Best AI Models

Opus 5.5 is the model to beat. One day after launch it sits at the top of the Artificial Analysis Intelligence Index with 58 points at max effort, five points clear of GPT-6 Astra and Claude Fable 5.1. Even at high effort it scores 54, which still beats every non-Anthropic model.

The price is the surprise. At $4 input and $20 output per million tokens it costs 20% less per token than Opus 5 did, and Anthropic says it costs about 40% less to run on typical work because it uses fewer tokens. Fable 5.1 and GPT-6 Astra both cost $10/$50, so Opus 5.5 gives you more intelligence for well under half the price.

Anthropic says it performs at the level of Fable 5.1 on most work and matches Fable and Mythos in biology and cybersecurity. Its own coding numbers back that up: 66.4% on Terminal-Bench 4.0 versus 55.8% for Fable 5.1. It also has a 1M-token context window and 128K output.

The catch: it is brand new, so independent coding tests and Arena votes will take a few weeks to settle. For now, it is our default pick for developers and heavy users of the Claude app.

Muse Spark 1.3 is the model that put Meta back in the frontier race. It scores 48 on the Artificial Analysis Intelligence Index, the same as GPT-6 Sol, and it is far faster: about 213 tokens per second at max effort, roughly twice the speed of GPT-6 Sol. On the Arena human-vote board it scores 1493, close to Claude Opus 5.

Pricing is low. The standard API tier costs $1.25 input and $4.25 output per million tokens. A Contributor tier drops that to $0.10/$0.20. Launch coverage reports that Meta can use Contributor traffic to train its products, so read the terms and keep private data off it. The Meta Model API is compatible with the OpenAI SDK and has built-in web search with citations.

Meta reports 75.4% on DeepSWE 1.1 and 98.5% on a long-context retrieval test inside its 1M-token window. Those are Meta's numbers, and independent coding results still trail the Claude and GPT-6 leaders.

Muse Spark is closed, unlike Meta's older Llama models. Bloomberg reported that Meta is rolling it out to its free Meta AI assistant and its social apps; developers get it through Muse Code and the Meta Model API. For developers who want speed and low cost with near-frontier smarts, it is a strong choice. On our weights its price and speed lift it to second place overall, even though GPT-6 Astra and Claude Fable 5.1 score higher on intelligence.

GPT-6 Astra is OpenAI's top model and the second-smartest model you can use today. It scores 53 on the Artificial Analysis Intelligence Index at max effort, tied with Claude Fable 5.1, and 52 at xhigh. It handles text and images, reads up to 1.05M tokens, and writes up to 128K tokens in one reply. Through the API it works with web search, code interpreter, computer use and MCP tools.

The weak spot is cost. At $10 input and $50 output per million tokens it costs 2.5 times as much as Claude Opus 5.5, which now scores higher. Long prompts get pricier still: past 272K input tokens, input costs double and output costs 1.5 times as much.

Access is also narrower than the name suggests. In regular ChatGPT it appears as GPT-6 Pro and is limited to the Pro plans (from $100; the $200 tier has been closed to new sign-ups since 10 September 2026) plus Business and Enterprise. Plus users only get it inside ChatGPT Work and Codex, and Free and Go users do not get it. OpenAI rolled it out gradually, so some eligible accounts waited.

Pick Astra if your team already builds on OpenAI tools and needs the top tier. Otherwise GPT-6 Sol gives most of the quality for a fifth of the price.

GPT-6 Sol is the best deal from OpenAI. It sits one step below Astra, uses training methods developed for Astra, and costs $2 input and $10 output per million tokens, a fifth of Astra's price and about half of GPT-5.6 Sol's.

On the Artificial Analysis Intelligence Index it scores 48 at max effort, level with Meta's Muse Spark 1.3 and one point above GPT-5.6 Sol. At lower effort levels the score falls quickly (44 at xhigh, 43 at high), so use max effort for hard problems. OpenAI says Sol at xhigh effort scored 33.2% on AutomationBench against 26.9% for Claude Opus 5, at a much lower cost per task.

The context window is 1.05M tokens, though input tops out at 922K, and output can reach 128K. Through the Responses API it gets web search, file search, code interpreter and computer use.

The odd part is access. At launch Sol was in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, but not in the regular ChatGPT chat window. For developers that doesn't matter. For API coding agents where Opus 5.5 is too expensive, Sol is our pick. Its sibling GPT-6 Luna ($0.10/$0.50) handles simple, high-volume jobs.

Fable 5.1 is Anthropic's largest model and, until Opus 5.5 arrived three weeks later, was the top model on most leaderboards. It still scores 53 on the Artificial Analysis Intelligence Index, level with GPT-6 Astra. On the Arena human-vote board, Fable models hold the top spot (Fable 5 high at 1506) and Fable 5.1 max sits at 1498.

Anthropic reports 60.9% on Humanity's Last Exam without tools and 77.9% on OSWorld 2.0 (partial credit), both clear gains over Fable 5 and Opus 5. Anthropic still recommends it for the most demanding reasoning and long-horizon agent work, or when Opus 5.5 at high effort falls short on your own tests.

Fable 5.1 costs the same $10/$50 as Fable 5, but cache reads fell 75% to $0.25 per million, and Anthropic estimates typical workloads cost about 25% less. It is also slower than Opus 5.5 (about 61 to 66 tokens per second at top effort).

The honest take: Opus 5.5 now scores higher for 40% of the price, so most people should switch. Keep Fable 5.1 for jobs where you have proof it does better. Its twin, Claude Mythos 5.1, is the same model with looser safeguards, open only to vetted US security and life-science groups.

GLM-5.3 is the second-highest open-weight model on the Artificial Analysis Intelligence Index, scoring 45 at max effort (Xiaomi's MiMo-V2.6-Pro scores 46), one point behind Grok 4.7 and level with Qwen3.8 Max. You can download the 753-billion-parameter weights from Hugging Face or pay Z.ai $1.40 input and $4.40 output per million tokens.

Z.ai built GLM-5.3 by extending post-training on the GLM-5.2 base. Its main gains are coding and cybersecurity. Z.ai reports 66.9% on DeepSWE 1.1 and 84.5% on CyberGym, a vulnerability-finding test. Those security skills are why Z.ai held back the weights for about two weeks for a safety review.

The model has a 1M-token context window and three reasoning levels (low, high, max). Speed is middling at about 61 tokens per second on Artificial Analysis. On Arena it scores 1483, level with GPT-5.6 Sol.

The weights use a custom GLM-5.3 license, so read it before shipping a product. Self-hosting a 753B model also needs a multi-GPU server. For most teams, the cheap API is the easy route, and running it yourself is the option when data can't leave your servers.

Google's best model you can use today is not a Pro model. Gemini 3.5 Pro was promised for June and has slipped repeatedly. Google now says it is testing with partners, with no public date. That leaves Gemini 3.8 Flash as the flagship.

It is a very good flagship for a Flash model. It scores 41 on the Artificial Analysis Intelligence Index, well below the leaders, but on the Arena human-vote board it scores 1493, level with Muse Spark 1.3 and Claude Opus 5. People like its answers more than its test scores suggest.

Speed and price are where it wins. It outputs about 283 tokens per second, the fastest model in this list. Until 31 December 2026 it costs $0.75 input and $3.75 output per million tokens. From 1 January 2027 that doubles to $1.50/$7.50, so budget for the rise. The Gemini API also has a free tier.

It reads text, images, audio, video and PDFs across a 1M-token window, but output caps at about 65K tokens, half of what Claude and GPT-6 allow. Use it for chat apps, document processing and fast agents. Pick Opus 5.5 or Astra for the hardest reasoning.

Grok 4.7 is a solid step up from Grok 4.6 at the same price. SpaceXAI built it on a larger base model and trained it longer on tasks that take hours, not single replies. It is also meant to check its own work more often before answering.

It scores 46 on the Artificial Analysis Intelligence Index, level with the top open-weight model (MiMo-V2.6-Pro) and two points below GPT-6 Sol. The biggest gain is in agentic coding: SpaceXAI reports 37.6% on Terminal-Bench 4.0, up from 20.3% for Grok 4.6, though independent testing by Artificial Analysis found 26%, per The Decoder. That is still far behind Claude Opus 5.5's 66.4% on the same test.

The price is $2 input and $6 output per million tokens for prompts under 200K tokens, doubling above that. That undercuts GPT-6 Sol on output. The context window is 500K tokens, half of most rivals. Speed is a strength: Artificial Analysis measured about 188 tokens per second.

The model launched on the SpaceXAI API, in Cursor, in Grok Build and in the Grok app on the same day. A Fast variant runs about twice as quickly at about twice the price. Pick Grok 4.7 if you already use Cursor or the Grok ecosystem. Otherwise Sol or Opus 5.5 is the stronger buy.

Qwen3.8 Max is Alibaba's largest model: 2.4 trillion parameters, of which about 95 billion are active for each token. That design, called mixture of experts, keeps it cheaper to run than its size suggests. The 0902 snapshot scores 45 on the Artificial Analysis Intelligence Index, level with GLM-5.3.

The API costs $2 input and $6 output per million tokens, and it accepts text, images and video. Alibaba reports 67.7 on SWE-bench Pro and 86.1 on OSWorld-Verified, and says the September update more than doubled its TerminalBench 3.0 score, from 11.3 to 29.0.

It is also the first Max-class Qwen you can download. But the open release (Qwen3.8-2.4T-A95B) is text-only, always runs in thinking mode, and supports 262K tokens natively rather than the API's 1M. It uses a custom license, not the Apache 2.0 license of smaller Qwen models.

The weak point is speed: about 39 tokens per second on Artificial Analysis, among the slowest here. Choose it if you want a strong multimodal API at a low price and like having downloadable weights as a backup.

DeepSeek V4 is the family to pick when price matters most. It comes in two main versions on the API. V4.1 Flash, released on 10 September, costs $0.15 input and $0.60 output per million tokens off-peak. V4 Pro 0813 costs $0.66/$1.98 off-peak. Prices double during peak hours in China's working day (01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays).

Both use the MIT license, the most permissive in this list. You can run them, change them and sell products built on them without a revenue cap.

On the Artificial Analysis Intelligence Index, the new V4.1 Flash scores 39 and runs at about 219 tokens per second. Oddly, the bigger V4 Pro 0813 scores lower at 36. DeepSeek reports 80.6% on SWE-bench Verified and 67.9 on Terminal-Bench 2.0 for V4 Pro, but independent tests so far put it well behind the US leaders and behind GLM-5.3 and Kimi K3.

Both support a 1M-token context and very long outputs of up to 384K tokens. For bulk jobs like tagging, summarizing or first-draft code, nothing here beats it on cost. For hard reasoning, spend more on a model higher up this list.

Kimi K3 is the largest open-weight model released so far, with 2.8 trillion parameters and 104 billion active per token. It scores 44 on the Artificial Analysis Intelligence Index at max effort, one point behind GLM-5.3 and Qwen3.8 Max. On Arena it scores 1485, a little above GLM-5.3.

Moonshot's own numbers point to agent work as its strength: 91.2 on BrowseComp (web research), 84.8 on OSWorld-Verified (using a computer) and 88.3 on Terminal-Bench 2.1. It also handles images well. Treat these as vendor claims until more independent results arrive.

The drawback is price. At $3 input and $15 output per million tokens, Kimi K3 costs more than GPT-6 Sol ($2/$10), which scores higher. Cache hits drop input to $0.30, which helps agents that resend the same context. Speed is slow, at about 37 tokens per second.

The weights are downloadable under the Kimi K3 License. It is fine for most uses, but companies that sell the model as a hosted service and earn over $20 million a year need a separate agreement. Use it through the Kimi app or self-host it when you need an open model with strong agent skills.