Best AI Models 2026: Ratings by Task & Strength

What changed · reviewed 10 September 2026

Figure	Was	Now	Source & date
Alibaba flagship	Qwen3.7-Max	Qwen3.8-Max — Intelligence Index 58	Report AI, Chinese AI Models 2026, 3 Sep 2026 · HIGH
Chinese models in global top four	none listed	Qwen3.8-Max (58) and Kimi K3 (57.1)	Report AI, Chinese AI Models 2026, 3 Sep 2026 · HIGH
US-frontier composites	Sol 58.9 · Opus 4.8 55.7 · Terra 55.0	unchanged — pending re-verification	Carried forward from Jul 2026; not yet re-checked against the primary leaderboard

Methodology

Best for: agentic workflows, science & reasoning

R&M ●●●●● · Code ●●●●● · Write ●●●●● · MM ●●●●○ · Ctx ●●●●○ · Speed ●●●○○ · Price ●●○○○ · Rel ●●●●○ · Agent ●●●●●

Generally available July 9, 2026, the Sol flagship tops the Artificial Analysis Intelligence Index at 58.9 and sets state-of-the-art marks on agentic evals — BrowseComp 92.2%, OSWorld 2.0 62.6%, and the Coding Agent Index at 80. Its clear weakness is price: at $5 / $30 per million tokens it is the most expensive flagship here, so the cheaper Terra ($2.50 / $15) and Luna ($1 / $6) tiers carry most everyday work.

Best for: agentic coding & long-form writing

R&M ●●●●● · Code ●●●●● · Write ●●●●● · MM ●●●○○ · Ctx ●●●●○ · Speed ●●●○○ · Price ●●●○○ · Rel ●●●●● · Agent ●●●●●

Launched May 28, 2026 at $5 / $25, Opus 4.8 leads the coding-focused SWE-bench Pro board (69.2%) and scores 88.6% on SWE-bench Verified, and it holds the top tier of the LMArena text leaderboard within a statistical tie. Its reliability and writing quality are class-leading; its softest dimension is multimodal, where it trails Gemini on vision, audio and video. For teams that want Opus-class coding cheaper, Sonnet 5 ($3 / $15, $2 / $10 intro) is the value sibling.

Best for: multimodal understanding & 2M-token context

R&M ●●●●● · Code ●●●●○ · Write ●●●●○ · MM ●●●●● · Ctx ●●●●● · Speed ●●●●○ · Price ●●●●○ · Rel ●●●●○ · Agent ●●●●○

Gemini 3.1 Pro pairs the frontier’s largest usable window — 2M tokens — with the best multimodal scores in the field: MMMU-Pro 81.0% and GPQA Diamond 94.3%, plus a LiveCodeBench Pro Elo near 2,439. It sits in the LMArena top three in a statistical tie with Opus. Coding trails Claude and GPT-5.6 by a hair on agentic SWE tasks, but for anything spanning documents, images, audio and video, nothing else is close.

Best for: long-horizon agents & multilingual work

R&M ●●●●● · Code ●●●●○ · Write ●●●●○ · MM ●●●●○ · Ctx ●●●●○ · Speed ●●●●○ · Price ●●●●○ · Rel ●●●●○ · Agent ●●●●●

Alibaba’s current flagship, and the strongest non-US frontier model. Our 3 September 2026 review put it at an Artificial Analysis Intelligence Index of 58 — inside the global top four. That score has been volatile across successive captures (53 → 56 → 58), so treat the precise value as MEDIUM confidence rather than a settled number. It supersedes Qwen3.7-Max (May 2026), whose figures this page previously carried and which we retain for comparison: GPQA Diamond 92.4%, HMMT Feb 2026 97.1%, SWE-bench Verified 80.4%, and a demo run of 35 straight hours across 1,158 tool calls. The top Max tier is proprietary, but the open-weight Qwen line makes the ecosystem exceptionally deployable — including Qwen3.8-27B (index 52), which runs on a single consumer GPU. Note: the nine-criteria ratings in the matrix above are inherited from the Qwen3.7-Max assessment and are due for re-rating.

Best for: cheapest near-frontier reasoning, open weight

R&M ●●●●● · Code ●●●●○ · Write ●●●●○ · MM ●●●○○ · Ctx ●●●●○ · Speed ●●●●○ · Price ●●●●● · Rel ●●●●○ · Agent ●●●●○

Shipped April 24, 2026 under an MIT license, the 1.6T-parameter V4-Pro is the standout open-weight value: MMLU-Pro 87.5, GPQA Diamond 90.1, SWE-bench Verified 80.6% (tied with Gemini 3.1 Pro as the top open entry) — at roughly $0.87 per million output tokens. It is weakest on multimodal, and NIST’s CAISI review judged its all-round capability to lag the closed frontier by about eight months, but on price-per-point of reasoning nothing beats it.

Best for: real-time agentic tool use

R&M ●●●●○ · Code ●●●●○ · Write ●●●●○ · MM ●●●○○ · Ctx ●●●○○ · Speed ●●●●○ · Price ●●●●○ · Rel ●●●○○ · Agent ●●●●●

Released July 8, 2026, Grok 4.5 lands around #4 on the Artificial Analysis Intelligence Index (score 54) and takes the single top spot on agentic tool use, backed by Terminal-Bench 2.1 at 83.3% and SWE-bench Pro at 64.7%. At $2 / $6 it is priced aggressively. The trade-offs are a smaller 500K context window and a shorter reliability track record than the incumbents.

Best for: open-weight, ultra-long-context self-hosting

R&M ●●●○○ · Code ●●●○○ · Write ●●●○○ · MM ●●●○○ · Ctx ●●●●● · Speed ●●●●○ · Price ●●●●● · Rel ●●●○○ · Agent ●●●○○

Announced April 8, 2026, Llama 5 is a ~600B open-weight model with a headline 5M-token context. Its value proposition is freedom — free to self-host with the longest context of any release here. But coverage as of July 2026 is mixed: reporting disputes whether Llama 5 is a full frontier generation, and independent boards place its reasoning and coding a clear tier below the closed leaders. It shines when you need weights you own plus enormous context, not when you need the top benchmark score.

Best for: EU data-sovereign, efficient open deployment

R&M ●●●○○ · Code ●●●○○ · Write ●●●●○ · MM ●●●○○ · Ctx ●●●○○ · Speed ●●●●○ · Price ●●●●○ · Rel ●●●○○ · Agent ●●●○○

Mistral Large 3 (December 2025) is a 675B-parameter open-weight MoE — the largest from a Western lab — scoring MMLU-Pro 73.11 and MATH-500 93.6 on independent evals. It is a tier below the US and Chinese frontier on raw capability, but its European base, permissive weights, and efficient Medium 3.5 and Small 4 tiers make it the default for cost-sensitive, data-sovereign EU deployments.

What is the best AI model in 2026?

There is no single winner, and the answer now depends on which month you ask. As of July 2026, GPT-5.6 Sol held the top overall composite (Artificial Analysis Intelligence Index 58.9), with Claude Opus 4.8 and Gemini 3.1 Pro in a statistical tie just behind on LMArena. Our 3 September 2026 review then placed two Chinese models inside the global top four — Alibaba’s Qwen3.8-Max (58) and Moonshot’s Kimi K3 (57.1). The US-frontier composites on this page are carried forward from July and are pending re-verification, so treat the ordering at the very top as provisional. The best model still depends on the task — see the matrix above.

What is the best AI model for coding?

Claude Opus 4.8 leads SWE-bench Pro (69.2%) and is the reference for agentic coding, with GPT-5.6 Sol essentially tied on agentic coding evals. For value, Claude Sonnet 5 delivers near-Opus coding at roughly 40% lower cost.

What is the cheapest capable AI model?

DeepSeek V4-Pro — an MIT-licensed open-weight model at about $0.87 per million output tokens that still posts MMLU-Pro 87.5 and GPQA Diamond 90.1, near-frontier reasoning for a fraction of the flagship price.