LLM deployment

Self-hosted (Llama / Mistral) vs Managed API (OpenAI / Anthropic)

Two paths to production LLM. The right answer depends on volume, sensitivity and team capacity.

Quick verdict

Managed API for most enterprises under 100M tokens/day. Self-hosted for extreme volume, regulated air-gap, or genuine model customisation.

What each actually does

Managed APIs (OpenAI, Anthropic, Google) win on accuracy at frontier tasks, ops simplicity and speed to production. You trade unit economics at very high volume.

Self-hosted open-source (Llama 3.3, Mistral, Qwen) wins on unit cost at scale, full data control and deep customisation. You take on GPU ops, model updates and eval.

In practice, most enterprises are still better served by managed APIs. Self-hosting pays off at extreme volume or when regulation genuinely rules out managed.

Self-hosted (Llama 3.3 / Mistral / Qwen)

Open-source models running on your GPUs — full control, custom fine-tuning, no per-token vendor bill.

Strengths
  • Predictable cost at extreme scale
  • Fully air-gapped deployment possible
  • Deep fine-tune / LoRA freedom
  • No third-party data exposure
Weaknesses
  • Accuracy trails frontier on hardest tasks
  • GPU ops and MLOps burden
  • You maintain the model + fixes
  • Slower to adopt new capabilities

Managed API (OpenAI / Anthropic / Google)

Frontier models exposed as APIs with enterprise tenancy options.

Strengths
  • Best-in-class accuracy on hard tasks
  • Zero infra ops
  • New models arrive automatically
  • Broad ecosystem and tooling
Weaknesses
  • Per-token cost at scale
  • Some data always leaves your infra (even with zero-retention)
  • Vendor lock-in on prompts / features
  • Regional availability varies

Side-by-side

CriterionSelf-hosted (Llama / Mistral)Managed API (OpenAI / Anthropic)
Accuracy (hard tasks)Very strong (behind frontier)Best-in-class
Cost at 10B tokens/moLow (amortised GPUs)Very high
Ops burdenHigh (GPU + MLOps)Near zero
Data controlFull (air-gap possible)Zero-retention available but data leaves
CustomisationFull fine-tune / LoRALimited (prompt, RAG, some fine-tune)
Time to productionWeeks–monthsDays

Choose Self-hosted (Llama / Mistral) when

  • Genuinely air-gapped / sovereign deployment
  • Extreme volume (>1B tokens/day)
  • Deep fine-tuning is the moat
  • GPU infra and MLOps team already exist

Choose Managed API (OpenAI / Anthropic) when

  • Standard enterprise volume (<100M tokens/day)
  • Frontier accuracy required
  • No dedicated MLOps team
  • Time-to-market matters most
The pragmatic answer

Use both, on purpose

The dominant pattern in 2026: managed APIs for reasoning-heavy queries, self-hosted small models for high-volume classification, extraction and routing. Model routing gets 70–90% of self-hosting's cost win at 5% of the ops burden.

Frequently asked questions

Can I fine-tune managed models?

Yes — OpenAI, Anthropic and Google all offer fine-tuning. It's more limited than self-hosted LoRA but often good enough.

How much does self-hosting really cost?

GPU rent for a Llama 3.3 70B endpoint at moderate load: $5–15k/month. Add MLOps, retraining and updates — real all-in is often 2–3x list.

What about hybrid — self-host small + managed big?

That's the recommended default in 2026. Route by query type. Best economics with minimum ops.

Ready to Build Your AI Product?

Talk to a senior AI consultant from T7 about your industry, workflow, or product idea. Free, no commitment — reply within one business day.

  • · AI feasibility & architecture review
  • · Product / MVP roadmap
  • · Integration & automation strategy