Updated August 2026. I first wrote this in September 2025, when the answer was "GPT-5 for front-end, Claude Sonnet 4 for back-end". A year is a geological era in this space, so I have rewritten it around the current model lineup. The advice on how to pick has aged fine; the model names did not.
I have been testing most of the frontier AI models for coding tasks since GPT-3, and it is still hard to name a single best one. Official benchmarks released by the companies are not very useful for picking a model for your use case - every lab now leads at least one leaderboard, which tells you more about the leaderboards than the models.
Even if you stick with one model, quality can drift: models get quantised or throttled during peak hours, and newer snapshots get swapped in behind an unchanged API version number.
What actually matters when picking
- The programming language. All frontier models are strong at TypeScript and Python; quality drops off unevenly for less common languages, and which model drops off least changes per language.
- Front-end versus back-end. This gap has narrowed since 2025, but it still exists - see the per-model notes below.
- Where you run it. A model that shines in a chat window can be mediocre inside an agentic tool. In 2025, GPT-5 kept looping inside Cursor while Sonnet was brilliant there; today the split is between agent harnesses - Claude models are at their best in Claude Code, OpenAI models in Codex.
- Thinking/reasoning mode. Worth it for debugging and multi-file changes; a waste of tokens (and money) for boilerplate.
- Context window. Gemini 3.1 Pro handles 1M tokens at the lowest frontier price, and Anthropic offers 1M-token context on the Sonnet line too. For whole-repo questions this matters; for day-to-day coding, agentic search has mostly replaced "paste the entire repo".
The models, as of August 2026
Claude (Anthropic)
The Claude 5 family - Sonnet 5 and Opus 5, plus the new top-tier Fable 5 - is my default for back-end and agentic coding. Fable 5 currently tops SWE-bench at around 95%, and for once the benchmark matches my experience: it finds logic bugs across files that other models miss. The catch is price; Sonnet 5 is the sensible daily driver, and most of my coding happens with it inside Claude Code, which has become the strongest agentic harness of the bunch.
GPT (OpenAI)
GPT-5.6 leads the terminal-oriented benchmarks (Terminal-Bench, Artificial Analysis coding index), and the Codex models remain excellent at front-end work - the "beautiful designs" reputation from GPT-5 still holds. If your workflow lives in Codex, you will be well served. Outside their own harness they are less consistent, same as Claude models are outside theirs.
Gemini (Google)
I was harsh on Gemini in the 2025 version of this post, and Google has fixed the biggest complaint: Gemini 3.1 Pro is actually available, at the lowest price per token of any frontier model, with a genuine 1M context window. It is the value pick and the large-codebase pick. My remaining gripe is unchanged - the Google Cloud console is still a maze compared to the OpenAI and Anthropic dashboards.
Grok (xAI)
Grok 4 remains good for researching bugs and general technical research - I still use it mostly in the chat app rather than the API. For writing production code it trails the three above.
What I actually run at GetSite
For generating websites in production, GetSite now runs on Claude Opus 5 by default, with Sonnet 5 as the faster option. GPT-5.2 handles the chatbot and other menial tasks, Gemini 3 Flash covers the cheap high-volume calls, and GPT Image 2 generates images. For my own coding, it is Sonnet 5 in Claude Code, with Fable 5 reserved for the gnarly multi-file problems.
That is four providers in one product, and it is deliberate: the leaderboard reshuffles every couple of months, so build your stack model-agnostic and re-evaluate quarterly. The best model for coding is a moving target - the best habit is keeping your switching costs low.