Which LLM for coding and software development in May 2026?

The short answer

For coding workloads right now, use GPT-5.3-Codex from OpenAI. It's the only model in the current lineup purpose-built for code generation, and at $4.81/M tokens with 76 tok/s output speed, it sits at a reasonable price-performance point for most development teams. If budget matters more than peak quality, Kimi K2.6 at $1.42/M tokens delivers 53.9 quality for a fraction of the cost.

The decision gets more nuanced depending on whether you're running batch code reviews, powering an IDE copilot, or generating boilerplate at scale. I'll break down each scenario below.

Why GPT-5.3-Codex is the default pick

GPT-5.3-Codex (OpenAI) scores 53.6 on the quality index at $4.81/M input tokens and outputs at 76 tok/s. Those numbers tell a specific story: it's not the highest-quality model available, but it's explicitly tuned for code. General-purpose models like GPT-5.5 score higher overall (60.2) but cost $11.25/M tokens, and that quality premium reflects broad capability, not necessarily better function signatures or tighter diffs.

For code-heavy pipelines where structured output matters, a model trained on code distributions will produce fewer parser failures and more syntactically correct completions per attempt. Fewer retries means lower effective cost, even if the sticker price is higher than budget alternatives.

When to pick something else

Not every coding task needs a code-specialist model. Here's where I'd deviate.

Fast iteration loops in an IDE

If you're building a copilot-style integration where inference latency directly affects developer experience, Gemini 3.5 Flash at 219 tok/s is nearly three times faster than GPT-5.3-Codex. It scores 55.3 on quality and costs $3.38/M tokens. For inline completions and short suggestions where the model generates 50-200 tokens at a time, the speed difference is the difference between fluid and sluggish.

Budget batch processing

Running large-scale code migrations, automated refactoring, or test generation across thousands of files? Kimi K2.6 (MoonshotAI) at $1.42/M tokens is the cheapest model with a quality score above 53. At scale, the cost gap between $1.42 and $4.81 compounds fast. On a 10B-token monthly workload, that's $33,900 saved. Kimi K2.6 is also open source, which matters if you need to self-host for compliance.

Maximum quality, cost no object

GPT-5.5 at 60.2 quality is the strongest model available. At $11.25/M tokens and 65 tok/s, it's expensive and not especially fast, but for complex architectural reasoning or multi-file refactors where correctness on the first pass saves engineering hours, the premium can pay for itself.

Comparison table

Decision table

The open-source angle

Two open-source models deserve attention. Kimi K2.6 (MoonshotAI) is the better-documented option with published pricing and speed. Qwen3.7 Max (Alibaba) scores higher at 56.6 quality and costs $3.75/M tokens via API, but has no published throughput figure, which makes capacity planning harder if you're self-hosting. If you can benchmark Qwen3.7 Max on your own hardware and confirm acceptable latency, it's the stronger open-source choice on quality alone.

What I'd actually deploy

For a team shipping a coding assistant product today, I'd route traffic by task type: Gemini 3.5 Flash for inline completions (speed matters, tokens are short), GPT-5.3-Codex for multi-turn code generation and review (code tuning matters, context is longer), and Kimi K2.6 for any offline batch work. This routing strategy keeps per-request costs low where quality differences are imperceptible and reserves the specialized model for tasks where it earns its price.

Use the LLM Selector to filter by coding performance and budget, or browse current rankings on Explore to see how these models stack up as new benchmarks arrive.