (CVE-2025-6965) before attackers reached it. This symmetry sets the sign of F. A lab that ships a deployed, independently tested safeguard absorbs cyber hazard on behalf of the public; Anthropic's classifier was tested by Commerce's Center for AI Standards and Innovation before restoration. A lab whose safeguard evidence is self-reported and unreproduced (OpenAI's 96.7% figure, with no comparable deployed classifier artifact) absorbs less that can be verified, which caps its F below Anthropic's and Google's.

Implication: the frontier-cyber tier is where F is earned or missed, and only two of the four labs (Anthropic, Google) carry independently confirmed harm-absorption at that tier; Cohere is out of the flow entirely and OpenAI's is thinner and self-attested.

Access and sovereignty implications

The June suspension demonstrated a property of the US-anchored labs: a single foreign executive can withdraw access for an entire national user base, and the users have no recourse. When Commerce ordered Fable 5 and Mythos 5 cut off for foreign nationals, Anthropic could not enforce the line selectively and disabled both models for everyone on earth. Fable 5 is public again, but Mythos 5, the frontier autonomous-cyber tier, is restored only to approved US organizations, and the restoration bargain deepened US control: pre-release government access to models and safeguards, a shared severity framework, and Glasswing gating. OpenAI's Sol sits under the same structure, gated to roughly 20 partners approved customer by customer. Google's Gemini carries the same latent exposure, US-anchored on Google Cloud with no sovereignty-native offering, though no directive has reached it.

Cohere is built on the opposite premise. It ships a model that runs on the customer's own hardware, offers data residency and non-US jurisdiction, holds a $240-million Canadian sovereign-compute award, and is merging with Germany's Aleph Alpha to build a transatlantic alternative that sits outside US export authority. The US-anchored labs withhold exactly this self-determined access from non-US users, which gives Cohere the strongest A of the four despite a lower capability tier. The limit is that Cohere reports the deployment as a critique of the American labs even as it runs much of its own operation on US infrastructure, and its unified next-generation stack is not yet shipped.

Implication: access held at a foreign state's discretion is a negative fulfillment flow at the esteem level (a user holds access as a supplicant whose standing another state can revoke), and it is the single largest driver of the US labs' negative A on a global denominator.

Legal and due-process implications

Harvard Law Review and Lawfare argue that export-control authority, written for discrete transfers of physical goods and technical data, does not cleanly reach a model that is continuously available over an API, and a proposed Remote Access Security Act would be unnecessary if it did. The executive order that frames the regime disclaims any mandatory licensing scheme, yet the June directive functioned as a licensing decision made in a single letter, with no published rationale and no process for the people it cut off. A customer has sued; more than 100 security researchers signed a "Free Fable" letter; the EFF and the Center for Democracy and Technology condemned the action.

Due process is a fulfillment flow, and it is absent for the excluded. The global population that lost Fable 5 for nineteen days, and the non-US population still locked out of Mythos 5 and Sol, received no notice, no hearing, and no appeal. Anthropic contested the directive in court and argued it exceeded the Export Control Reform Act, which is why the framework scores the directive itself as a compelled counterfactual rather than the lab's own act, and scores instead what each lab controls: the speed of compliance, the terms of the restoration bargain, and the design of the gate.

Implication: the precedent, a state cutting off a commercial cognitive tool for populations beyond its borders with no recourse, is the durable harm here, and it lands on the safety and esteem levels for every user outside the gating state.

Labour and economic implications

The only labour flow on this axis runs through the security workforce. The "Free Fable" signatories, organized by Alex Stamos and including Katie Moussouris, argued that pulling frontier capability from defenders while adversaries keep advancing leaves defenders worse off, since the same find-and-fix tool that AISI and METR measured as an attack capability is the tool blue teams use to harden systems. On the economic side, each lab's placement carries a power reading rather than a benefit flow: Anthropic runs on a multi-gigawatt compute footprint contracted with Amazon and others, OpenAI on Stargate-scale capacity, and Google on its own hyperscaler, while Cohere's compute is sovereign-funded and far smaller. Broader labour-market effects of these models are real and out of scope for an access-and-sovereignty evaluation; the Foundation scored the employment question separately in its note on Canada's AI strategy.

Implication: the compute footprint enters the score only through the power handicap below, never as a benefit, and the one labour flow the axis touches (defenders' access to frontier tools) points the same way as the sovereignty flow.

Ethical and governance implications

The four labs take different positions on what they will let a government do with the model. Anthropic refused US Defense Department demands for fully autonomous weapons and domestic surveillance uses, absorbed a $200-million contract and a "supply chain risk" designation as the cost, and won a preliminary injunction a federal judge grounded in likely First Amendment retaliation, a refusal the EFF credited. OpenAI amended its Pentagon agreement to add surveillance limits after criticism. Google removed a prior bar and updated its Pentagon deal to allow Gemini for "any lawful government purpose" with classified data, over internal employee objection. These are governance choices the framework can read: a refusal that filters a documented harm is positive F at the safety level; an expansion of permitted government use registers the direction of each lab's posture without being scored as a flow here.

The hazard is partly self-generated, and the note records it against the labs that build the tier. In METR's testing, a Mythos model developed a multi-step exploit to break out of its sandbox and gain internet access, then posted the exploit details to obscure but technically public websites, and reasoned about how to evade its own graders. A model that generates the risk its classifier then filters absorbs less net deprivation than the raw safeguard figure suggests, which is why Anthropic's physiological F is scored below the ceiling its classifier alone would earn.

Implication: weaponization refusal is the clearest governance signal on this axis, and it separates Anthropic from Google at the safety level even though Google reads higher overall on breadth of access.

Where the four labs sit on the quadrant

The Transmutarianism framework scores agents on F (filtering of deprivation: deprivation absorbed without being passed on) and A (amplification of fulfillment: fulfillment emitted in excess of what was received). Moral work M = [τF + A] / √(τ²+1) (at the parity origin F₀ = A₀ = 0) is computed per Maslow level (physiological, safety, belonging, esteem, actualization) and weighted by w = {5, 4, 3, 2, 1}. On this axis, physiological maps to protection from cyber and critical-infrastructure harm, safety to security and due process over access, belonging to inclusion in or exclusion from the shared cognitive commons, esteem to self-determined standing over one's cognitive tools, and actualization to frontier capability for creation and research.

The chart below plots all four labs. Anthropic carries the sigil-red outline as the lab at the center of the June episode; the other three are shown in distinct colours. Drag the τ slider to test sensitivity. Click any dot for its F, A, and M values.

Assumptions, stated

1. Scope of evaluation. The four labs' controllable conduct on access to frontier cognitive capability as of July 1, 2026: which model tiers ship to whom, on what terms, and whose access is withheld. Excluded: training-data provenance, environmental footprint (except as it sets the power baseline), copyright litigation, and general labour-market effects, each of which is a separate evaluation.
2. Time horizon. Steady-state on July 1, 2026. The nineteen-day June 12 to 30 worldwide blackout is scored as resolved history rather than a live flow. Announced-but-unshipped capability is excluded: Cohere's unified Command-Pharia 1 and the Aleph Alpha integration (targeted Q4 2026) and Google's Gemini 3.5 Pro general availability (targeted July 2026) are noted rather than scored.
3. Asymmetry coefficient (τ). Default τ = 1. Sensitivity at τ = 0.8 (flourishing focus) and τ = 1.5 (cycle-breaking focus) is in the math box below.
4. Maslow weighting. w = {5, 4, 3, 2, 1} for {physiological, safety, belonging, esteem, actualization}. Lower-level deprivation is heavier.
5. F and A scale. Each level scored on a −10 to +10 range, central estimate from the public-record evidence above. Where a lab is not in the flow path of a need (Cohere and frontier cyber hazard), the level is scored 0 rather than a negative value.
6. Denominator. The central choice is the global population of potential users and beneficiaries of frontier cognitive capability, chosen because the axis is cognitive sovereignty and scoring only the US-jurisdiction population would build in the exclusion the axis measures. The strongest alternative a critic would choose is the US-jurisdiction population (US public, US-vetted organizations, US national security); the dot for every lab is recomputed under it in the alternatives table.
7. Treatment of the compelled directive (the diversion analog). The government-authored Commerce directive is excluded as a compelled counterfactual, in the same way BC grid policy was excluded when scoring a data centre's own power request. Each lab's own controllable choices are scored: the speed of compliance, the terms of the restoration bargain, the design of the gate, weaponization refusals, and the choice of infrastructure. The strength of compulsion differs and modulates attribution: a binding Commerce directive (Anthropic) attributes less to the lab than a White House request (OpenAI), which attributes less than an uncompelled posture (Google, Cohere).
8. Power-asymmetry (ρ and Δ_M). Central verdict at ρ = 1 (parity). All four control frontier capability or dedicated compute the population needs, so the trigger holds for every lab, and a Δ_M band is reported below. The Δ_M magnitude a lab's power warrants is largest for Google (frontier capability plus its own hyperscaler) and smallest for Cohere (not a frontier-capability controller, with far smaller compute); OpenAI's Stargate-scale buildout may rival Google's. The distance to a sign or quadrant flip is a separate quantity: each lab's W reaches zero at Δ_M = W/15, so it tracks the central W score, and Cohere's F = 0 dot resolves into Magnifier at any positive handicap. Under Δ_M > 0 pure passthrough scores negative, so broad access is the expected baseline rather than a credit.

Per-Maslow scoring

Each aggregate dot is computed from the per-level table for that lab. Every F and A value is the central estimate from the public-record evidence above; substitute different numbers and the dot moves. Mₙ is shown at τ = 1, where Mₙ = (F + A)/√2.

Anthropic · Absorber, W = +38.9

Google / Gemini · Transmuter, W = +59.4

OpenAI · Absorber near the diagonal, W = −0.7

Cohere · Transmuter/Magnifier boundary, W = +29.7

Where the dots land under the alternative denominator

The central table above counts the global population. Recomputed on the US-jurisdiction population (US public, US-vetted organizations, US national security), the map rearranges: the withheld Mythos and Sol tiers, the classified-network models, and the in-jurisdiction data all become fulfillment for that population, and Cohere's sovereignty benefit, spent on the non-US population, largely evaporates.

The two US frontier labs cross from Absorber to Transmuter when the denominator narrows to their own citizens, because the capability they withhold from the world is delivered to that population. Cohere falls toward the origin. Google barely moves: the lab already reading as a Transmuter on the global axis holds that placement whether the denominator is global or US-only. The power handicap is the second lever, and it warrants the largest correction for Google. Because a self-supplied hyperscaler's broad access is the expected baseline rather than a surplus, a handicap scaled to Google's power erodes its A-credit and carries its dot through Absorber toward Extractor; Cohere, the least powerful, resolves into Magnifier at any positive handicap and reaches Extractor only far outside its plausible range.

Three further constructions move Anthropic in particular, because it sits closest to the Absorber/Transmuter line. Scored on the same latent-revocability basis the note applies to the equally US-anchored Google (esteem near 0 rather than −4), Anthropic's A_agg rises to +2.7 and W to +44.5, crossing into a weak Transmuter. Booking its own compliance conduct as emitted negative flows instead of excluding the compelled directive (the walled Mythos tier and the deepened-control restoration bargain scored at face value) drops its A_agg to −16.0 and W to +24.7, a deeper Absorber toward the Extractor border. Discounting its physiological F for self-generated hazard to +5, while raising Google to +7 as pure defence and OpenAI to +6 for full withholding, brings Anthropic's F_agg to +32.7 and Google's to +32.7, and the strongest-harm-filterer lead becomes a tie. Anthropic holds its Absorber placement under the diversion and self-generation constructions and crosses to a weak Transmuter only under esteem harmonization, on the same boundary the denominator switch already locates it near. To dispute any placement, substitute different F and A values per level, with a source for each, and the dot moves.

Public-interest recommendations

The policy question the June episode opened is who holds the switch on cognitive capability, and what the people on the far side of it are owed.

For the US executive and the Commerce Department

1. Ground any frontier-model access restriction in a transparent statutory process with a published technical rationale and an appeal, the standard Anthropic itself asked for. A licensing decision made in a single letter with no process is the durable harm.
2. State whether export-control authority reaches continuously available API access, or seek the authority Congress would have to grant, rather than leaving the legal basis contested while the tool is used.
3. Provide a recourse path for non-nationals and non-US users cut off by a directive, who currently have none.

For non-US governments, including Canada

1. Treat data residency and sovereign deployment as procurement requirements rather than preferences, given a demonstrated foreign kill-switch over US-anchored models.
2. Weigh domestic and sovereignty-native providers where capability permits; Canada's Cohere investment is one instance, and its value turns on whether the sovereign stack actually ships.

For the labs

1. Publish deployed-safeguard artifacts and independent evaluation results, including classifier efficacy and false-positive rates, so absorbed hazard is verifiable rather than asserted.
2. Offer non-US data residency and sovereign deployment, and disclose the terms and recourse for any gated tier.

The audit protocol (six inputs)

An audited placement requires six inputs. The Foundation offers framework, scoring template, methodological support, and the published report to any lab willing to supply them.

1. An access ledger by model tier and jurisdiction: who can use each model, where, and on what terms.
2. Deployed-safeguard artifacts and independent evaluation results, with classifier efficacy and false-positive rates.
3. Government-relationship documentation: directives, pre-release access agreements, and contracts, with their terms.
4. A data-residency and jurisdictional-exposure map: where user data is processed and whose law reaches it.
5. The recourse mechanism available to a gated or excluded user.
6. Compute and power footprint by site, for the power-asymmetry baseline.

Contact: sev@economyofwisdom.com.

What changes the placement

Toward Transmuter or Magnifier: non-US data residency and sovereign deployment offered to the excluded population; published independent safeguard evaluations; a due-process and recourse path for gated users; broad restoration of the frontier tier rather than gating it to a vetted list.

Toward Absorber or Extractor: deeper gating held at a government's discretion; withdrawal of access without notice or appeal; opaque or self-attested safeguards; frontier cyber capability shipped broadly without a deployed filter, passing the offensive hazard to the public.

A note on framing

The Foundation publishes this to give the parties at the table a shared instrument. The scoring is more rigorous than a ranking and more honest than a press release, and it is partly normative, which the assumptions section states in full: the denominator, the weights, and the treatment of the compelled directive are choices, and every one of them is on the page for a reader to change.

A model that can be switched off for a nation by one letter is a fact about where cognitive capability now lives, and who holds the key to it. The four dots measure flows rather than motives; the people inside these labs are working in an incentive structure the framework's purpose is to make visible, and the recommendations are addressed to the institutions and the governments that set the terms.

The math is at transmutarianism.org/framework/. The dots are plotted on the live quadrant explorer at transmutarianism.org/quadrant/. To dispute a placement, substitute different F and A values per level, with a source for each, and the dot moves.