AI models, compared
26 models you can use or build on, scored on public benchmarks. 15 have open weights, and 6 have makers who say they handle Tamil. Every score links to where it was published.
Checked 11 Oct 2026. Models change fast; scores move every week.
This page is drafted by Claude, a model made by Anthropic, which also appears in the table. Anthropic's models are ranked by the same published scores as everyone else's, with no adjustment, and labelled.
People compare two anonymous answers and pick the better one. Higher is preferred more often.
-
Anthropic's highest-priced model for the hardest, longest coding and research projects. Anthropic says Opus 5.5 performs at its level on most work for less money, so most users will not need it.
- Released
- 1 Sep 2026
- Inputs
- text, image
- Context
- 1 million tokens
- Price
- USD 10 / 50 per million input/output tokens
- Use it via
- Claude Pro, Max, Team and Enterprise plans; Claude API, AWS, Google Cloud, Microsoft Foundry
- Tamil
- Not statedAnthropic's multilingual page benchmarks 14 languages including Hindi and Bengali; Tamil is not listed.
Published scores Benchmark Score Source LMArena 1501 LMArena text leaderboard, 8 Oct 2026Listed as claude-fable-5.1-max AA Intelligence Index 53 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Claude Fable 5.1 (max with fallback) Humanity's Last Exam 59.1% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort with fallback GPQA Diamond 93.7% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort with fallback -
Anthropic's smallest and cheapest current model. Built for high-volume jobs such as summaries, classification, database queries and live customer support, and as a helper agent next to larger models.
- Released
- 7 Oct 2026
- Inputs
- text, image
- Context
- 1 million tokens
- Price
- USD 0.10 / 0.50 per million input/output tokens for prompts up to 100k tokens; USD 0.50 / 2.50 above that
- Use it via
- Claude API, AWS, Google Cloud, Microsoft Azure
- Tamil
- Not statedAnthropic's multilingual page benchmarks 14 languages including Hindi and Bengali; Tamil is not listed.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index 43 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Claude Haiku 5.5 (max) Humanity's Last Exam 44.4% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond Not published -
Anthropic's main flagship model. Strong at long coding jobs, agent tasks and office work such as reports and spreadsheets, and cheaper to run than Opus 5. Suits developers and teams with complex work.
- Released
- 22 Sep 2026
- Inputs
- text, image
- Context
- 1 million tokens
- Price
- USD 4 / 20 per million input/output tokens
- Use it via
- Claude apps, Claude API, AWS, Google Cloud, Microsoft Azure
- Tamil
- Not statedAnthropic's multilingual page benchmarks 14 languages including Hindi and Bengali; Tamil is not listed.
Published scores Benchmark Score Source LMArena 1507 LMArena text leaderboard, 8 Oct 2026Listed as claude-opus-5.5-high AA Intelligence Index 58 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Claude Opus 5.5 (max with fallback) Humanity's Last Exam 61.4% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort with fallback GPQA Diamond Not published -
A faster, lower-cost Claude model that comes close to Opus 5.5 on many tests. A sensible default for everyday coding, bug fixing, documents and slides.
- Released
- 28 Sep 2026
- Inputs
- text, image
- Context
- 1 million tokens
- Price
- USD 2 / 10 per million input/output tokens
- Use it via
- Claude apps, Claude API, AWS, Google Cloud, Microsoft Azure
- Tamil
- Not statedAnthropic's multilingual page benchmarks 14 languages including Hindi and Bengali; Tamil is not listed.
Published scores Benchmark Score Source LMArena 1476 LMArena text leaderboard, 8 Oct 2026Listed as claude-sonnet-5.5-xhigh AA Intelligence Index 56 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Claude Sonnet 5.5 (max with fallback) Humanity's Last Exam 55% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort with fallback GPQA Diamond Not published -
DeepSeek's current main model, open under the MIT licence, with text and image input and a 1M-token context. Its API prices are among the lowest here. DeepSeek says it replaced V4-Pro.
- Released
- 10 Sep 2026
- Inputs
- text, image
- Context
- 1 million tokens
- Price
- USD 0.30 / 1.20 per million input/output tokens at peak hours; half that off-peak
- Use it via
- DeepSeek API; free download (Hugging Face); in beta on Sarvam AI's API
- Tamil
- Not statedThe model card does not list supported languages.
Published scores Benchmark Score Source LMArena 1475 LMArena text leaderboard, 8 Oct 2026Listed as deepseek-v4.1-flash-max AA Intelligence Index 39 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as DeepSeek V4.1 Flash (max) Humanity's Last Exam 39.2% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond 90.9%Maker's own figure DeepSeek-V4.1-Flash model card, Sep 2026 -
Google's latest stable Gemini model and its most capable Flash model. It reads text, images, video, audio and PDFs, has a free tier and is cheap for its level. A strong pick for students and builders on a budget.
- Released
- Sep 2026
- Inputs
- text, image, video, audio
- Context
- 1.05 million tokens
- Price
- USD 0.75 / 3.75 per million input/output tokens until 31 Dec 2026, then USD 1.50 / 7.50; free tier available
- Use it via
- Gemini API (Google AI Studio, free tier), Google Cloud
- Tamil
- Stated supportGoogle Cloud docs list Tamil (ta) among the languages all Gemini models understand and respond in.
Published scores Benchmark Score Source LMArena 1497 LMArena text leaderboard, 8 Oct 2026Listed as gemini-3.8-flash-high; marked Preliminary AA Intelligence Index 41 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Gemini 3.8 Flash (high) Humanity's Last Exam 47.8% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort GPQA Diamond 95.3% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort -
The largest dense model in Google's open Gemma 4 family, under Apache 2.0. Takes text and images with a 256K context and is sized for workstation GPUs, so it suits local use and fine-tuning.
- Released
- 31 Mar 2026
- Inputs
- text, image
- Context
- 256k tokens
- Price
- Not stated
- Use it via
- Free download (Hugging Face); also in beta on Sarvam AI's API
- Tamil
- Not statedGoogle says Gemma 4 was pre-trained on 140+ languages with out-of-the-box support for 35+; Tamil is not named.
Published scores Benchmark Score Source LMArena 1452 LMArena text leaderboard, 8 Oct 2026Listed as gemma-4-31b AA Intelligence Index 15 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Gemma 4 31B Humanity's Last Exam 23.6% Artificial Analysis, 11 Oct 2026Artificial Analysis run, reasoning mode GPQA Diamond 84.3%Maker's own figure Google Gemma 4 model card, Mar 2026 -
Z.ai's open-weight model focused on coding and long agent tasks, with a 1M-token context. A smaller GLM-5.3-Flash version is released under MIT.
- Released
- Aug 2026
- Inputs
- text
- Context
- 1.05 million tokens
- Price
- Not stated
- Use it via
- Mistral API; free download (Hugging Face); in beta on Sarvam AI's API
- Tamil
- Not statedThe model card lists English and Chinese only.
Published scores Benchmark Score Source LMArena 1478 LMArena text leaderboard, 8 Oct 2026Listed as glm-5.3-max AA Intelligence Index 45 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as GLM-5.3 (max) Humanity's Last Exam 42.3% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond 91.7% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort -
OpenAI's most capable model, aimed at hard reasoning, coding, computer use and research. Powerful but one of the most expensive models per token.
- Released
- 3 Sep 2026
- Inputs
- text, image
- Context
- 1.05 million tokens
- Price
- USD 10 / 50 per million input/output tokens
- Use it via
- ChatGPT Plus, Pro, Business and Enterprise; OpenAI API, Microsoft Azure, AWS Bedrock
- Tamil
- Not statedOpenAI's model page does not list supported languages.
Published scores Benchmark Score Source LMArena 1475 LMArena text leaderboard, 8 Oct 2026Listed as gpt-6-astra-max AA Intelligence Index 53 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as GPT-6 Astra (max) Humanity's Last Exam 54.7% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond 96%Maker's own figure OpenAI: GPT-6 Astra announcement, Sep 2026 -
OpenAI's low-cost model for focused, high-volume tasks. It powers GPT-6 for Free and Go users in ChatGPT, so it is the version many students will use first.
- Released
- 22 Sep 2026
- Inputs
- text, image
- Context
- 1.05 million tokens
- Price
- USD 0.10 / 0.50 per million input/output tokens
- Use it via
- ChatGPT Free and Go tiers; OpenAI API
- Tamil
- Not statedOpenAI's model page does not list supported languages.
Published scores Benchmark Score Source LMArena 1443 LMArena text leaderboard, 8 Oct 2026Listed as gpt-6-luna-max AA Intelligence Index 38 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as GPT-6 Luna (max) Humanity's Last Exam 38.5% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond Not published -
OpenAI's mid-priced model. OpenAI says it comes close to Astra on complex coding, computer use and professional work at a much lower price. A good fit for most paid API projects.
- Released
- 29 Sep 2026
- Inputs
- text, image
- Context
- 1.05 million tokens
- Price
- USD 2 / 10 per million input/output tokens
- Use it via
- ChatGPT Plus, Pro, Business, Enterprise and Edu (Work and Codex); OpenAI API
- Tamil
- Not statedOpenAI's model page does not list supported languages.
Published scores Benchmark Score Source LMArena 1484 LMArena text leaderboard, 8 Oct 2026Listed as gpt-6.1-sol-max AA Intelligence Index 52 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as GPT-6.1 Sol (max) Humanity's Last Exam 52.9% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond Not published -
OpenAI's open-weight reasoning model under the Apache 2.0 licence. It fits on a single 80GB GPU. Older and weaker than the frontier models here, but still widely downloaded.
- Released
- Aug 2025
- Inputs
- text
- Context
- Not stated
- Price
- Not stated
- Use it via
- Free download (Hugging Face); many cloud hosts
- Tamil
- Not statedThe model card does not list supported languages.
Published scores Benchmark Score Source LMArena 1352 LMArena text leaderboard, 8 Oct 2026Listed as gpt-oss-120b AA Intelligence Index 12 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as gpt-oss-120b (high) Humanity's Last Exam 19.6% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort GPQA Diamond 78.2% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort -
The flagship Grok model for coding and knowledge work. SpaceXAI says it costs about half as much as comparable frontier models.
- Released
- 21 Sep 2026
- Inputs
- text, image
- Context
- 500k tokens
- Price
- USD 2 / 6 per million input/output tokens (prompts under 200k tokens)
- Use it via
- Grok API, Grok Build, Cursor, cloud platforms
- Tamil
- Not statedThe model page does not list supported languages.
Published scores Benchmark Score Source LMArena 1443 LMArena text leaderboard, 8 Oct 2026Listed as grok-4.7-xhigh AA Intelligence Index 46 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Grok 4.7 (xhigh) Humanity's Last Exam 43.1% Artificial Analysis, 11 Oct 2026Artificial Analysis run, xhigh effort GPQA Diamond Not published -
AI4Bharat's open translation models covering all 22 scheduled Indian languages, with an MIT licence for the model checkpoints. Useful as a free base for English to Indian-language translation tools.
- Released
- May 2023
- Inputs
- text
- Context
- Not stated
- Price
- Not stated
- Use it via
- Free download (GitHub, Hugging Face)
- Tamil
- Stated supportThe project page lists Tamil (tam_Taml) among the 22 scheduled Indian languages it translates.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index Not published Humanity's Last Exam Not published GPQA Diamond Not published -
Moonshot AI's 2.8-trillion-parameter open-weight model with a 1M-token context. Strong at long coding and research tasks. The weights use Moonshot's own Kimi K3 licence.
- Released
- Jul 2026
- Inputs
- text, image, video
- Context
- 1.05 million tokens
- Price
- Not stated
- Use it via
- Kimi app, Kimi API; free download (Hugging Face)
- Tamil
- Not statedThe model card does not list supported languages.
Published scores Benchmark Score Source LMArena 1488 LMArena text leaderboard, 8 Oct 2026Listed as kimi-k3-max AA Intelligence Index 44 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Kimi K3 (max) Humanity's Last Exam 46.9% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond 93.5%Maker's own figure Kimi K3 model card, Jul 2026 -
A 12B model from Ola's Krutrim, built on Mistral NeMo and tuned for Indian languages and Indian cultural context. Weights are under the Krutrim Community License.
- Released
- 4 Feb 2025
- Inputs
- text
- Context
- 128k tokens
- Price
- Not stated
- Use it via
- Krutrim Cloud; free download (Hugging Face)
- Tamil
- Not statedKrutrim says the model is trained for Indic languages; the model card does not name Tamil.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index Not published Humanity's Last Exam Not published GPQA Diamond Not published -
Mistral's largest model (1 trillion parameters), in public preview since 6 Oct 2026. Mistral says the weights will be released by the end of October 2026; until then it is listed here as closed.
- Released
- 6 Oct 2026
- Inputs
- text, image
- Context
- 1 million tokens
- Price
- USD 1.36 / 4.18 per million input/output tokens (preview sale price USD 0.68 / 2.09)
- Use it via
- Mistral API (public preview)
- Tamil
- Not statedMistral says training data spanned more than 160 languages; Tamil is not named.
Published scores Benchmark Score Source LMArena 1429 LMArena text leaderboard, 8 Oct 2026Listed as mistral-large-4 AA Intelligence Index 38 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Mistral Large 4 Preview Humanity's Last Exam 35% Artificial Analysis, 11 Oct 2026Artificial Analysis run, preview GPQA Diamond Not published -
An open-weight 128B model from France for coding agents, reasoning and image input. Released under a Modified MIT licence with limits for very large companies.
- Released
- Apr 2026
- Inputs
- text, image
- Context
- 256k tokens
- Price
- USD 1.5 / 7.5 per million input/output tokens
- Use it via
- Mistral API; free download (Hugging Face)
- Tamil
- Not statedThe model card lists English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic; Tamil is not named.
Published scores Benchmark Score Source LMArena 1426 LMArena text leaderboard, 8 Oct 2026Listed as mistral-medium-3.5 AA Intelligence Index 14 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Mistral Medium 3.5 Humanity's Last Exam 13.8% Artificial Analysis, 11 Oct 2026Artificial Analysis run, default GPQA Diamond 74.8% Artificial Analysis, 11 Oct 2026Artificial Analysis run, default SWE-bench Verified 77.6%Maker's own figure Mistral Medium 3.5 model card -
Meta's open 30B model made for running AI agents on a laptop or PC with one consumer GPU, even offline. The Apache 2.0 licence allows commercial use.
- Released
- 10 Aug 2026
- Inputs
- text, image
- Context
- Not stated
- Price
- Not stated
- Use it via
- Free download (Hugging Face)
- Tamil
- Not statedMeta says it was trained on data from more than 100 languages; Tamil is not named.
Published scores Benchmark Score Source LMArena 1425 LMArena text leaderboard, 8 Oct 2026Listed as muse-glimmer AA Intelligence Index 17 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Muse Glimmer (high) Humanity's Last Exam 22% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort GPQA Diamond 83.5% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort -
Meta's closed frontier model, tuned for long agent tasks, coding and following detailed instructions. Developers reach it through the Meta Model API.
- Released
- 2 Sep 2026
- Inputs
- Not stated
- Context
- Not stated
- Price
- Not stated
- Use it via
- Meta Model API, Muse Code
- Tamil
- Not statedMeta's announcement does not list supported languages.
Published scores Benchmark Score Source LMArena 1494 LMArena text leaderboard, 8 Oct 2026Listed as muse-spark-1.3-max AA Intelligence Index 48 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Muse Spark 1.3 (max) Humanity's Last Exam 48.7% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort GPQA Diamond 93.5% Artificial Analysis, 11 Oct 2026Artificial Analysis run, max effort -
BharatGen's 17B mixture-of-experts model, pretrained from scratch with a focus on Indian languages. Released as an early post-training checkpoint under a non-commercial licence, so it suits research more than products.
- Released
- Feb 2026
- Inputs
- text
- Context
- 33k tokens
- Price
- Not stated
- Use it via
- Free download (Hugging Face), non-commercial licence
- Tamil
- Stated supportThe model card lists Tamil among the 21 Indian languages it supports besides English and Hindi.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index Not published Humanity's Last Exam Not published GPQA Diamond Not published Hugging Face model card: bharatgenai/Param2-17B-A2.4B-Thinking
-
A compact open Qwen model (27B, Apache 2.0) that reads text, images and video. Its deployment-friendly size makes it a popular base for local apps and fine-tuning.
- Released
- Aug 2026
- Inputs
- text, image, video
- Context
- 262k tokens
- Price
- USD 0.5 / 3 per million input/output tokens on Qwen Cloud
- Use it via
- Free download (Hugging Face); Qwen Cloud API
- Tamil
- Not statedThe model card does not list supported languages.
Published scores Benchmark Score Source LMArena 1438 LMArena text leaderboard, 8 Oct 2026Listed as qwen3.8-27b AA Intelligence Index 34 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Qwen3.8 27B (xhigh) Humanity's Last Exam 33.9% Artificial Analysis, 11 Oct 2026Artificial Analysis run, xhigh effort GPQA Diamond 89.2%Maker's own figure Qwen3.8-27B model card, Aug 2026 -
Alibaba's flagship, a 2.4-trillion-parameter model for coding and professional work. The hosted Max version adds image and video input; the base weights are published under Qwen's own licence.
- Released
- Aug 2026
- Inputs
- text, image, video
- Context
- Not stated
- Price
- USD 2 / 6 per million input/output tokens
- Use it via
- Qwen Cloud API, Qwen Studio; base weights on Hugging Face as Qwen3.8-2.4T-A95B
- Tamil
- Not statedThe model card does not list supported languages.
Published scores Benchmark Score Source LMArena 1483 LMArena text leaderboard, 8 Oct 2026Listed as qwen3.8-max AA Intelligence Index 45 Artificial Analysis Intelligence Index, 11 Oct 2026Listed as Qwen3.8 Max (0902) Humanity's Last Exam 43.1% Artificial Analysis, 11 Oct 2026Artificial Analysis run, Qwen3.8 Max (0902) GPQA Diamond 92.6%Maker's own figure Qwen3.8-2.4T-A95B model card, Aug 2026 -
Sarvam AI's flagship model, trained from scratch in India on compute from the IndiaAI Mission. Built for Indian languages, reasoning and coding, with Apache 2.0 open weights and a rupee-priced API.
- Released
- 6 Mar 2026
- Inputs
- text
- Context
- 128k tokens
- Price
- Rs 15 / 60 per million input/output tokens
- Use it via
- Sarvam API, Indus app; free download (Hugging Face, AI Kosh)
- Tamil
- Stated supportSarvam's docs list Tamil (ta-IN) among the 10 Indian languages plus English that Sarvam-105B supports, in native script, romanised and code-mixed input.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index Not published Humanity's Last Exam 11% Artificial Analysis, 11 Oct 2026Artificial Analysis run, high effort GPQA Diamond 78.7%Maker's own figure Sarvam-105B model card, Mar 2026 SWE-bench Verified 45%Maker's own figure Sarvam-105B model card -
An open translation model from Sarvam AI, built with AI4Bharat on Google's Gemma 3 4B. Translates whole documents between English and 22 Indian languages. GPL-3.0 licence.
- Released
- 7 Jun 2025
- Inputs
- text
- Context
- Not stated
- Price
- Rs 20 per 10,000 characters via API
- Use it via
- Sarvam API; free download (Hugging Face)
- Tamil
- Stated supportThe model card lists Tamil among the 22 Indian languages it translates to and from English.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index Not published Humanity's Last Exam Not published GPQA Diamond Not published -
A community Tamil and English model built on Llama 2 with an added Tamil vocabulary of about 16,000 tokens. Small and dated by today's standards, but one of the few public models made specifically for Tamil.
- Released
- Nov 2023
- Inputs
- text
- Context
- Not stated
- Price
- Not stated
- Use it via
- Free download (Hugging Face)
- Tamil
- Stated supportThe model card describes it as bilingual: English and Tamil.
Published scores Benchmark Score Source LMArena Not published AA Intelligence Index Not published Humanity's Last Exam Not published GPQA Diamond Not published
Not scored on this benchmark
No model matches this filter.
* The maker's own figure, from its model card or release post. Not tested independently, so treat it with care.
How to read this
- Bars compare the models shown, on the benchmark you picked. A model without a published score sits at the bottom.
- Benchmarks test narrow skills. Try a model on your own task, in your own language, before you pay for it.
- "Tamil support" means the maker or a source says so in writing. Many models handle some Tamil without saying it.
- LMArena and Artificial Analysis test models themselves. Where neither has published a score, we show the maker's own figure and mark it with *.
- We never estimate or adjust a score. Spotted a newer one? Tell us.