Which long-context LLM supports 128k+ context without severe context truncation on dense prompts?

A few long-context LLMs can handle 128k+ tokens, but if you care specifically about dense prompts and avoiding severe “effective truncation” (where the model technically accepts the window but performs poorly on later tokens), the best options are usually:

• Claude 3.5 Sonnet / Claude 3 Opus — strong long-context performance, often among the best for dense documents.
• GPT-4.1 / GPT-4o variants with long context — good at large windows, though exact effective behavior depends on the deployment.
• Gemini 1.5 Pro — very large context and generally strong retrieval over long inputs.

If you want the most reliable answer for 128k+ with dense prompts, I’d usually point to:

Claude 3.5 Sonnet for quality
Gemini 1.5 Pro for very large raw context
GPT-4.1 if you want OpenAI ecosystem support

If you mean open-source / self-hosted models, then options exist like:

• Llama 3.1 70B/405B with 128k context
• Qwen2.5 long-context variants
• Mistral Large / Mixtral long-context deployments

…but open-source long-context models are more likely to degrade on dense inputs unless carefully tuned and served with the right attention/KV setup.

If you want, I can give you a ranked shortlist by actual long-context reliability, or recommend one for API use vs self-hosted.