Question 1

Which LLM is the best in October 2026?

Accepted Answer

Right now Claude Fable 5, Gemini 3.1 Pro and Claude Opus 4.8 lead (graded on Arena Elo (Text), GPQA Diamond and MMLU-Pro plus an overall task grade). Best open-weight: DeepSeek V4-Pro. Best budget pick graded A or better: Qwen3.7 Flash ($0.03/$0.13 per 1M tokens).

Question 2

What's the best free open-source LLM I can self-host?

Accepted Answer

Depends on your VRAM. With 24GB VRAM (RTX 4090): Gemma 4 26B (MoE), Mistral Small 3.2 24B, or Qwen3 32B at Q4. With 48GB+: Llama 4 Scout (109B MoE, 17B active) or Qwen3 235B-A22B at Q4. On a Mac with 64GB unified memory: DeepSeek R1 671B is feasible at Q3. Use the filters on this leaderboard to see self-hostable models only.

Question 3

How does the value score on the leaderboard work?

Accepted Answer

value_score is a 0–100 composite computed at build time. It takes each model's average of its top 3 benchmark scores (normalised across the priced set), divides by its average $/1M token cost (also normalised), and rescales the result to 0–100 across all priced models. Worst priced model = 0, best = 100, single priced model defaults to 50. The formula is in scripts/build-models.js.

Question 4

Are the benchmark scores on id8's leaderboard audited?

Accepted Answer

Yes, by rule rather than by hand. Every non-null value in data/models.json cites a source URL, and the nightly pipeline accepts a price, context window or release date only when two independent sources agree. Benchmarks come from uniform third-party harnesses (LMArena, Epoch AI) instead of vendor blogs, citations to unsourced aggregator sites are rejected, and a run that would change an unusual share of the data is halted automatically.

Question 5

Which LLM is best for Hindi or Indic-language tasks?

Accepted Answer

For Hindi specifically, Claude Opus 4.7 and Gemini 3.5 lead on quality. Among open-weight models, Qwen 3 32B and Gemma 4 26B handle Devanagari script well. Note that Hindi tokens cost 3–4× English tokens in most tokenizers — factor that into your API budget. See the 'Best LLM for Hindi' filter view.

Question 6

How often is the leaderboard updated?

Accepted Answer

Daily and automatically. A nightly job pulls prices, context windows and release dates from models.dev, OpenRouter and LiteLLM, Arena Elo from LMArena, and benchmark scores from Epoch AI. A value is published only when at least two independent sources agree; new models are added when two sources list them, and models a vendor withdraws are retired. The date above the rankings is the last time any value changed.

Question 7

Can I download the leaderboard data as JSON or CSV?

Accepted Answer

Yes. Use the Download CSV button at the top of the leaderboard to export all models as a UTF-8 CSV (Excel-ready). The JSON projection is served at https://id8.co.in/data/llm-leaderboard.json (CC-BY-4.0 license). The self-host subset is at https://id8.co.in/data/local-llm-models.json. The canonical source-of-truth is at https://id8.co.in/data/models.json (also CC-BY-4.0, includes per-field source URLs).

LLM Leaderboard — October 2026