Start with what you want to get done. We'll match your task to the evidence, not to the biggest marketing budget.
Here are the options worth considering.
We've ranked these using the tests that are relevant to your task. Each test has limits, so treat this as a useful starting point, not a guarantee.
In short
For coding and building things, our pick is Claude Opus 5.5. Gemini 4 Argon and MiMo-V2.6-Pro are too close to call against it. If you would rather use a model with open weights, MiMo-V2.6-Pro is the best of those we can see. We have verified prices for Gemini 4 Argon only; for the rest we cannot say what they cost.
For Coding and building things, we looked at coding votes first.
Best-supported choice
Scored highest on coding votes among the models with enough evidence, and none of its other relevant results put it behind the leaders.
Worth knowing: Not enough data for fixing real software bugs, so this recommendation rests on fewer tests.
Price: Price not independently verified
How we know: scores and sources
See everything about Claude Opus 5.5
Also worth a look
Good-enough alternative Close call with our top pick
Rated Top tier on coding votes.
Worth knowing: Not enough data for fixing real software bugs, so this recommendation rests on fewer tests.
Price: - Google AI Pro: GBP 18.99 per month (price source, verified 28 Oct 2026)
- Google AI Ultra (from): GBP 79.99 per month (price source, verified 28 Oct 2026)
How we know: scores and sources
Full profile for Gemini 4 Argon
Good-enough alternative Open weights Close call with our top pick
Rated Top tier on coding votes.
Worth knowing: Not enough data for fixing real software bugs and difficult coding tasks, so this recommendation rests on fewer tests.
Price: Price not independently verified
How we know: scores and sources
Full profile for MiMo-V2.6-Pro
About cost. We can compare performance, but we don't yet have enough verified pricing data to tell you which option is cheapest. We never use API prices as a stand-in for consumer subscription prices, and we never assume a free tier has the same features, limits or model access as a paid plan. New to tokens and “per million”? Read how AI pricing works.
Other models with enough evidence (63)
These have a result for the main test but were not shown above. They rated Middle or Behind the leaders on it, or fell outside the top few. Listed in score order.
Models we could not rank for this task (11)
Missing evidence is not a low score. These models are left out of the ranking because the tests for this task have no result for them.
- amazon-nova-experimental-chat (26-02-10): Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- Dola Seed 2.0 Pro: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- ERNIE 5.0 (0110): Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- ERNIE 5.1: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- GPT-5.5 Instant: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- Grok 4.20 Beta 1: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- Grok 4.20 Multi-Agent Beta (0309): Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- Muse Spark: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4): Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- Qwen3.5-Max-Preview: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
- Qwen3.7-Max-Preview: Not enough data for Arena WebDev: building web apps, SWE-bench Verified (Epoch runs) and FrontierCode (reported by Cognition).
How this was decided: our recommendation rules. Results are relative to the 77 models listed here.
Your choice stays in your browser. Nothing you select here is sent anywhere.