If Your India Team Cannot Call Claude Fable 5: Export Controls and How to Architect Around Them

FoundrySoft is an India studio that builds for US companies. I am writing this as someone who has to staff projects under the rule, not as an export lawyer. Get counsel. Then use this as the engineering conversation that counsel is rarely in.

In June 2026 the US Department of Commerce issued an export-control directive that, as reported, bars foreign nationals (inside or outside the United States) from accessing some of Anthropic's most capable models, named in coverage as Claude Fable 5 and Claude Mythos 5, including Anthropic's own foreign-national staff. The politics around it are loud. The engineering consequence is quiet and immediate: a lot of "we standardized on Claude" diagrams are now illegal to operate as drawn.

Fable 5.1's cache-price cut does not help if your developers are not allowed to call it.

What "access" usually means in practice

Not just "can they open console.anthropic.com." Access includes:

• API keys in CI, staging, and laptop env files
• Prompt Playground and Workbench
• Log streams and traces that include prompts and completions
• Fine-tune and eval jobs
• Vendor-side debugging where a foreign-national engineer pastes a failing prompt into the restricted SKU

If an India-based engineer can make the model run, they have access. Forwarding a US colleague's laptop over a screenshare is still access. So is a shared Notion doc of "example Fable 5 outputs."

What breaks on a typical project

1. Local inner loop. Engineers in Chennai iterate against Fable 5, ship, and a US-only production key is supposed to make it fine. It does not. The iteration was the access.
2. Evals in CI. Nightly jobs on an India-hosted runner calling the restricted SKU.
3. On-call. The person who got paged is in IST and the runbook says "replay on Fable."
4. Vendor lock as process. Playbooks, snippets, and "the Claude way" of tool calling, all built around a model half the team cannot touch.

Patterns that still ship

Pattern A: restricted SKU is a production-only, US-person path. India builds and evals on Opus 5 (if permitted for your facts), GPT-5.6 Sol, Gemini 3.1 / 3.8, or open weights. A small US-person set owns the Fable/Mythos route: production config, break-glass, and a shadow eval that does not leak prompts back into India-readable logs. This is operationally annoying and legally cleaner.

Pattern B: do not use the restricted SKUs at all. For many products Opus 5 plus Sol plus a cheap long-context model is enough. Fable 5.1's cache win is real (cache pricing post) but it is not worth a compliance incident. We have shipped entire agents without Fable.

Pattern C: self-host an open model for the India loop. DeepSeek V4.1-Flash and other MIT/Apache weights can live in an India VPC. Production in the US can still call a US-permitted hosted SKU. Two models, one tool contract. See self-host vs API.

Anti-pattern: one global key, "we'll be careful." Keys leak. Laptops leave. Contractors join. Careful is not a control.

What to put in the SOW

• Named model SKUs allowed in development versus production
• Where keys, traces, and prompts may be stored (region and nationality)
• Who is on-call for a restricted-SKU incident
• A substitution clause: if a SKU becomes newly restricted, the vendor (us, or you) has N days to reroute without treating it as a change request tax

If you are the US buyer, ask every offshore partner this in the first call. If they have not heard of the June directive, they have not thought about your risk.

Frequently Asked Questions

Does this apply to Claude Opus 5? The reported June directive named Fable 5 and Mythos 5. Confirm current SKU lists with Anthropic and counsel. Do not assume last quarter's allow-list.

Can we use a US VPN from India? That is trying to evade an export control, not "architecture." Do not.

What about Azure or Bedrock endpoints? The restriction is about who may access the model, not which cloud logo is on the invoice. Same question: who is the human (or the CI identity) making the call.

How do we keep evals honest if India cannot see Fable outputs? Run a US-side eval job, store aggregate metrics (pass rate, not prompts) in a place India can read. Or drop Fable from the matrix. Aggregate-only metrics are how you avoid rebuilding access in a dashboard.

---

FoundrySoft staffs US-facing AI work from Chennai with model routing that matches export reality, not the keynote. See hire AI agent developers or talk to us.