LLM leaderboard
This LLM leaderboard compares 20 AI models on four things you can check: the official price per 1 million tokens, what 1,000 real tasks cost us, how fast each model answers through its own API, and its score on NS-ID, our Bahasa Indonesia benchmark. We run every test ourselves; no score is bought.
Short answer
On our October 25, 2026 run, the top NS-ID tier (models within 3 points of the best score of 99.3 out of 100) was Kimi K3.1 Pro, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, Grok 4.7, Claude Sonnet 5.5, Qwen3.8 Max (0902), GPT-6 Luna, DeepSeek V4.1 Flash, Perplexity Sonar, DeepSeek V4 Pro and ChatGPT (chat-latest). The lowest cost per 1,000 tasks was Llama 4 Maverick at $0.0644. The fastest direct APIs were Claude Haiku 5.5, DeepSeek V4.1 Flash and Claude Sonnet 5.5 (about 235 tokens per second, measured from our app server; gaps under 5% count as a tie). 16 of 20 models scored 90 or more, so a gap of a few points is not a real ranking difference.
NS-ID
Quality here means our Bahasa Indonesia suite (NS-ID v1). We have not built an English suite yet, so the table shows no English score instead of borrowing one from somewhere else.
Scroll the table sideways to see every column.
Models within 3 points of the top score of their tier share a tier, because gaps that small are within normal run to run variation. Inside a tier, rows are sorted by score. Scores are dated test results of the API route named, not a judgement of the company.
| Tier | Model | NS-ID score | Input / 1M | Output / 1M | Price source | Cost / 1,000 tasks | Tokens/sec | First token | Access path | Tested |
|---|---|---|---|---|---|---|---|---|---|---|
| 4 | Mistral Small 4Mistral AI, mistralai/mistral-small-2603 | 84.5 | $0.15 | $0.60 | Sourcechecked October 25, 2026 | $0.0756 | OpenRouter intermediary | October 25, 2026 | ||
| 3 | Command A+Cohere, command-a-plus-05-2026 | 87.8 | Sourcechecked October 25, 2026 | $4.09 | 210.5 | 5,709 ms | Cohere API direct | October 25, 2026 | ||
| 3 | Llama 4 MaverickMeta, meta-llama/llama-4-maverick | 88.2 | $0.19 | $0.65 | OpenRouter pricechecked October 25, 2026 | $0.0644 | OpenRouter intermediary | October 25, 2026 | ||
| 3 | Mistral Medium 3.5Mistral AI, mistralai/mistral-medium-3-5 | 89.6 | $1.50 | $7.50 | Sourcechecked October 25, 2026 | $0.92 | OpenRouter intermediary | October 25, 2026 | ||
| 3 | Gemini 3.5 Flash-LiteGoogle, google/gemini-3.5-flash-lite | 90.7 | $0.30 | $2.50 | Sourcechecked October 25, 2026 | $0.26 | OpenRouter intermediary | October 25, 2026 | ||
| 2 | Gemini 3.8 FlashGoogle, google/gemini-3.8-flash | 93.0 | $0.75 | $3.75 | Sourcechecked October 25, 2026 | $3.09 | OpenRouter intermediary | October 25, 2026 | ||
| 2 | Claude Haiku 5.5Anthropic, claude-haiku-5-5 | 93.2 | $0.10 | $0.50 | Sourcechecked October 25, 2026 | $0.23 | 234.7 | 681 ms | Anthropic API direct | October 25, 2026 |
| 2 | Gemini 4 ArgonGoogle, google/gemini-4-argon | 94.6 | $2.00 | $10.00 | Sourcechecked October 25, 2026 | $1.64 | OpenRouter intermediary | October 25, 2026 | ||
| 1 | ChatGPT (chat-latest)OpenAI, chat-latest | 96.4 | $5.00 | $30.00 | Sourcechecked October 25, 2026 | $3.39 | 139.6 | 966 ms | OpenAI API direct | October 25, 2026 |
| 1 | DeepSeek V4 ProDeepSeek, deepseek-v4-pro / a307abda487cd1b463329ccb945ce396 | 96.8 | $1.32 | $3.96 | Sourcechecked October 25, 2026 | $3.68 | 121.0 | 12,919 ms | DeepSeek API direct | October 25, 2026 |
| 1 | Perplexity SonarPerplexity, sonar | 97.1 | $0.25 | $2.50 | Sourcechecked October 25, 2026 | $3.38 | 197.6 | 1,199 ms | Perplexity API direct | October 25, 2026 |
| 1 | DeepSeek V4.1 FlashDeepSeek, deepseek-flash / e72d3c7f663d3e1383b00b3884cb4501 | 97.1 | $0.30 | $1.20 | Sourcechecked October 25, 2026 | $0.60 | 233.8 | 5,089 ms | DeepSeek API direct | October 25, 2026 |
| 1 | GPT-6 LunaOpenAI, gpt-6-luna | 97.4 | $0.10 | $0.50 | Sourcechecked October 25, 2026 | $0.097 | 135.5 | 7,825 ms | OpenAI API direct | October 25, 2026 |
| 1 | Qwen3.8 Max (0902)Alibaba (Qwen), qwen/qwen3.8-max-0902 | 97.6 | $2.00 | $6.00 | OpenRouter pricechecked October 25, 2026 | $4.00 | OpenRouter intermediary | October 25, 2026 | ||
| 1 | Claude Sonnet 5.5Anthropic, claude-sonnet-5-5 | 97.6 | $2.00 | $10.00 | Sourcechecked October 25, 2026 | $3.77 | 231.9 | 7,750 ms | Anthropic API direct | October 25, 2026 |
| 1 | Grok 4.7xAI (SpaceXAI), x-ai/grok-4.7 | 98.0 | $2.00 | $6.00 | OpenRouter pricechecked October 25, 2026 | $3.89 | OpenRouter intermediary | October 25, 2026 | ||
| 1 | GPT-6 AstraOpenAI, gpt-6-astra | 98.0 | $10.00 | $50.00 | Sourcechecked October 25, 2026 | $6.61 | 49.2 | 2,925 ms | OpenAI API direct | October 25, 2026 |
| 1 | GPT-6.1 SolOpenAI, gpt-6.1-sol | 98.5 | $2.00 | $10.00 | Sourcechecked October 25, 2026 | $1.33 | 110.6 | 4,916 ms | OpenAI API direct | October 25, 2026 |
| 1 | Claude Opus 5.5Anthropic, claude-opus-5-5 | 99.0 | $4.00 | $20.00 | Sourcechecked October 25, 2026 | $9.68 | 158.5 | 9,218 ms | Anthropic API direct | October 25, 2026 |
| 1 | Kimi K3.1 ProMoonshot AI, moonshotai/kimi-k3.1-pro / fpv0_24af6fd9 | 99.3 | $2.00 | $10.00 | OpenRouter pricechecked October 25, 2026 | $9.28 | OpenRouter intermediary | October 25, 2026 | ||
Full NS-ID resultsScore change log
Speed ranking (direct APIs only)
We send the same Indonesian prompt 10 times to each model through the vendor's own API and keep the median. Models we can only reach through OpenRouter are left out of this ranking, because OpenRouter adds its own routing hop and the time would not be the vendor's.
| # | Model | Tokens/sec (median) | Time to first token (median) | API | Measured |
|---|---|---|---|---|---|
| 1 | Claude Haiku 5.5 | 234.7 | 681 ms | Anthropic API | October 26, 2026 (10x) |
| 1 | DeepSeek V4.1 Flash | 233.8 | 5,089 ms | DeepSeek API | October 26, 2026 (10x) |
| 1 | Claude Sonnet 5.5 | 231.9 | 7,750 ms | Anthropic API | October 26, 2026 (10x) |
| 4 | Command A+ | 210.5 | 5,709 ms | Cohere API | October 26, 2026 (10x) |
| 5 | Perplexity Sonar | 197.6 | 1,199 ms | Perplexity API | October 26, 2026 (10x) |
| 6 | Claude Opus 5.5 | 158.5 | 9,218 ms | Anthropic API | October 26, 2026 (10x) |
| 7 | ChatGPT (chat-latest) | 139.6 | 966 ms | OpenAI API | October 26, 2026 (10x) |
| 7 | GPT-6 Luna | 135.5 | 7,825 ms | OpenAI API | October 26, 2026 (10x) |
| 9 | DeepSeek V4 Pro | 121.0 | 12,919 ms | DeepSeek API | October 26, 2026 (10x) |
| 10 | GPT-6.1 Sol | 110.6 | 4,916 ms | OpenAI API | October 26, 2026 (10x) |
| 11 | GPT-6 Astra | 49.2 | 2,925 ms | OpenAI API | October 26, 2026 (10x) |
Not in the speed ranking
- Gemini 4 Argon: Our direct Gemini API key hit its monthly spending cap, so this model ran through OpenRouter and is not timed.
- Gemini 3.8 Flash: Our direct Gemini API key hit its monthly spending cap, so this model ran through OpenRouter and is not timed.
- Gemini 3.5 Flash-Lite: Our direct Gemini API key hit its monthly spending cap, so this model ran through OpenRouter and is not timed.
- Llama 4 Maverick: We can only reach this model through OpenRouter (no direct vendor key), so its time would include OpenRouter's routing hop.
- Grok 4.7: We can only reach this model through OpenRouter (no direct vendor key), so its time would include OpenRouter's routing hop.
- Kimi K3.1 Pro: We can only reach this model through OpenRouter (no direct vendor key), so its time would include OpenRouter's routing hop.
- Qwen3.8 Max (0902): We can only reach this model through OpenRouter (no direct vendor key), so its time would include OpenRouter's routing hop.
- Mistral Medium 3.5: Our direct Mistral API key has no request quota, so this model ran through OpenRouter and is not timed.
- Mistral Small 4: Our direct Mistral API key has no request quota, so this model ran through OpenRouter and is not timed.
How to read this leaderboard
- Price per 1M tokens is copied from the vendor's official pricing page, or from OpenRouter's published price list for models we can only reach through OpenRouter (marked "OpenRouter price"); the link and the date we checked it sit next to every price. Prices older than 90 days are marked "needs re-check" and leave the price sort.
- Cost per 1,000 tasks comes from real usage: the amount billed for this model's answers in the NS-ID run, divided by the number of answers, times 1,000. Models with long hidden reasoning cost more per task than their token price suggests.
- Speed is measured by us, not copied from vendors: time to the first visible token and visible tokens per second, median of the streamed calls.
- The NS-ID score is the mean of six category scores on a 0 to 100 scale. Writing tasks are scored by AI judges from other vendors and carry the label model-graded.
Questions about this leaderboard
Which is the best LLM right now?
For Bahasa Indonesia on our October 2026 run, the 12 models in the top tier were within 3 points of each other, which is inside normal run-to-run variation. Pick by price and speed among the top group rather than by the first row alone.
Why is there no English score?
We only publish scores from tests we run ourselves. Our first suite is in Bahasa Indonesia; an English suite is planned. Until then we show no English score rather than copy one from another site.
Why are some models missing from the speed ranking?
We could only reach them through OpenRouter, a router that adds its own hop between us and the model provider. Their speed would measure OpenRouter as much as the model, so we leave them out and say why.
Can a vendor pay to rank higher?
No. The scoring code has no input for payment, sponsorship or affiliate deals, and an automated test checks that a sponsor flag does not change any score. Every score change is listed in the public score change log.