Language Models Language Models General-purpose chat and reasoning models — the models most people mean when they say “AI.” Frontier flagships, their budget tiers, and the open-weight and regional alternatives.
Anthropic 2026.09.01
+ Leads on long-horizon agentic coding, with strong root-cause debugging across multi-step sessions + Strong agentic scientific-research performance, including protein-design work Anthropic reports near 50% hit rates on − Anthropic reports reduced visibility into very long-context and multi-agent runs versus shorter sessions − Can still bypass approval gates or auto-mode classifiers on some agentic tasks, per Anthropic's own testing Anthropic 2026.07.24
+ Anthropic's recommended model for complex, high-stakes work, with strong reasoning-to-cost + Zero-data-retention eligible, useful for regulated or enterprise deployments − Sits below the flagship tier in branding despite strong practical scores − Costs meaningfully more per token than the Sonnet tier for everyday tasks Anthropic 2026.09.22
+ Tops LMArena outright on hard prompts (1,545) and on the WebDev arena (1,827), the highest rating on either board + Anthropic reports roughly 40% lower cost per workload than Opus 5 and over 30% faster output, at $4/$20 per million tokens − Released weeks ago, so creative writing is still below this site's vote floor and no hallucination testing exists yet — its overall score is pending − Still the premium tier: many times the price of the fast models that most high-volume work runs on Anthropic 2026.06.09
+ 3rd of 380 on LMArena's creative-writing arena, and 7th of 399 on hard prompts, on tens of thousands of votes + Built for long-horizon agentic work, with a 1M-token context and up to 128K output tokens per request − Anthropic released it and then temporarily suspended it in June 2026 before restoring access − Superseded by Fable 5.1 in September 2026 at the same $10/$50 per million tokens Google 2026.09.02
+ Highest-rated model in the directory on LMArena's hard prompts after Anthropic's Opus line, on 14,320 votes + Google reports 90.8% on Terminal-Bench 2.1, up from 81.6% for 3.7 Flash, at $0.75/$3.75 per million tokens − Built on 3.7 Flash rather than a new base model, and Google tells you to stay on 3.7 for efficiency-first work − Spends more thinking tokens by design, so the headline price understates what a task actually costs Moonshot AI 2026.07.16
+ The largest open-weight release yet at 2.8T parameters with 104B active, served by 17 independent providers + Moonshot reports 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.1, and LMArena ranks it 8th of 133 on WebDev − $3/$15 per million tokens, several times the price of the open-weight models it competes with − Its licence is only open to a point: model-as-a-service businesses over $20M a year need a separate agreement Xiaomi 2026.09.22
+ The strongest open-weight model measured here: 10th of 399 on LMArena's hard prompts, MIT-licensed at 1T parameters with 42B active + Native text, image, video and audio input with a 1M-token context, at $0.44/$0.87 per million tokens − Xiaomi is new to frontier releases, so its production track record is thin next to the established labs − 27th on creative writing against 10th on hard prompts — clearly stronger at reasoning than at expressive writing Meta 2026.09.02
+ Meta's first frontier line since it moved off Llama, natively multimodal across text, image and video at $1.25/$4.25 per million tokens + Meta reports roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on the same agentic work − Proprietary, ending the open-weight position Llama gave Meta, and served by a single provider on OpenRouter − 10th of 409 on LMArena's overall text arena — strong, but behind the Anthropic and Google models above it Alibaba 2026.08.03
+ 6th of 133 on LMArena's WebDev arena, the highest-rated non-frontier model for building working web apps + 2.4T mixture-of-experts with about 95B active and a 1M-token context, at $2/$6 per million tokens − Loses to GLM-5.3 on seven of the eight benchmarks both makers publish, at several times the price − Served by a single provider on OpenRouter, so there is no fallback if it is unavailable Zhipu AI 2026.08.14
+ Open weights across 32 independent providers at $0.38/$1.19 per million tokens, the cheapest model at this rating + Z.ai reports state-of-the-art open-weight results on Terminal-Bench 3.0 and roughly 50% better coding than GLM-5.2 − A post-training upgrade on the GLM-5.2 base rather than a new model, so the gains are tuning rather than scale − 17th of 399 on LMArena's hard prompts: clearly behind the frontier it undercuts on price DeepSeek 2026.09.10
+ MIT-licensed and openly downloadable at 552B parameters with only 8B active, served by 27 independent providers + DeepSeek reports 74.2% on Deep-SWE v1.1, ahead of both Opus 5 (74.0%) and GPT-5.6 Sol (73.0%) − 36th of 409 on LMArena's overall text arena, well behind the frontier models it matches on published coding benchmarks − A Flash-tier model by design: DeepSeek's own Pro line is stronger on maths and long reasoning OpenAI 2026.07.09
+ Best-in-class agentic reasoning, with search, code execution, and computer use in one API + Disciplined, low-hallucination output on long, multi-step tasks − Long-context requests above roughly 272K tokens get repriced sharply higher − Slower to produce a first answer than most rivals at max reasoning DeepSeek 2026.04.24
+ Genuinely open-weight frontier model you can self-host, unlike closed rivals + Very low API cost, especially during off-peak pricing windows − V4 Pro is text-only; image input is limited to an experimental Flash variant − Trails the closed frontier labs on aggregate intelligence benchmarks Google 2025.12.17
+ Among the fastest decoding speeds of any frontier-class model + Aggressively priced for high-throughput production traffic − Newer Flash versions (3.5 to 3.8) have since replaced it as Google’s default fast model − Higher reasoning settings add noticeable cost and latency OpenAI 2026.09.03
+ Rated #1 of 128 models on LMArena’s WebDev arena for building working web apps, well clear of every rival + OpenAI reports state-of-the-art results on FrontierMath, ARC-AGI-3, and OSWorld 2.0 computer use − In everyday use people prefer many rivals: #49 of 402 on LMArena’s hard prompts, well behind Claude − Uses an 'opaque recurrence' reasoning style that OpenAI says is harder to monitor for chain-of-thought audits Anthropic 2026.06.30
+ Fast and capable, priced for everyday production use + Strong default for coding and long documents without Opus-level cost − Enabling maximum thinking mode can quietly balloon token spend − Trails Opus on the hardest multi-step reasoning problems Google 2025.11.18
+ Native multimodal input across text, image, audio, video, and PDFs + Very large context window with strong long-document recall − Slower first response than Google's own Flash tier − No native image or audio output, understanding only OpenAI 2026.09.22
+ 5th of 133 on LMArena's WebDev arena, OpenAI's strongest showing for building working web apps after Astra + 93.5% factual consistency on Vectara's hallucination leaderboard, ahead of GPT-6 Astra's 91.3% − 156th of 409 on LMArena's overall text arena — in blind comparisons people prefer a great many cheaper models − Sits below Astra in OpenAI's own line-up, so it doesn't carry the flagship's headline capabilities OpenAI 2026.09.22
+ Priced for volume at $0.10/$0.50 per million tokens, a twentieth of Sol, with the same 1.05M-token context + Built for high-volume clerical work — pulling fields from documents, summarising tickets — where throughput matters more than depth − 160th of 409 on LMArena's overall text arena and 155th on creative writing: not a model to reach for on hard or expressive work − OpenAI positions it two tiers below Astra, so complex reasoning and long agentic runs are outside its brief SpaceXAI 2026.08.12
+ Leads several agentic and terminal-automation benchmarks + Native, real-time X and web search built into the model − Reasoning mode cannot be turned off, adding cost and latency to every call − Hits a steep pricing cliff once context passes roughly 200K tokens Google 2026.04.02
+ Purpose-built for on-device and edge deployment + Small variants, from 4B to 12B, run well on modest hardware − Not intended to compete with frontier models on hard reasoning − Limited context window compared to Gemini's cloud models Zhipu AI 2026.02
+ Coding performance that rivals or matches Claude on several benchmarks + Open-weight, giving flexibility for self-hosted or fine-tuned deployments − Smaller ecosystem of tooling and integrations than the big three labs − Less battle-tested in production outside China-based deployments SpaceXAI 2026.09.21
+ SpaceXAI reports Terminal-Bench 4.0 nearly doubling to 38.0% from Grok 4.6's 20.3% on multi-hour terminal tasks + Holds Grok 4.6's $2/$6 per million tokens despite a larger base model and a longer reinforcement-learning run − 151st of 399 on LMArena's hard-prompts arena — the gains its maker reports on agentic benchmarks don't show up in blind human preference − SpaceXAI's own figures put it behind Claude Fable 5.1 on CursorBench 4.0, 46.3% against 51.8% Anthropic 2025.10.15
+ Cheapest Claude tier, well suited to high-volume, simple tasks + Low latency, good for chat-style and classification workloads − Smaller context window (200K tokens) than Sonnet 5 and Opus 5 (1M) − Noticeably weaker on hard reasoning than Sonnet or Opus Mistral AI 2026.04.29
+ Strong multilingual performance, especially across European languages + Apache 2.0 license permits unrestricted commercial use − Smaller research budget than the US/China frontier labs shows in ceiling performance − Less name recognition can complicate enterprise procurement Anthropic 2026.09.28
+ Anthropic reports at least 30% faster output than Sonnet 5, with fewer tokens needed for many of the same tasks + Available across Anthropic, AWS, Google Cloud, and Azure at $2 input/$10 output per million tokens − Released on September 28, so this directory has no qualifying independent capability measurements for it yet − Anthropic positions Opus 5.5 above it for complex work requiring careful judgment Mistral AI 2026.03.16
+ One open-weight model that combines reasoning, image input, and agentic coding, under Apache 2.0 + Mixture-of-experts design activates only 6.5B of 119B parameters per token, keeping inference efficient − A small-tier model: trails frontier flagships on the hardest reasoning tasks − At 119B total parameters, self-hosting needs far more memory than earlier Small models Baidu 2025.11.13
+ Deep integration with Baidu's search and China-market ecosystem + Strong Mandarin-language understanding and generation − Weaker mindshare and tooling support outside China − English-language performance trails the Western frontier labs NVIDIA 2026.06.04
+ Fully open: weights, training data, and recipes released under the permissive OpenMDW-1.1 licence + 550B-parameter hybrid Mamba-Transformer (55B active) built for long-running agents − At 550B parameters, self-hosting needs a multi-GPU cluster − Its most efficient format (NVFP4) is designed for NVIDIA hardware AI21 Labs 2026.01.08
+ Hybrid Transformer-Mamba architecture gives very high throughput and low memory use at long context + One of the largest context windows among open-weight models, aimed at enterprise long-document work − Needs its own proprietary quantization approach to hit those efficiency numbers − Much smaller track record and mindshare than Llama, Qwen, or DeepSeek Tencent 2024.11
+ Strong presence in the Chinese consumer and enterprise market via Tencent's ecosystem + Broad multimodal family spanning text, image, and video under one brand − Limited adoption and tooling outside China − Less benchmarked against Western frontier models in independent tests Reka AI 2025.03
+ Compact, efficient multimodal model that's easy to self-host + Good cost-to-performance ratio for mid-tier multimodal tasks − Trails frontier labs on the hardest reasoning benchmarks − Much smaller brand recognition and community than the major labs ByteDance 2025.01
+ Deeply integrated into ByteDance's own consumer apps and ecosystem + Competitive general-purpose performance at low cost − Primarily built for the Chinese market, with limited presence elsewhere − Less transparent benchmarking against global frontier models Upstage 2025.07
+ Strong performance for its size, popular in Korean-language and enterprise RAG use cases + Efficient enough to run cost-effectively at scale − Narrower global mindshare than the major US/China labs − Falls behind frontier models on the hardest general reasoning tasks 01.AI 2024.10
+ Fast, low-cost model that performs well for its price point + Reasonable multilingual coverage including English and Chinese − Trails top-tier frontier models on complex reasoning and coding − Smaller developer ecosystem than Qwen or DeepSeek ← Next section → Coding & Agentic