Best AI Models for Coding (September 2026): Ranked by Price and Performance

Claude Opus 5.5 is the best AI model for coding right now. Anthropic released it on 22 September and it went straight to #1 on both Artificial Analysis indexes: 66 on the Coding Agent Index and 58 on the Intelligence Index. It also costs less than Opus 5 did, at $4/$20 per million tokens instead of $5/$25.

OpenAI shipped GPT-6 Sol and GPT-6 Luna the same day. Sol scores 57 on coding at $2/$10. Luna scores 41 at $0.10/$0.50, which makes it the cheapest model with a published Coding Agent score.

The price spread is enormous. Claude Fable 5.1 and GPT-6 Astra list at $10/$50. Luna lists at $0.10/$0.50, a 100x gap. Picking the wrong model means burning money or getting worse code.

Rankings below use the Artificial Analysis Coding Agent Index v1.5 and Intelligence Index v4.3.2. AA rebased its index in September 2026, so scores are lower than in earlier versions of this article. You can check every number in our benchmark dashboard.

Update (September 26, 2026): Both indexes are measured at each model's maximum reasoning effort, which is slower and far more expensive than how most people run these models. The costs and speeds below are now for each model's everyday setting (its strongest setting that answers within 45 seconds), with thinking tokens included. That changes the value picture a lot, so check the cost table before choosing. For a live ranking with an effort switch and monthly costs, see best AI model for coding.

The Full Rankings

Only models with a published Coding Agent Index v1.5 score are ranked here. The top four are separated by 9 points. Per task, GPT-6 Sol costs about a sixth of Claude Fable 5.1 and about a third of Opus 5.5.

1. Claude Opus 5.5: Best Coding Model Overall

• Coding Agent Index: 66 (#1)
• Intelligence Index: 58 (#1)
• Pricing: $4.00 / $20.00 per 1M tokens ($0.20 cache reads)
• Context: 1M tokens
• Max output: 128K tokens

Opus 5.5 leads both indexes by a clear margin: 4 points on coding, 5 on intelligence. It replaces Opus 5 at a 20% lower price, and it's 60% cheaper per token than Fable 5.1 or GPT-6 Astra.

Thinking is always on. You control depth with effort levels from low to max, and the score moves a lot with them: 58 at max, 56 at xhigh, 54 at high, 51 at medium, 42 at low. The default is medium. If you're running it on hard refactors at the default, you're leaving 7 points on the table. At high effort a full answer takes about 41 seconds; extra high takes over two minutes.

Best for: Hard bugs, multi-file refactors, long agent runs, architecture work. The model to run in Claude Code for anything that matters.

2. GPT-6 Astra: OpenAI's Top Coding Model

• Coding Agent Index: 62
• Intelligence Index: 53
• Pricing: $10.00 / $50.00 per 1M tokens
• Context: 1.05M tokens
• Max output: 128K tokens

Astra launched on 3 September and ties Claude Fable 5.1 on both indexes. Same list price, but at max effort AA spent about $3.26 per Intelligence Index task on Astra against $7.63 on Fable 5.1. It uses fewer tokens to get there.

At medium effort it answers in about 15 seconds, the quickest of the top four, but it still costs more per task than Opus 5.5, which beats it by 4 points.

Best for: Teams locked into OpenAI who need its strongest model for the hardest problems.

3. Claude Fable 5.1: Agentic Knowledge Work

• Coding Agent Index: 62
• Intelligence Index: 53
• Pricing: $10.00 / $50.00 per 1M tokens ($0.25 cache reads)
• Context: 1M tokens
• Max output: 128K tokens

Fable 5.1 (1 September) replaced Fable 5, which now scores 50. The 5.1 release cut cache reads by 75% to $0.25 per million, which helps agents that re-read the same codebase all day.

Its lead is in agentic knowledge work, not code. For coding, Opus 5.5 scores higher and costs 60% less per token. Three weeks after launch, Fable 5.1 is hard to justify for a coding workload.

Best for: Research-heavy agent work where code is only part of the job.

4. GPT-6 Sol: Best Value Coding Agent

• Coding Agent Index: 57
• Intelligence Index: 48
• Pricing: $2.00 / $10.00 per 1M tokens
• Context: 1.05M tokens
• Max output: 128K tokens

Sol matches GPT-5.6 Sol on intelligence (48 vs 47) at half the price. It scores 57 on the Coding Agent Index, and at its everyday setting 1,000 coding tasks cost about $109, a third of Opus 5.5.

Claude Sonnet 5 lists at the same $2/$10, but it thinks so much that 1,000 tasks cost about $350, more than Opus 5.5. If you want one mid-priced model for daily coding, Sol is it.

Best for: Daily coding when you want a single model. Feature work. Agent loops that run all day.

5. Grok 4.7: Fastest in the Top Five

• Coding Agent Index: 56
• Intelligence Index: 46
• Pricing: $2.00 / $6.00 per 1M tokens
• Context: 500K tokens

Grok 4.7 landed on 21 September from SpaceXAI (xAI's new name on Artificial Analysis). It sits one point behind GPT-6 Sol on coding and answers in about 12 seconds, the quickest of the top five.

Its low list price is misleading. It thinks a lot, so 1,000 coding tasks cost about $362, more than Opus 5.5. The 500K context window is also half what most rivals offer, though enough for most single-repo work.

Best for: Interactive coding where you want fast answers from a strong model and cost is secondary.

6. Muse Spark 1.3: Screenshots and Video In

• Coding Agent Index: 54
• Intelligence Index: 48
• Pricing: $1.25 / $4.25 per 1M tokens
• Context: 1M tokens

Meta's Muse Spark 1.3 (2 September) is Meta's highest-scoring model on AA's index, 8.5 points above Muse Spark 1.2. It accepts text, image, and video input (Meta's model page), so you can hand it a screenshot or a screen recording of the bug. It writes quickly once it starts, but thinks first, so a full answer takes about 36 seconds.

It scores 3 points below Sol on coding and costs about $174 per 1,000 tasks against Sol's $109. It's proprietary, unlike the Llama models.

Best for: Frontend work from screenshots. Bugs that are easier to show than describe.

7. GLM-5.3: Budget Pick With a Top-Tier Score

• Coding Agent Index: 54
• Intelligence Index: 45
• Pricing: $1.40 / $4.40 per 1M tokens
• Context: 1M tokens

Z AI's GLM-5.3 (18 August) ties Muse Spark on the Coding Agent Index at a similar list price. At its everyday setting it answers in about 39 seconds and costs about $207 per 1,000 tasks. Pick it if you prefer Z AI's stack or need a second provider for fallback.

Best for: Budget coding agents. A fallback model at the same quality tier as Muse Spark.

8. Kimi K3: The Open-Weight Option

• Coding Agent Index: 52
• Intelligence Index: 44
• Pricing: $3.00 / $15.00 per 1M tokens
• Context: 1M tokens

K3 debuted at #1 on Arena's Frontend Code Arena in July. Two months later, GPT-6 Sol beats it on both AA indexes for less money through the API.

Open weights are the reason to pick it now. Through the API it's slow, at over a minute per answer. If you need to run a strong coding model on your own hardware, K3 is still one of the better choices. See our complete Kimi K3 breakdown for hardware and hosting.

Best for: Teams that need to self-host. Frontend-heavy work.

9. GPT-6 Luna: Budget Coding Powerhouse

• Coding Agent Index: 41
• Intelligence Index: 37
• Pricing: $0.10 / $0.50 per 1M tokens
• Context: 1.05M tokens

Luna is the cheapest model with a published Coding Agent score. It matches GPT-5.6 Luna on intelligence (37) at half the price, answers in about 23 seconds, and costs about $7 per 1,000 coding tasks.

41 won't solve hard bugs. For agent sub-tasks, simple edits, and code explanation, nothing with a coding score comes close on price.

Best for: Agent sub-tasks. High-volume coding workflows. Budget-conscious teams.

Other Coding Options Worth Knowing

Gemini 3.8 Flash ($0.75/$3.75)

Scores 42 on the Coding Agent Index and 41 on the Intelligence Index. It thinks a lot, so despite the low list price, 1,000 coding tasks cost about $148, more than GPT-6 Sol. Google's fourth Flash model in under four months, so expect a replacement soon.

Qwen3.8 Max ($2/$6)

Alibaba's September release scores 45 on intelligence but 43 on coding. Grok 4.7 costs the same and scores 56 on coding. Only worth it if you're already on Alibaba Cloud.

DeepSeek V4 Pro ($1.32/$3.96)

Scores 43 on the Coding Agent Index and 36 on intelligence. Muse Spark scores 11 points higher on coding for about the same money.

MiMo-V2.6-Pro ($0.43/$0.87)

Xiaomi's open-weight model scores 46 on the Intelligence Index and costs about $20 per 1,000 coding tasks, the most intelligence per dollar of anything here. At about a minute per answer it suits background agents. AA hasn't published a Coding Agent score for it, so test it on your own tasks first.

GLM-5.3 Flash ($0.15/$0.50)

Scores 42 on intelligence for about $26 per 1,000 coding tasks. No Coding Agent score yet, and slow at about a minute per answer.

DeepSeek V4.1 Flash ($0.30/$1.20)

Scores 39 on intelligence, answers in about 12 seconds and costs about $33 per 1,000 coding tasks. No Coding Agent score yet. Good for fast, simple edits.

Claude Sonnet 5 ($2/$10)

Scores 38 on the Intelligence Index with no published coding score. It lists at the same price as GPT-6 Sol but thinks so much that it costs more per task than Opus 5.5. If you want to stay with Anthropic, use Opus 5.5.

Models to Stop Using

• GPT-5.6 Sol, Terra, and Luna: GPT-6 Sol and Luna match or beat them at half the price.
• Claude Opus 5: Opus 5.5 scores 7 points higher and costs 20% less.
• Claude Fable 5: replaced by Fable 5.1 at the same price.

Best Free AI Model for Coding

DeepSeek R1 (Free via OpenRouter)

Available for free on OpenRouter. Its reasoning makes it passable for algorithmic problems. It scores 11 on the current Intelligence Index, and the 64K context limit and slow speed are real constraints.

For learning and experiments on non-critical tasks, you can't beat the price. For anything else, GPT-6 Luna at about $7 per 1,000 coding tasks is the better deal.

Cost Comparison: Real Numbers

Running 50 coding tasks per month, averaging 45,000 tokens per task, at each model's everyday setting with thinking tokens included:

Counting thinking changes the order. Grok 4.7, Qwen3.8 Max, Kimi K3 and Claude Sonnet 5 all look cheaper than Opus 5.5 on list price, and all cost more per task. GPT-6 Sol gives you a 57 on coding for about a third of Opus 5.5's bill. Our cost calculator runs these numbers for your own volume.

A tiered setup (Opus 5.5 for hard tasks, Luna for simple ones) cuts the bill further. Luna scores 41 on the Coding Agent Index to Opus 5.5's 66, so give it simple edits and lookups, and check its output on your own tasks before you route more to it. It costs about $7 per 1,000 coding tasks against Opus 5.5's $296.

The Coding Agent Index

The Artificial Analysis Coding Agent Index scores models running inside coding agent harnesses, such as Claude Code and Codex, on agentic coding. Version 1.5 is the equal-weight average of three benchmarks: DeepSWE v1.1, Terminal-Bench 4.0, and SWE-Atlas-QnA (AA methodology).

A model's raw intelligence score doesn't tell you how well it works as a coding agent. Tool integration, function calling, and token efficiency all change the result. Qwen3.8 Max is the clearest example: 45 on intelligence, 43 on coding. Grok 4.7 goes the other way, with 46 on intelligence and 56 on coding.

AA only publishes this index for some models. Where a model has no score, we don't guess one.

How to Configure Your Agent

A balanced September 2026 config for agent setups:

{
  agents: {
    defaults: {
      model: {
        primary: "anthropic/claude-opus-5.5",
        fallbacks: ["openrouter/openai/gpt-6-sol"]
      },
      subagents: {
        model: "openrouter/openai/gpt-6-luna",
        maxConcurrent: 2
      },
      thinkingDefault: "medium",
      maxConcurrent: 3
    }
  }
}

Opus 5.5 handles primary tasks at its default medium effort. Raise it to high or max for the hard ones. GPT-6 Sol is the fallback, and Luna handles sub-agent work. Using the task math above, 50 primary tasks a month on Opus 5.5 come to about $15 at high effort, thinking included. Luna sub-agents add cents.

Our Recommendation

For most developers in September 2026: Claude Opus 5.5 as your primary model, GPT-6 Luna as your sub-agent. Opus 5.5 is the best coding model available and the cheapest of the top three. Luna fills the budget slot at $0.10/$0.50.

If Opus 5.5 is too much for your volume, GPT-6 Sol scores 57 on coding for about a third of the cost per task. Grok 4.7 scores 56 and answers fastest, but costs more per task than Opus 5.5. For background agents, MiMo-V2.6-Pro runs about $20 per 1,000 tasks.

If you're on a tight budget, Luna handles simple coding work for under a cent per task.

FAQ

What is the best AI model for coding in 2026?

Claude Opus 5.5. It scores 66 on the Artificial Analysis Coding Agent Index and 58 on the Intelligence Index, #1 on both, and costs about $296 per 1,000 coding tasks at high effort. For value, GPT-6 Sol scores 57 on coding for about $109. For budget work, GPT-6 Luna scores 41 for about $7.

What is the cheapest AI model that's good for coding?

GPT-6 Luna at $0.10/$0.50 per million tokens. It scores 41 on the Coding Agent Index, and it's the cheapest model with a published score. Gemini 3.8 Flash scores 42 on coding but thinks a lot, so it costs far more per task. For open weights, MiMo-V2.6-Pro scores 46 on the Intelligence Index at $0.43/$0.87. For free, DeepSeek R1 is available on OpenRouter.

Is DeepSeek R1 good enough for coding?

For simple tasks, boilerplate, and learning, yes. It scores 11 on the current Intelligence Index and has a 64K context limit, so it struggles with production code and large codebases. GPT-6 Luna scores 37 for about $7 per 1,000 coding tasks.

What models does Claude Code use?

Claude Code runs Anthropic's Claude models. Claude Opus 5.5 is the strongest option, with a Coding Agent Index score of 66. Claude Sonnet 5 at $2/$10 is the cheaper one. You can configure different models for different task types.

How do I reduce my AI coding costs?

Use a tiered model setup. Route complex tasks to Claude Opus 5.5 or GPT-6 Sol and simple tasks to GPT-6 Luna or DeepSeek V4.1 Flash. Our Cost Calculator can estimate your savings.

Model data sourced from Artificial Analysis Intelligence Index v4.3.2 and Coding Agent Index v1.5. Costs are for each model's everyday reasoning setting, thinking tokens included. Updated September 26, 2026. For the live version, see best AI model for coding. For the overall model rankings beyond coding, see our Best LLM in 2026 guide and LLM Leaderboard. For the GPT-5.6 tiers that GPT-6 replaced, see our GPT-5.6 Sol vs Terra vs Luna guide. For general-purpose AI agents (automations, knowledge work, not coding), see Grok Bot vs Hermes vs OpenClaw. For cost optimization, see how to reduce AI API costs. Use our Model Selector for personalized recommendations based on your workload.