Best LLMs for coding: developer favorites

[Last updated: July 2026]

The best LLMs that developers use for coding stand out by combining deep understanding of programming languages with practical capabilities that enhance a developer's workflow. They solve complex problems and deliver code that can be used to build production applications faster—not just vibe code a prototype.

At the frontier, Anthropic leads on software engineering. Claude Fable 5 tops SWE-Bench Pro at 80%, and Claude Opus 5 comes within 0.5% of it on CursorBench at half the cost per task. GPT-5.6 Sol owns the terminal at 88.8% on Terminal-Bench 2.1.

Open-weight models are closing the gap. Kimi K3 beat Fable 5 on frontend code and swept six of seven frontend categories. GLM 5.2 comes within a point of Claude Opus 4.8 on tool orchestration under a plain MIT license. DeepSeek V4-Flash runs a million tokens of context at $0.14 per million input tokens.

Most of these models are available in developer tools like Claude Code, Cursor, Codex, and Copilot. Most developers keep a favorite for daily work and switch models for specific tasks.

Here are the LLMs developers use most for coding.

Developers’ favorite LLMs for coding

Most experienced developers run several models through platforms like Cursor, Claude Code, or Codex CLI, matching the model to the task rather than standardizing on one.

• Daily production coding: Claude Opus 5 for the balance of capability and cost, Claude Sonnet 5 for higher volume
• Hardest long-horizon work: Claude Fable 5, where the extra spend is justified by the task
• Terminal and shell-heavy work: GPT-5.6 Sol, which leads Terminal-Bench 2.1
• Front-end development: Kimi K3, which swept six of seven frontend categories, or GPT-5.6 Sol
• Parallel agentic tasks: GPT-5.6 Sol with ultra, or Claude Opus 5 for long-running agents
• Agentic loops and multimodal: Gemini 3.6 Flash
• High-volume, cost-sensitive work: DeepSeek V4-Flash or Mistral Small 4
• Self-hosted or air-gapped: GLM 5.2 on MIT, or Nemotron 3 if you need published training data
• Data residency requirements: Mistral Large 3 or Small 4 under Apache 2.0
• Sub-agent orchestration: Claude Haiku 4.5 at $1 and $5

The Pragmatic Engineer's March 2026 survey of 906 software engineers found 46% named Claude Code as the tool they love most, nearly 2.5x the share of Cursor at 19% and more than 5x GitHub Copilot at 9%.

Claude Code went from well-liked underdog to the most-used AI coding tool in eight months from its May 2025 launch. That survey is a few months old now, and the model names inside it have all been replaced, but the tool preference has held.

There may be a new coding LLM leading the pack by the time you read this. These are the consistent favorites.

Anthropic Claude coding LLMs

Anthropic runs four tiers, with a Mythos-class tier sitting above Opus. Fable 5 leads the benchmarks, Opus 5 is the value pick and the default on Claude Max, Sonnet 5 handles high-throughput work, and Mythos 5 exists but you probably can't get access to it.  

Opus 5 is the LLM most teams should run for coding tasks. It's priced identically to Opus 4.8 and comes close to Fable 5's capability at half the cost per task. It's also new state of the art on Frontier-Bench v0.1, more than doubling Opus 4.8, and posts roughly three times the next-best model on ARC-AGI-3.

A few practical notes on these models:

• Fable 5 refuses some work. Requests involving cybersecurity, biology, chemistry, or model distillation route to Opus 4.8 instead of answering. Anthropic says that triggers in under 5% of sessions, but if your work sits in those domains your rate will be higher.
• Mythos-class models require 30-day data retention for business traffic. Opus 5 and below have no retention requirement.
• Sonnet 5's introductory pricing expires August 31, 2026, after which it moves to $3 and $15.
• Knowledge cutoffs differ. Opus 5 is trained through May 2026, Fable 5 and Sonnet 5 through January 2026.
• Fast mode runs Opus 5 at about 2.5x default speed for twice the base price.

OpenAI GPT LLMs for coding

The GPT-5.6 family replaced OpenAI's entire lineup with three capability tiers instead of a flagship plus mini variants. Sol is the frontier model, Terra handles everyday work, and Luna covers high volume. All three run the same context window, so the difference is capability and price.  

Terra is the value pick here, not Sol. It lands within a couple of points of Sol on both benchmarks at half the price. Sol earns its premium on the hardest work, where it takes the Artificial Analysis Coding Agent Index at 80, a new state of the art, and where the ultra setting runs four agents in parallel to push Terminal-Bench from 88.8% to 91.9%.

Two features change how you'd use these in practice. Programmatic Tool Calling lets the model write in-memory programs that filter tool output instead of routing everything back through context, which Clio measured at 38% fewer prompt tokens on multi-step document analysis with no quality loss. And Qodo found GPT-5.6 beat GPT-5.5 on their agentic code-review F1 score using roughly 3x fewer tokens per pull request at about half the median latency.

A few practical notes before you pick one:

• Luna is weak on long context. It scores 41.3% on MRCR v2 recall against Sol's 73.8%. The window is the same size; the reliability isn't. Verify it before pointing it at a large codebase.
• Cache billing changed with this generation. Cache writes now bill at 1.25x the uncached input rate while reads keep the 90% discount, with a 30-minute minimum cache life.
• ultra trades tokens for results. Four parallel agents by default, so budget accordingly. Developers can build the equivalent through the multi-agent beta in the Responses API.
• OpenAI hasn't published context limits or knowledge cutoffs for this family. The 1M and 128K figures in the table are inferred from their own long-context benchmarks, which run to 1M tokens.

Google LLMs for coding 

Google's stable coding lineup is built entirely on Flash. Gemini 3.1 Pro is still labeled Preview in Google's own docs, and Gemini 3 Pro Preview has been shut down. If you need a production commitment from Google today, the Flash line is where it lives.  

Neither stable Flash model has an independent coding benchmark, and the only Gemini with published SWE-Bench Pro and Terminal-Bench numbers is the one Google still calls Preview. Those numbers also trail Anthropic and OpenAI by a wide margin.

Gemini 3.6 Flash is the one to use if you're picking a Gemini for coding. It's the only one that supports computer use, Google positions it specifically for rapid coding cycles and iterations, and it has a lower output price than the older 3.5 Flash despite being more capable.

A few practical notes:

• Flash stopped being the budget option. $1.50 and $7.50 is real money against where the Flash tier started. Google folded frontier capability into the Flash name instead of maintaining a separate cheap tier.
• Batch and Flex cut input in half, to $0.75, which matters for high-volume code analysis that isn't latency-sensitive.
• Preview status carries real risk. Tighter rate limits and as little as two weeks of deprecation notice. Gemini 3 Pro Preview was shut down, which is what that risk looks like in practice.
• Gemini 3.5 Flash doesn't support computer use. That's the main functional gap versus 3.6 Flash.

Favorite open-weights LLMs for coding

Every model here ships downloadable weights, and two of them now beat frontier models on specific tasks. Read the licenses before you commit to one, because "open" covers very different terms across this list.  

GLM 5.2 is the one to reach for first. It's the strongest open-weight coding model available, it runs a million tokens of context, and it ships under a plain MIT license with no revenue cap and no acceptable-use policy attached. Hugging Face published a direct comparison putting it at 81.0% on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0%, and within a single point of Opus 4.8 on tool orchestration. Nvidia publishes an official NVFP4 quantization, and it deploys through Dell Enterprise Hub, Microsoft Foundry, or AWS SageMaker.

Kimi K3 is the frontend specialist. It beat Claude Fable 5 on frontend code and swept six of seven frontend categories at $3 and $15 against Fable 5's $10 and $50. It also has native vision, which is what makes it good at turning designs and screen recordings into working interfaces.

A few things to check before you commit to any of these:

• Kimi K3's license is not plain MIT. Model-as-a-Service operators above $20 million over any consecutive 12 months need a separate agreement with Moonshot, and products above 100 million monthly active users or $20 million monthly revenue must display "Kimi K3" in the interface. Internal-only use is exempt.
• Mistral Medium 3.5 is off the table for most enterprises. Its Modified MIT license only authorizes use if your global consolidated revenue is at or below $20 million for the preceding month. The test runs monthly and follows third-party fine-tunes.
• Inkling's Apache 2.0 comes with a separate use policy that prohibits automated decisions affecting financial, legal, employment, healthcare, and housing rights, and Thinking Machines can update it unilaterally. Thinking Machines also says plainly that Inkling is not the strongest model available today, open or closed.
• Mistral deprecated Devstral 2 roughly seven months after launch, along with the Magistral and Pixtral lines. An open license means you keep the weights you downloaded, not that anyone keeps maintaining them.
• Qwen 3.8 is not an open-weight model. Alibaba has published no weights, no license, and no repository for it, and the only shipped variant is closed and subscriber-gated. Qwen 3.6 is the newest release with published weights. Treat any Qwen 3.8 spec you read elsewhere as unverified.
• Self-hosting Inkling is expensive. Running it in BF16 needs roughly 2TB of aggregate VRAM, or eight B300s, dropping to about 600GB quantized to NVFP4.

One thing to watch in that table: three of the benchmark cells cite SWE-bench Verified and one cites SWE-bench Pro, which is why I labeled each one and added the note underneath. Mixing those two unlabeled is the exact error we pulled out of the live post, and it would be easy to reintroduce here.

Why teams keep an open-weights model ready for cybersecurity

On July 16, 2026, Hugging Face disclosed a security incident driven end to end by an autonomous AI agent system, running tens of thousands of automated actions at machine speed. Reconstructing it meant building a timeline from more than 17,000 recorded events.

Their first attempt used frontier models behind commercial APIs. It didn't work. Forensic analysis means submitting real attack commands, exploit payloads, and command-and-control artifacts, and the providers' safety guardrails blocked the requests. As Hugging Face put it, those guardrails "cannot tell an incident responder apart from an attacker."

They ran the investigation on self-hosted GLM 5.2 instead. It worked, and no attacker data or credentials left their environment.

The lesson generalizes past security. Any workflow where a model refusal is expensive, whether that's security tooling, penetration testing, or handling code you can't send to a third party, is a workflow worth having a self-hosted fallback for. Vet it and keep it warm before you need it, not during an incident.

Which LLM should my team use for coding?

Different models excel in different areas:

• Production development at scale: Claude Opus 5
• Highest capability regardless of cost: Claude Fable 5
• Architecture and large refactors: Claude Opus 5
• Terminal, shell, and CLI work: GPT-5.6 Sol
• Front-end and UI quality: Kimi K3 or GPT-5.6 Sol
• Parallel multi-agent pipelines: GPT-5.6 Sol with ultra
• Multimodal and agentic loops: Gemini 3.6 Flash
• High-volume or cost-sensitive workloads: DeepSeek V4-Flash or Mistral Small 4
• Self-hosted with no license restrictions: GLM 5.2
• Regulated industries and data residency: Mistral Large 3 or Small 4, self-hosted

Most experienced developers take a multi-model approach, using specialized strengths for specific tasks through Claude Code, Cursor, Codex, and Copilot.

Need help choosing which coding LLMs to use at your company?

We can help you decide which coding LLMs are best for your teams, tech stack, and goals. Our senior engineering teams already use AI coding tools to build production-ready software.

And if you need to hire an AI-augmented team of senior software engineers to speed up delivery, we can assemble them for you in 4-6 weeks. It’ll be faster to get started, more cost-efficient than internal hiring, and we’ll deliver high-quality results quickly.

Zappos, Twilio, and Veho are just a few companies that trust us to build their software and enterprise systems.

You can schedule a time to talk with us here. No hassle, no expectations, just answers.