LLM Landscape — Closed · Open · Korean-Specialized · Evaluation · Pricing

The LLM market shifts fast. Closed-API and open-weight, English-centric and multilingual, cloud and self-hosted, models specialized for Korean — all sit alongside each other.

1. Closed (API · weights private)

Even within the same provider, model capability shifts quickly across generations · dates.

2. Open weights

Model families whose weights can be downloaded and run for inference. License conditions differ per model.

The degree of "open" varies per model. There's a difference between weights-only public, training code public, and even training data public. Hence the view that "open weights" is more accurate than "open source."

3. Korean-specialized · Korean-company models

Korean ability holds more meaning per-model. The observation is that even the same global model can vary widely in Korean across generations.

4. Reasoning models · multimodal · context length

Reasoning models — From late 2024, the trend of OpenAI o1 · o3, DeepSeek R1, Claude's extended thinking, Gemini 2.5's thinking mode. The model goes through longer internal reasoning before responding, spending more tokens · time accordingly.

Multimodal — Models that take images · audio · video · documents as input alongside text are now standard. GPT-4o · Gemini · Claude 3.x.

Context length expansion:

A larger context isn't always the answer. Position effects like "lost in the middle" come along with cost · latency.

5. Evaluation sites

Limits of evaluation:

• Suspicion of training-data leakage (cases where benchmarks ended up in training data).
• Many evaluations are English-centric.
• A single score doesn't directly tie to your domain performance.

6. Pricing models

Per-token billing (API) — Most closed models price input · output tokens separately. Output tokens are usually more expensive. With the introduction of context caching · prompt caching, cached input gets a discount.

Cost per request ≈ (input tokens × input rate) + (output tokens × output rate)

Subscription model (consumer) — ChatGPT Plus / Team / Enterprise · Claude Pro / Team · Gemini Advanced · Perplexity Pro. Bundles UI · quota · extra features.

Self-hosted — Open weights + your own GPU or cloud GPU. Per-call cost disappears, but GPU time · MLOps people · model updates · evaluation · operational burden grow. For small light workloads, API; for high traffic · strong data control needs, self-hosted. The threshold is per-workload.

Data usage policy — Even within the same provider, free tier · paid API · enterprise policies differ. Check the terms and the model card every time.

7. Cloud cost gate

Having an API key is not the same as approving spend. Keep a local or non-billed path as the default and allow an external provider only when a separate cost gate is open. If a runtime store still contains an old cloud selection, a closed gate must downgrade it to the default path.

Opening a cloud AI feature

1Default

Block cloud providers from both selection and execution.

2Approval

Decide transmitted data, input and output token caps, monthly budget, and stop conditions.

3Enable

Open the application gate and the provider spend limit together.

4Observe

Track requests, tokens, errors, and cost, then close the gate at the limit.

Cost protection layers

Application gate

Prevents a key or stale setting from starting external calls.

Request budget

Caps input length, output tokens, and per-user request frequency.

Provider limit

Caps organization-wide usage with spend limits and alerts.

Provider dashboards can lag behind real usage. Keep the pre-request application gate instead of relying on the dashboard budget alone.

8. The thread of selection

• Fast and cheap, in volume — GPT-4o-mini / Claude Haiku / Gemini Flash / small open models.
• Quality first — GPT-4 / Claude Sonnet · Opus / Gemini Pro / large open models.
• Reinforced reasoning — o1 · o3 / Claude extended thinking / Gemini Thinking / DeepSeek R1.
• On-device · privacy — Small variants of Llama · Gemma · Phi · Qwen + LM Studio · Ollama.
• Heavy Korean — Korean-specialized models · multilingual-strong global models, with your own domain evaluation.

9. Spots where you often get stuck

Aliasing volatility — Aliases like gpt-4 · gemini-1.5-pro-latest point at different models depending on the time. Operations should pin to dates.

Benchmark over-trust — #1 isn't #1 in your domain.

License differences — Open weights doesn't mean all are commercially usable. Check the model card.

Data training use — Free / paid / enterprise can have different policies. Don't let sensitive info into the input.

Generation-change regression — A new model doesn't beat the old in every aspect. Sometimes regression appears in your tasks.

Advertised vs actual context length — Sometimes the advertised limit and the per-model input · output limit differ.

Tokens of reasoning models — Whether thinking tokens are included in the response or billed separately differs per provider.

"AGI" · "superhuman" expressions — Marketing expressions should be filtered out when interpreting evaluation results.

Closing thoughts

The LLM landscape shifts fast, so single-model dependence creates regression risk. Model pinning, a domain eval set, validated provider switching, a default-closed cost gate, and cost monitoring form a safer operational baseline.

Next

• (end of ai)

References: LMArena · LiveBench · Open LLM Leaderboard · OpenAI Models · Anthropic Models · Gemini Models · Meta Llama · Mistral · DeepSeek.