Adya
On Friday evening, the US government told Anthropic to switch off its two most powerful models.
Fable 5 and Mythos 5. Gone for every user on the planet within hours. The Commerce Department sent a directive saying foreign nationals should not have access — including foreign nationals working at Anthropic. The company could not reliably separate foreign users from everyone else in real time, so the only way to comply was to disable both models worldwide.
Anthropic disputes it publicly. It calls the issue a narrow potential jailbreak already available from other models, and it is working to get access restored — one official suggested the lockdown could lift in a few weeks. So read this as a live and contested situation, not a permanent ban. It very likely gets reversed once the dust settles. That is exactly why the panic reaction is the wrong reaction.
The Indian AI timeline saw this and reached for the obvious conclusion: we need our own foundation models. We cannot depend on America.
I want to make the opposite case. Friday was not an argument for Indian foundation models. It was an argument against depending on any single model, from anyone, ever. And the way you escape that dependence is not by spending hundreds of millions to build your own frontier model. It is by getting world class at the two layers above the model.
Section 01The wrong lesson, and the right one
The instinct is understandable. Watch a foreign government reach in and switch off the tools your business runs on, and “build our own” feels like the only safe answer. It is the wrong one — it solves for the wrong failure.
The wrong lesson
- A hundreds-of-millions race for two national treasuries
- Press-release launches that quietly lose on usage
- A copy of the very thing that just got switched off
The right lesson
- Win the two layers above the model: harness and fine-tuning
- Run on any model — open or frontier
- Swap the engine underneath in an afternoon
Section 02What actually makes a model business win
I run an agentic AI company. We test models for a living, inside our own harness, against real long-horizon enterprise tasks — Opus 4.6, 4.7 and 4.8, GPT 5.5, Gemini, Qwen. We tested Fable 5 itself before it went dark.
Four things decide who wins, and it is worth being honest about which ones India can actually play in.
Train the frontier model
The best researchers in the world and hundreds of thousands of advanced chips wired together. Not one, not a thousand — hundreds of thousands. That is a capex line only two countries can write today: the US and China. India cannot, and pretending otherwise is how you set fire to public money.
US & China onlyInference and distribution
Serving millions of users needs enormous compute parallelised across huge fleets, and distribution decides the rest. If your model is even one point worse than the best in the world, people quietly switch away. Look at the Indian model attempts that led on a press release and lost on usage. Inference is unforgiving.
Capital + reachArchitecture and harness
Here the story changes. India is genuinely good at building smaller, domain-specific models in the fifty-billion-parameter range — and more of the value now sits in the harness: the orchestration, tooling, memory and control layer wrapped around the model. This is engineering, taste and domain depth. We have all three.
India can winFine-tuning
Adapting open weights to your own data — cheaply, repeatably, on the tasks you actually run. Just as important as the harness, and just as winnable.
India can winSection 03The proof is already on the table
You do not have to take my word that the value has moved up the stack. The two cleanest examples of 2026 are companies beating frontier labs without owning a frontier model.
Fine-tuned a Chinese open model for one job
Cursor took Kimi K2.5, an open-source model from Moonshot AI, and fine-tuned it for agentic coding. Composer 2.5 reaches roughly Opus 4.7 level on Cursor’s own coding benchmark — GPT 5.5 still leads on shell-heavy work, but on the task that matters it matches a frontier model.
A fine-tuned open model beat Opus on legal
Harvey runs on Claude, and Opus 4.7 leads its legal benchmarks when used directly. But with Fireworks AI, a fine-tuned open model beat Opus on a 100-task slice of their Legal Agent Benchmark — and their strongest setup used a frontier model only as a sparse advisor the cheaper model calls when it needs help.
Matching a frontier model on the task you care about, at one-tenth the price, by fine-tuning open weights — that is the entire thesis in one product. Harness plus fine-tuning, proven on a public benchmark, in one of the highest-stakes domains there is.
The winning design does not put the expensive model in the engine. It puts a fine-tuned open model in the engine and a frontier model in the passenger seat.
Section 04A word on Fable 5, since everyone is talking about it
Fable 5 is a genuinely strong long-context, long-horizon model. I am not going to pretend otherwise. But it has very little impact on normal product work or day-to-day operations. Where it shines is the extreme end — very advanced simulations, deep strategy, multi-day agentic runs. And the token economics are brutal: for ordinary enterprise work it can cost more than hiring ten engineers.
Take GPT 5.5, put it inside a good harness, and you get a better result than Fable 5 used raw out of the box. Fable 5 only pulls clearly ahead when you also wrap it in a strong harness and point it at a genuinely long-horizon task. That is a narrow, expensive, extreme use case — not the centre of the market. The centre is won by harness and fine-tuning, not raw frontier horsepower.
Section 05What India should actually do
Stop measuring our ambition by whether we can build a model that competes with Anthropic and OpenAI on raw capability. That is a hundreds-of-millions-of-dollars race for two national treasuries, and we are not in it. Measure it instead by three things that are squarely within our reach.
- Build excellent small and domain-specific models — the range where Indian engineering is already strong.
- Build fine-tuning studios that let any business adapt open weights to its own data, cheaply.
- Build the best harnessed systems in the world, so the model becomes a swappable component rather than a single point of failure.
This is why we built two things into the platform.
Point an agent at your data. Get a fine-tuned model.
Describe what you want, and it fine-tunes a model for you — small, cheap, and owned by you, on your own cloud or hardware. For a lot of enterprises that is millions of dollars a year saved, with no loss of quality on the tasks they actually run.
The harness that keeps you provider-independent
Run on any model and swap the engine underneath in an afternoon, with a control layer that reroutes when a provider goes dark — so if one model is switched off, your apps never feel it.
Section 06The real lesson
Friday proved that renting a model you do not control is fragile. It can be switched off by a vendor, a regulator or a billing event — none of which you decide. The answer is not to spend a fortune building a copy of the thing that just got switched off. The answer is to become so good at the layers above the model that you can run on any model, swap it in an afternoon, and never again be held hostage by one company or one government.
That is not a fantasy. It is engineering — and it is the kind of engineering India is good at. Let us go build it.
Get the next one in your inbox
Essays on enterprise AI, seen from India — the harness, fine-tuning, and what actually wins. No noise, no hype.
Start building on Vanij