Best LLM in 2026: Which Model to Use Right Now

Claude Opus 5.5 is the best LLM right now. It scores 58 on the Artificial Analysis Intelligence Index, five points clear of the next model, and costs $4/$20 per million tokens. That's cheaper than Opus 5 was in July.

Update (September 26, 2026): The 58 is Opus 5.5 at max effort, the setting benchmark headlines use. At the setting most people run, it scores 53.6, still ahead of Claude Fable 5.1 (51.2) and GPT-6 Astra (49.6). For value, MiMo-V2.6-Pro scores 46.3 for about $0.13 per task. The LLM leaderboard has both views with real costs per task.

Most people still don't need it. GPT-6 Sol scores 48 for $2/$10. Muse Spark 1.3 scores the same 48 for $1.25/$4.25 and runs at about 219 tokens per second.

That's the short answer. Here's the full ranking and when each model makes sense.

Best LLM Rankings (September 2026)

Quality scores from the Artificial Analysis Intelligence Index v4.3.2. Total cost is input plus output price per 1M tokens. Pricing and speed from Artificial Analysis (first-party APIs), September 2026. AA rebased its index in September 2026, so every score here is lower than in earlier versions of this article.

Four of the top ten are superseded models you shouldn't start a project on. Below the top ten, the cheap end got much stronger:

The Quick Answer: Which LLM Should You Use?

For most people: GPT-6 Sol or Muse Spark 1.3.

Both score 48. Sol costs $2/$10 and has the OpenAI tooling around it. Muse Spark costs $1.25/$4.25, runs twice as fast, and takes text, image, and video input (Meta's model page). If speed or price is the constraint, pick Muse Spark. If you already build on OpenAI, pick Sol.

For volume: GPT-6 Luna or GLM-5.3 Flash.

Luna scores 37 at $0.10/$0.50. GLM-5.3 Flash scores 42 at $0.15/$0.50. Either one makes a million-call month cheap. Luna is faster (132 vs 43 tok/s). GLM-5.3 Flash is smarter.

For hard problems: Claude Opus 5.5.

Architecture decisions, debugging unfamiliar codebases, long agent runs. Opus 5.5 leads both the Intelligence Index (58) and the Coding Agent Index (66). It costs less than Claude Fable 5.1 or GPT-6 Astra and beats both.

Skip Fable 5.1 and Astra for most work. Both score 53 at $10/$50. Opus 5.5 scores 5 points higher at 40% of the price. Fable 5.1 still leads AA's Agentic Index (57.9 against Astra's 51), and Astra ties it on intelligence at roughly 40% of the cost per task, according to AA. That's a narrow reason to pay $50 per million output tokens.

Best LLM by Use Case

Best LLM for Coding

Winner: Claude Opus 5.5 (Coding Agent Index: 66)

Opus 5.5 is #1 on AA's Coding Agent Index v1.5, which tests models on DeepSWE, Terminal-Bench 4.0, and SWE-Atlas-QnA. Claude Fable 5.1 and GPT-6 Astra tie behind it at 62, both at $10/$50.

Budget alternative: GPT-6 Sol (57) at $2/$10, or Grok 4.7 (56) at $2/$6. Grok is cheaper on output but slow at 40 tok/s. GPT-6 Luna scores 41 at $0.10/$0.50 and handles sub-agent work.

For the full breakdown, see our best AI models for coding guide.

Best LLM for Writing

Winner: GPT-6 Astra (EQ-Bench Creative Writing Elo: 2173, $60/1M)

On EQ-Bench Creative Writing v3, GPT-6 Astra (2173) and Claude Fable 5.1 (2162) lead. Opus 5.5 scores 2050, so the best model overall isn't the best writer on that test. For research synthesis, where you feed it several documents and want one coherent summary back, Opus 5.5's 58 on the Intelligence Index still makes it the stronger pick.

Cheaper option: GPT-6 Sol at $2/$10 scores 2125, and it's the value pick on our creative writing ranking. Claude Sonnet 5 costs the same and scores 1794.

Best LLM for Data Analysis

Winner: Claude Opus 5.5 for accuracy, GPT-6 Sol for volume.

Use Opus 5.5 when a wrong number costs you something: statistics, multi-step calculations, reasoning across a large file. Every top-tier model except Grok 4.7 (500K) now takes at least 1M tokens, and the GPT-6 models take 1.05M, so context size no longer decides this.

Budget: Muse Spark 1.3 scores 48 at under a quarter of Opus 5.5's price.

Full comparison: best AI for data analysis.

Fastest LLM

Winner: Gemini 3.8 Flash (~297 tok/s, quality 41)

Google's fourth Flash model in under four months is the fastest thing in the top tier. It costs $0.75/$3.75. If you want more intelligence at speed, Muse Spark 1.3 scores 48 at about 219 tok/s. DeepSeek V4.1 Flash runs at about 232 tok/s for $0.30/$1.20 and scores 39.

Best Open-Weight LLM

Winner: MiMo-V2.6-Pro (quality 46, $0.43/$0.87)

Xiaomi's model is the highest-scoring open-weight model on AA's index, the value pick on most of our best-for pages, and it ties Grok 4.7 on the Intelligence Index. Kimi K3 scores 44 and has a Coding Agent Index score of 52, but at $3/$15 through Moonshot's API it costs far more. See our Kimi K3 guide for the self-hosting trade-offs.

Best Free LLM

Winner: DeepSeek R1 (free via OpenRouter)

It has reasoning and handles basic coding, analysis, and writing. But it scores 11 on the current index, has a 64K context window, and responds slowly. Fine for learning and experiments. For $0.60 per million tokens combined, GPT-6 Luna scores 37, which is close enough to free that most people should just pay.

The Cost-Performance Sweet Spot

Quality and price don't move together. At the bottom you get a lot for very little. At the top, the best model is no longer the most expensive one.

Going from GPT-6 Sol (48) to Opus 5.5 (58) costs an extra $12/1M for 10 more points. That's the best upgrade at the top end. Going from Opus 5.5 to Fable 5.1 or Astra costs an extra $36/1M and loses 5 points.

Our recommendation: Start with GPT-6 Sol or Muse Spark 1.3 as your default and GPT-6 Luna for volume. Escalate to Opus 5.5 when the cheaper model fails the task. You rarely need to go past it.

What Changed in 2026

Here's what's different since the July version of this ranking:

Anthropic took #1 with Opus 5.5. Released 22 September, it scores 58 and costs $4/$20, a 20% cut from Opus 5's $5/$25. Claude Fable 5.1 (1 September) replaced Fable 5 and cut cache reads to $0.25 per million tokens.

OpenAI shipped GPT-6. Astra (3 September) ties Fable 5.1 at 53. Sol and Luna (22 September) match GPT-5.6 Sol and Luna on quality at half the price. There's no reason left to run GPT-5.6. Our GPT-5.6 Sol vs Terra vs Luna guide covers the older tiers.

Meta reached the frontier. Muse Spark 1.3 (2 September) scores 48 and is one of the fastest frontier-class models. It's proprietary, unlike Meta's older Llama releases.

Cheap models got good. GLM-5.3 Flash scores 42 at $0.15/$0.50. MiMo-V2.6-Pro scores 46 at $0.43/$0.87.

xAI is now SpaceXAI on the leaderboards. Grok 4.7 (21 September) scores 46 and 56 on the Coding Agent Index at $2/$6. Its agent platform, Grok Bot, launched in August 2026.

How We Test

We rank models using three data sources:

1. Artificial Analysis Intelligence Index (v4.3.2): Independent benchmark built from 10 evaluations, weighted 30% agents, 20% coding, 30% general, and 20% scientific reasoning. This is the "quality score" in our rankings.

2. Coding Agent Index (v1.5): Also from Artificial Analysis. Tests models on agentic coding with DeepSWE, Terminal-Bench 4.0, and SWE-Atlas-QnA. AA only publishes it for some models, so we don't quote a coding score where there isn't one.

3. Measured cost and speed: AA's cost per task (thinking tokens included) and seconds to a full answer for each reasoning setting. We use them to pick each model's everyday setting: the strongest one that answers within 45 seconds.

Pricing and speed come from Artificial Analysis, which measures each provider's first-party API. Speed is median output tokens per second.

Compare Models Live

See the full rankings with interactive sorting and filtering in our LLM Leaderboard. Filter by task type, sort by any metric, find the best model for your specific workload.

For personalized recommendations based on your budget and use case, try the AI Model Selector.