By Ajay Arora, Chief Technology & Strategy Officer, Centroid
Why Anthropic built a model it wouldn’t sell, how the program evolved from April to today, and where the real services opportunity lies for a technology consultancy.
Setting the stage
There’s a version of the Claude Mythos story that reads as just another frontier-model release. That version misses the point. The more useful way to understand Mythos — especially for a consulting firm deciding how to position around it — is a deliberate experiment in how to bring dangerously capable technology into the world. Anthropic built a model it judged too powerful to release, said so publicly, spent months proving its value in a tightly controlled setting, and only then shipped a public version engineered to suppress the very capabilities that made it dangerous.
For technology leaders, interesting questions aren’t really about benchmarks. They’re about the reasoning that produced this unusual sequence of decisions, the roadmap that has unfolded over the past two months, and — most importantly — what all of it means for firms that advise clients on how to adopt AI safely and profitably. This write-up works through all three: the thinking behind Mythos, the roadmap to today, and a detailed consulting outlook.
Part 1 — The thinking behind Mythos
A capability that arrived before its safety case
The intellectual starting point for Mythos was recognition, not ambition. When Anthropic trained the model that became Claude Mythos Preview, it discovered the system was strikingly capable at computer security — able to find and exploit software vulnerabilities at a level that rivaled or exceeded all but the most skilled human experts. In internal testing it autonomously surfaced thousands of high-severity flaws, including vulnerabilities in every major operating system and web browser.
That single fact reframed everything. A model this good at security is not a neutral tool. It is the best defender assistant ever created, and the most efficient attacker’s accomplice ever created — and it is the same model in both cases. This is the dual-use problem in its purest form: the identical query that helps a security engineer harden critical infrastructure could help an adversary breach it. Once Anthropic accepted that framing, the conventional playbook of “announce, benchmark, ship” became untenable. You cannot responsibly hand a weapons-grade capability to anyone with a credit card.
Decoupling capability from distribution
The first and most consequential idea behind Mythos, then, was to treat building the capability and distributing it as two separate decisions. Most of the industry collapses these into one — if you can build it, you ship it. Anthropic pried them apart. It kept the capability internal and in a vetted partner program and held the public release hostage to a single condition: the existence of safeguards strong enough to reliably prevent misuse.
This decoupling is the conceptual key to the whole story. It bought Anthropic optionality. It lets the company demonstrate value, gather evidence, build trust, and develop safeguards — all before the dangerous capability ever reached the open market. For a consultancy, this is also the single most transferable idea in the entire saga, and I’ll return to it.
Safety as a design constraint, not a disclaimer
The second idea was that safety had to be engineered into the product surface, not appended as a usage policy. Anthropic’s approach treats misuse prevention as an architectural layer — a set of classifiers that sit in front of the model and intercept dangerous requests — rather than as a legal disclaimer or a fine-tuning afterthought. The philosophy is that if a capability can cause serious harm, the constraint must be a mechanism, not a promise.
Responsible disclosure, applied to a model
The third strand of thinking borrows from security culture itself. In cybersecurity, when you discover a serious vulnerability, you don’t publish it to the world immediately; you disclose it responsibly to the parties who can fix it, give them a head start, and only then go public. Anthropic effectively applied that ethic to an entire model. Mythos Preview was the vulnerability disclosure; Project Glasswing was the coordinated head start for defenders; the public Fable release was the eventual broader disclosure, made only once the ecosystem had time to prepare. Viewed this way, the program is internally coherent rather than merely cautious.
A warning wrapped in a product
The final piece of thinking is that Mythos was meant to function as a signal to the whole industry. Anthropic argued openly that cheap, fast models with formidable cyber capabilities are imminent — that within roughly six to twelve months, competing labs would likely field their own Mythos-class systems, possibly without safeguards. The implication was pointed out: this capability is coming regardless, so the responsible move is to demonstrate what good stewardship looks like before the reckless version arrives. Claude Mythos was a product and a warning flare at the same time.
Part 2 — The roadmap to today
The decisions above played out over a compressed and remarkably eventful two months. Tracing the chronology makes the strategy legible.
April 2026 — Concept and controlled launch.
Anthropic introduced Claude Mythos Preview as a general-purpose frontier model occupying a new tier above its Opus class and simultaneously launched Project Glasswing — a collaborative initiative to secure the world’s most critical software. Rather than a public release, access was gated to roughly 50 vetted partners, a roster that read like an industry summit: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, among others. Each had to meet strict security requirements. The model stayed locked behind that gate.
Late April through May — Proof and parallel commercialization.
Glasswing partners put Mythos Preview to work, and the results became the program’s evidence base: more than 10,000 high- and critical-severity vulnerabilities identified across real codebases, with partners moving beyond detection into automated patching, pre-release security checks, penetration-testing simulations, and the rebuilding of legacy systems in memory-safe languages. Critically, Anthropic didn’t abandon the broader security market while Mythos stayed restricted — it shipped Claude Security, a product built on the publicly available Opus 4.8, which reportedly helped patch thousands of vulnerabilities within weeks. The pattern here matters: serve the mass market with a safe model, reserve the dangerous one for vetted defenders.
Early June (around June 2) — Scaling the coalition.
Anthropic expanded Glasswing to roughly 150 additional organizations across more than 15 countries, bringing the total to around 200 partners spanning energy, healthcare, water systems, telecommunications, and hardware manufacturing. The expansion followed weeks of consultation with the security industry, open-source maintainers, and the US government — and it landed alongside news that Anthropic had filed confidentially for an IPO, on the heels of a funding round near a trillion-dollar valuation.
June 4 — The recursive self-improvement warning.
Just before the public launch, Anthropic’s institute published “When AI Builds Itself,” co-authored by Jack Clark and Marina Favaro. The report disclosed that more than 80% of code merged into Anthropic’s own codebase was by then being written by Claude rather than human engineers, with per-engineer output up roughly eightfold since 2024. It argued that the industry is approaching “full recursive self-improvement” — AI systems capable of designing and building their own successors with diminishing human oversight — and called for a coordinated mechanism, a “brake pedal,” that would let frontier developers slow or pause in a verifiable, multilateral way. The framing drew explicit parallels to Cold War arms-control thinking. This is the context against which the Mythos commercialization should be read.
June 9 — The dual launch.
Anthropic released two models at once. Claude Fable 5 — a Mythos-class model made safe for general use — went public. Claude Mythos 5 — the same underlying model with certain safeguards lifted — remained restricted to vetted partners. The names encode the architecture: “Fable” (Latin fabula, “that which is told”) and “mythos” (its Greek cousin) are the same story; the safeguards are the only difference. Fable launched as state-of-the-art on nearly all tested benchmarks, with its advantage widening on longer, more complex, autonomous tasks.
Summer 2026 — A staged, capacity-managed rollout.
Fable 5 is fully available through the Claude API, on consumption-based enterprise plans, and via major clouds including AWS, priced at $10 per million input tokens and $50 per million output tokens with a 1-million-token context window. For subscription plans, the rollout is deliberately staged: included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22, shifting to usage credits on June 23, with the stated intent to restore it as a standard feature once capacity allows. Mythos 5 is upgrading existing Glasswing partners, and Anthropic has signaled trusted-access programs forming both cybersecurity and biology.
The throughline across all of it: capability proven in the dark, released into the light only with brakes attached, against a backdrop of the company simultaneously warning that the brakes for the whole industry don’t yet exist.
Part 3 — What the future looks like for the offering
Reading Anthropic’s stated intentions and trajectory, several forward dynamics are reasonably clear.
Safeguards will narrow, and access will widen.
Anthropic has been explicit that its current classifiers are tuned conservatively — catching some benign requests, triggering under 5% of sessions — and that it intends to reduce those false positives over time. Expect the practical envelope of what Fable will answer to expand steadily. In parallel, trusted-access programs will formalize: a systematic application path for cybersecurity organizations to reach Mythos-class capability with cyber safeguards lifted, and a separate track for vetted life-science researchers to access the model with biology and chemistry safeguards removed. The pattern is a widening series of concentric rings of access, each gated by vetting rather than payment alone.
Model cadence will accelerate, and capability will compound.
The recursive self-improvement disclosure is essentially a statement that Anthropic’s own development loop is speeding up because the models are now doing most of the engineering. That implies shorter intervals between capability jumps and a Mythos line that grows more capable faster than past generations did. For anyone building this, the planning assumption should be that today’s frontier is a floor, not a ceiling, and that the gap between “restricted” and “public” capability will keep being renegotiated.
Competition will commoditize the capability — unevenly.
Anthropic’s own forecast is that rival labs will reach Mythos-class capability within six to twelve months, some without comparable safeguards. The likely future is therefore bifurcated: safety-forward providers offering gated, governed access, and less-restricted alternatives appearing elsewhere. That divergence creates exactly the kind of ambiguity and risk that organizations pay advisors to navigate.
Governance and regulation will harden around it.
The data-retention requirement Anthropic introduced for Mythos-class traffic — mandatory 30-day retention across first- and third-party surfaces, with new access-logging and deletion controls — is a preview of where the compliance conversation is heading. Combined with the brake-pedal proposal and active government consultation, the future of this offering is inseparable from an emerging governance regime. Buyers will increasingly need to reconcile capability gains against data-handling, auditability, and dual-use-risk obligations.
In short: more capable, more available, more contested, and more regulated — all at once. That combination is the opportunity.
Part 4 — The consulting perspective
This is where the analysis turns practical. A Mythos-class capability landing into a market that is simultaneously hungry, cautious, and under-governed is close to an ideal setup for a technology consultancy. The opportunity is not to resell tokens — it’s to broker capability, manage risk, and translate raw model power into governed business outcomes. Several distinct service lines follow.
1. Security modernization and legacy remediation at scale.
The single most validated use case is the one Glasswing proved: using Mythos-class capability to find, prioritize, and fix vulnerabilities across large, old codebases, and to rebuild legacy systems in memory-safe languages. Most enterprises cannot access Mythos 5 directly, and Fable 5 deliberately blocks offensive cyber work. A consultancy can sit in the middle — running governed vulnerability-lifecycle programs (detection, automated patch authoring, pre-release checks, remediation verification) using the publicly available models and security products, while helping the most critical clients pursue trusted-access eligibility where appropriate. This is durable, high-value, outcome-based work.
2. Trusted access readiness and brokerage.
Because access to the unrestricted capability is gated by vetting rather than payment, there’s a genuine advisory niche in helping organizations qualify: meeting the security requirements, building the governance and data-handling posture that programs like Glasswing or the forthcoming biology track demand, and managing the application and onboarding. A firm that understands what “trusted” looks like can shepherd clients through a process most don’t know how to approach.
3. AI governance, safety, and dual-use risk advisory.
The 30-day retention requirement, the classifier-fallback behavior, the auditability expectations, and dual-use exposure together create a real compliance and risk surface. CTOs and CISOs need help answering concrete questions: What happens to our data on a Mythos-class model? How do we document and defend our use of a dual-use capability? How do we set an internal policy for where we will and won’t deploy it? This advisory line pairs naturally with existing security and risk practices and is differentiated precisely because the technology is so new.
4. Vertical capability build-outs.
Fable 5’s strengths point to specific industry plays. In financial services, its document-heavy reasoning, chart and table interpretation, and analytical judgment support governed research, due-diligence, and analysis workflows. In life sciences, the forthcoming biology trusted-access track and the model’s demonstrated hypothesis-generation and protein-design capabilities open accelerated-research engagements for vetted clients. In legal and professional services, its long-horizon document reasoning supports review and drafting augmentation. Each vertical is a packaged offering rather than a generic “AI transformation” pitch.
5. Architecture and integration patterns.
The graceful-degradation design — route high-risk requests to a safer fallback rather than refusing outright — is itself a reusable architectural pattern worth productizing as a consulting deliverable. So is the broader work of integrating long-horizon, autonomous agents into existing engineering and knowledge-work pipelines, with the human role shifting from doing to reviewing and directing. Helping clients redesign workflows, controls, and team structures around that shift is substantial change-management work.
6. Economics and build-vs-buy advisory.
The staged, credit-based subscription rollout and the premium token pricing make cost modeling non-trivial. Clients will need help forecasting consumption, deciding which workloads justify Mythos-class spend versus a cheaper model, and architecting tiered routing so the expensive capability is reserved for tasks that actually need it. This is bread-and-butter advisory that compounds as usage scales.
The unifying theme: the model itself is increasingly a commodity input. The scarce, defensible value is in governed translation — turning a powerful, double-edged, lightly-regulated capability into safe, auditable, measurable business outcomes for clients who can’t safely do it themselves.
Part 5 — How a firm should position now
A few concrete moves follow all of this.
Move early on the proven use case.
Security modernization and legacy remediation is the lowest-risk, highest-evidence entry point. It’s where the capability is most validated, where client pain is most acute, and where outcome-based pricing is most defensible. Build the reference engagement here first.
Invest in the governance muscle deliberately.
The differentiator over the next year won’t be access to the model — that will commoditize — it will be the ability to deploy it responsibly and defensibly. A firm that builds genuine depth in AI governance, dual-use risk, and the emerging compliance regime will be positioned for the regulatory hardening that is clearly coming.
Borrow the central insight: decouple capability from deployment.
The most transferable lesson from how Anthropic handled Mythos applies directly to client work. When advising on any powerful or sensitive capability, separate “can we do this” from “should we deploy this broadly, and under what controls.” Make controlled pilots, evidence-gathering, and staged rollout the default. That posture is exactly what nervous boards and risk committees want to hear, and it mirrors the stewardship the market is rewarding.
Hold the analysis honestly.
Part of being a trusted advisor is naming the tensions rather than just the upside. Anthropic is simultaneously a front-runner and the one calling for brakes; the safety narrative and the pre-IPO commercial narrative are genuinely entangled; the conservative safeguards that make the model defensible are also a real operational friction. Clients are better served by an advisor who surfaces these honestly than by one selling unqualified enthusiasm.
The technology is powerful. The deployment strategy is even more important.
Claude Mythos is best understood not as a model but as a thesis about how to introduce dangerous capability responsibly: build it, prove it quietly among defenders, warn the world, then release a braked version to everyone while keeping the unbraked one under vetting. The roadmap from the controlled launch to the dual release has executed that thesis with unusual discipline, all while the company warns that the industry’s collective brakes don’t yet exist.
For a consultancy, the takeaway is energizing rather than cautionary. A capability this powerful, arriving into a market this uncertain and this lightly governed, is precisely the environment in which expert, trustworthy advice is most valuable. The firms that win will be the ones that stop thinking of these models as products to resell and start treating them as powerful, double-edged inputs that clients need help deploying safely, economically, and defensibly. That is a services opportunity with years of runway — and it is open now.
Ready to turn AI into business value?
The central lesson of Anthropic’s Mythos strategy is simple: separate what AI can do from how it should be deployed.
That’s the same philosophy we bring to our clients.
We help organizations evaluate emerging AI, establish governance and risk controls, modernize software and security operations, and integrate advanced models into business workflows in ways that are secure, compliant, and measurable.
If your organization is exploring the next generation of AI, partner with a team that can help you deploy it responsibly—and realize its full business value. Reach out to us to discover how.
