The cyber security reverberations of Claude Mythos have drawn most of the commentary this past week, but there is a quieter shock buried in the same episode. For the first time since the current wave of frontier models emerged, an AI company has chosen to restrict access to its most capable tool rather than race it to market at the lowest possible price. If that choice sticks, every enterprise buyer that has treated AI capability as a commodity being deflated towards zero needs to revisit the assumption this quarter.
The deflationary era is ending
Until this month, the dominant pricing story around frontier AI was straightforward. Anthropic, OpenAI, Google, and a shrinking pack of challengers were locked in a deflationary spiral, cutting per-token prices at every major release and competing aggressively on coding, reasoning, and agentic benchmarks. Buyers priced the next eighteen months of capability on the assumption that the trend would continue: cheaper models, broader access, more automation surface area.
The Financial Times has now pointed out what treasurers and procurement leads should already have been modelling. Anthropic released Claude Mythos to a handful of tech customers only, citing cyber security concerns. OpenAI followed within days with a similar limited release. Both companies cited capacity constraints alongside safety considerations. Demand for top-tier coding agents, the FT reports, already outstrips supply. The economics are shifting from abundance to rationing, and the strategic implications for buyers are considerably less reassuring than the per-token chart suggests.
Strategic Reality: Commoditisation was the working assumption behind most enterprise AI budgets, vendor strategies, and build-versus-buy decisions made in the past eighteen months. If limited access and scarcity pricing become standard practice at the frontier, large parts of that planning now rest on a broken premise.
The numbers buyers need to see
The evidence is sparse, because the vendors are being deliberately opaque. But the publicly available signals align and are worth laying out explicitly.
| Signal | Implication |
|---|---|
| High-severity vulnerabilities claimed found by Opus 4.6 | 500 |
| Scope of Claude Mythos initial release | Small number of technology customers |
| OpenAI equivalent release profile | Also limited, announced within days |
| Anthropic credits for Mythos-based vetting | $100m (offered to customers) |
| Verification status of Mythos test results | Not disclosed; unable to be independently validated |
| Direction of historical frontier pricing | Deflationary — consistent downward pressure per token |
| Direction of current frontier access | Contracting — capacity allocation now a vendor choice |
The headline figure most boards will fixate on is the $100m in credits. It looks generous. Read in context, it is a tell. A company releasing a model broadly does not need to offer nine-figure credit packages to convince anyone of its value. The credit offer is a hedge against the accusation that the restricted release is profiteering dressed up as responsibility — an accusation the FT is the latest to raise.
What is really happening behind the safety framing
Bruce Schneier’s observation in the FT is the most important single sentence in the coverage so far. The warning Anthropic sounded about Mythos, he notes, could equally well have been sounded six months ago or six months from now. That is not a dismissal of the cyber threat. It is a reminder that the timing of the disclosure was a choice, and choices are strategic.
Three things are happening simultaneously. Each reinforces the others.
First, compute scarcity is real. Training and serving the largest models now consumes roughly the entire output of the advanced chip supply chain, and Anthropic is publicly straining to meet existing demand for its coding agent. Restricting Mythos is partly a capacity decision that would have been made even with no safety considerations at all.
Second, differentiation is hard. The persistent worry for model builders chasing an initial public offering is that there is no durable moat around their technology. If every frontier vendor ships broadly comparable capability at broadly comparable prices, margins collapse. The mystique around a restricted-access, security-critical model is a useful counterweight to that commoditisation narrative.
Third, customer economics are starkly asymmetric. A software vendor that can tell its customers “Mythos-vetted” has a meaningful pitch. A rival that cannot is exposed. For the AI company sitting in the middle of that asymmetry, access becomes a lever rather than a product feature.
Competitive Reality: The enterprise buyers who will get access to the next generation of scarce frontier models are not the ones with the largest budgets. They are the ones with the tightest, most strategic relationships with the vendors, and the most mature internal governance frameworks the vendors can defend on the front page of a newspaper.
Why this changes the enterprise AI operating model
Most UK organisations currently treat AI model selection as a technology procurement question: which provider, which model tier, which commercial terms. If the scarcity thesis holds, the question becomes structurally different. It moves closer to how banks handle clearing relationships or how manufacturers handle rare-earth supply: access first, price second, commodity comparison third.
The boards that will feel this first are those whose products depend on embedded AI features. If your security software, your legal research tool, or your customer communication platform relies on a specific vendor’s frontier model to deliver competitive capability, and that vendor can now choose who to sell scarce capacity to, you have a supplier concentration problem that did not exist in the same form twelve months ago.
Stakeholder exposure under scarcity pricing
The impact is not uniform across the organisation. Different functions face different risks, and several of them currently sit outside the usual AI governance conversation.
| Stakeholder | Exposure under scarcity | Current governance coverage | What needs to change |
|---|---|---|---|
| Chief technology officer | Direct risk on model access and substitution | Typically covered | Add access-risk modelling to vendor reviews |
| Chief information security officer | Dependency on AI-driven vulnerability vetting | Emerging | Formalise position on “AI-vetted” claims from suppliers |
| Chief financial officer | Budget assumptions built on deflating prices | Usually not covered | Stress test AI spend against flat or rising unit costs |
| Chief procurement officer | Single-vendor concentration across AI stack | Often not covered | Introduce concentration limits and second-source plans |
| Product leaders | Features dependent on specific vendor capability | Partially covered | Document which features require which tier of access |
| Legal and compliance | Disclosure obligations on AI supplier access | Minimal | Review whether supplier access risk is a material disclosure |
The CFO and CPO rows are where the strategic action is. Most UK boards have mature frameworks for supplier concentration in every domain except AI, where concentration has been normalised because the alternative — multi-vendor AI architectures — felt expensive and unnecessary during the deflationary phase. That calculation changes when access becomes the scarce resource rather than compute or budget.
Strategic recommendations by organisational maturity
The right response depends on where the organisation currently sits in its AI adoption. Below is a sequenced framework, ordered so that earlier stages build the foundations later stages need.
Early stage (pilots and initial deployments):
- Avoid commercial commitments that assume continued price deflation at the frontier.
- Document which pilots are model-agnostic and which depend on specific vendor capability.
- Negotiate contractual clarity on access tiers and escalation rights, not just rate cards.
- Build internal evaluation capacity that can compare models without vendor-supplied benchmarks.
Mid stage (AI embedded in core products or operations):
- Conduct an explicit supplier concentration review across the AI stack.
- Identify the three or four capabilities where loss of frontier access would be materially damaging, and document the substitution path for each.
- Engage directly with vendor strategic account teams on access commitments, not just commercial discounts.
- Revisit any public statements about AI capability that assume continued access to the top tier.
Mature stage (AI as core competitive advantage):
- Treat frontier access as a board-level supplier risk, with the same rigour as any other critical dependency.
- Maintain meaningful relationships with at least two frontier vendors, even where one is clearly preferred commercially.
- Invest in open-weight alternatives for capabilities that may become access-rationed, even if they currently lag on benchmarks.
- Build the internal discipline to walk away from a vendor relationship that becomes access-coercive, and make sure the vendor knows it.
Implementation Note: The most common failure mode in the next twelve months will be organisations that wait until they lose access to a critical capability before they start thinking about substitution. By then the market will be in the same position, and the substitution options will have contracted in parallel.
Hidden challenges leaders are underestimating
Beyond the immediate procurement response, four second-order problems deserve attention. Each compounds quietly and each is easier to address now than after an access shock.
Disclosure creep in marketing claims. Vendors whose products depend on restricted-access AI will be tempted to make “AI-vetted” or “Mythos-vetted” claims that outrun the actual access they have. Legal and marketing teams need a clear policy on substantiation, because the gap between marketing language and procurement reality will be the first place that customer trust breaks.
Regulatory asymmetry. UK and EU regulators are already examining concentration risk in cloud services. Frontier AI access, where a handful of vendors choose which customers can use the most capable tools, is a natural next topic. Organisations that build their AI operating model on vendor-selected access are assuming regulatory tolerance that cannot be taken for granted.
National security overlap. The FT piece notes Anthropic’s existing confrontation with the Pentagon over national security concerns. Similar tensions will emerge in the UK. Suppliers to regulated sectors and government may find that their preferred AI vendor’s access decisions are made with foreign national security considerations in mind, not commercial ones.
Skill atrophy under vendor dependence. Organisations that outsource all frontier capability to a single vendor stop building the internal capability to evaluate models, negotiate access, and design graceful degradation into their products. When access finally contracts, they discover they have lost the organisational muscle needed to respond quickly.
Warning: ⚠️ The combination of vendor concentration, regulatory uncertainty, and skill atrophy is the same combination that caught out organisations during the cloud concentration debates of the past decade. Treating frontier AI access as if it is a more benign version of the same dynamic is the error to avoid.
The strategic takeaway
Claude Mythos may or may not be the inflection point at which AI economics formally moved from deflation to scarcity. But it is the clearest signal yet that frontier vendors now consider limited access a viable strategic tool, and that the capacity constraints behind that choice are not going away on the timeline most enterprise buyers have assumed. The cyber security story that has captured the headlines is real. The economic story underneath it is, for most UK boards, the one that actually shapes the next eighteen months of decisions.
Three factors separate the organisations that will adapt well from those that will be surprised:
- Clarity on dependency: Know exactly which parts of the product and operating model depend on which vendor and which access tier, documented at a level a new CFO could understand in an afternoon.
- Substitution readiness: Have a credible second source for each critical capability, even if it is slightly more expensive or slightly less capable, so that no single vendor’s access decision can stop the business.
- Honest engagement with the vendor: Treat the relationship as a strategic supplier partnership, not a self-service cloud service, and invest accordingly in the governance and commercial relationships that give the organisation standing when access conversations get serious.
Next steps for your organisation
- Map current AI supplier concentration across product, security, and operations, and identify the three capabilities most exposed to an access shock.
- Stress test the next two years of AI budget against flat or rising unit costs rather than assumed deflation.
- Review any public or contractual commitments that assume continued access to a specific vendor’s top tier of capability.
- Establish a board-level view on whether frontier AI access is a material supplier risk that should appear in risk reporting.
- Identify one critical capability to build a second-source plan for this quarter.
The deflationary era of frontier AI was genuinely useful for enterprise buyers while it lasted. Treating it as a permanent condition, rather than as the opening phase of a market that is now repricing itself around scarcity, is the mistake that will cost the most in the next two years.
For further analysis of AI supplier risk and enterprise procurement strategy, see our insights archive. To discuss your organisation’s exposure to frontier AI access decisions, get in touch.
Analysis based on reporting by Richard Waters in the Financial Times, 20 April 2026 (original article), including commentary from US security expert Bruce Schneier and industry context around the limited release of Anthropic’s Claude Mythos and OpenAI’s comparable model.
Resultsense is a UK publication making sense of AI for professionals and businesses. This article represents independent editorial analysis.