Estimated reading time: 4 minutes
Key takeaways:
- Never build your stack around a single model. AI models can be deprecated, banned, or pivoted overnight. Treat them as accelerators, not foundations.
- Make your infrastructure model-agnostic: tools, prompts, and routing should sit outside the model so you can swap without rebuilding.
- Open-source models are no longer a compromise.
Two of Anthropic’s most powerful models to date had a shelf life of under a week. Claude Mythos 5 and Fable 5 launched on June 9, and were taken offline three days later.
What happened? The US government deemed the models a national security risk, and the Commerce Department barred foreign access. Anthropic, unable to reliably screen foreign-national status across its user base in real time, decided to shut down access entirely.
Such drastic product shifts used to be rare in enterprise tech. Now, they feel commonplace in today’s turbulent AI market.
“The AI market’s velocity is also its vulnerability,” Robert Keus, founder of GreenPT, a GPT-powered chat platform run on renewable energy, tells LeadDev. “Model providers can pivot, deprecate, or disappear overnight.”
Beyond sovereign bans, we’re seeing rapid model releases, pricing changes, and vendor experimentation. Today’s model might be made obsolete by the next version, or unavailable tomorrow.
Engineers can lose time wiring into new models, testing them, and adjusting workflows, only for the models to disappear. So, how can teams cope with uncertainty without rebuilding core development processes?
The experts say hedging your bets means reducing dependency on one model or provider, using open source models, and adopting abstraction layers for optionality.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
Avoid single-provider dependencies
“It is never a good strategy to build your AI stack around one specific model or one specific provider,” Sara Landi Tortoli, founder and AI product strategist at Polyweb, an AI consulting and media company, tells LeadDev.
For her, AI models are becoming a commodity and changing constantly. To reduce single-provider dependency, she recommends building a “model-agnostic harness.” Here, “harness” refers to the peripheral infrastructure supporting model interaction.
“Your tools, memory, data layer, permissions, prompts, routing, and evaluations should sit outside the model and be designed independently from it,” she says. “Then, only at the last step, you plug in the model.”
To her point, a recent Claude Code analysis found that only 1.6% of the codebase is AI decision logic. The remaining 98.4% is operational infrastructure like permissions, context management, tool routing, and recovery logic.
Use agnostic abstraction layers
Others encourage strategies to reduce proprietary AI dependency. “Beware of vendor lock-in, and use abstraction layers where possible,” Manik Surtani, CTO and co-founder of the Agentic AI Foundation (AAIF), tells LeadDev.
Instead of Anthropic’s Claude Code or OpenAI’s Codex, he recommends Goose, an open-source AI agent from AAIF. “Goose gives you flexibility of model, to help you navigate pricing changes, pivots and deprecation, government bans,” he says, noting it can run multiple models concurrently.
Another benefit to using an open-source agent is you get a vendor-neutral harness. “Goose isn’t incentivized to be token-verbose, which you can’t say about some proprietary agent harnesses,” he adds.
More like this
Consider open-source models
Another hedge against commercial vendor shifts is using open-source and open-weight models. This especially makes sense for simpler operations where state-of-the-art frontier models are unnecessary.
“Companies need digital autonomy – and that starts with open-source models,” says Keus. “Open-source models give you the portability and control to build a truly model-agnostic foundation.”
For him, the benefits are the ability to self-host, switch, audit, and fork models without rebuilding your infrastructure.
“At GreenPT, we run everything on open source,” adds Keus. “Even the newest coding models like GLM5.2 and Kimi2.7-code, are really good. We see that teams are switching more often.”
That said, open-source AI has its own caveats. To guide model selection, evaluation is becoming a necessary engineering skill.
“I believe that the real skill for engineering leaders now is knowing how to evaluate which model is best for a specific use case, at a specific quality bar, and at a specific price point,” says Tortoli.
Industry-standard evaluations and leaderboards have emerged for comparing AI models, which could aid decision-making.
Yet, industry benchmarking only goes so far, says Tortoli. “Industry benchmarks are useful for orientation, but companies really need their own task-specific evals.” She recommends a manual judge paired with a large language model- (LLM-) as-a-judge to scale the process.
Berlin • November 9 & 10, 2026
AI. Burnout. Big decisions.
The pressure is real. Berlin is where you work through it.
Expect more uncertainty
The software world should prepare for more shifts. As we’ve just witnessed, digital sovereignty adds to AI model uncertainty.
As Melissa Rosenthal shares on LinkedIn, national security mandates trump any pre-conceived notion of stability or enterprise assurances. “A national security export control sits above your SLA, your committed spend, and your roadmap.”
With this in mind, engineering leaders should prepare a vendor-agnostic foundation to keep using third-party AI vendors.
Keus advises engineering leaders to treat closed models as accelerators, not foundations. “Build on what you can own and replace,” he says. “That’s how you survive the next ban, pricing shift, or deprecation.”
