AI Model Power Rankings

Still #1 on Arena WebDev and well ahead on SWE-Bench Pro. Argon took the text board, but Opus 5.5 is the best model you can actually use today.

Scores 56 on Artificial Analysis's index, above Fable 5.1 and GPT-6 Astra, at $2 / $10. For most people it is the new sensible default.

#1 on Arena's text board and the lowest hallucination rate tested. It would rank higher if you could use it; for now, almost nobody can.

Replaces GPT-6 Sol a week after launch. OpenAI says it nearly matches Astra on coding for a fifth of the price, and Arena's WebDev voters put it in the top five.

Powers OpenAI's new Dots agents, but its 6.1 upgrade was shelved and the cheaper Sol now matches it on most coding tests.

Still excellent, but Sonnet 5.5 beats it on the independent index at a fifth of the price.

Still a top-tier fast model at 75 cents in. The price doubles on January 1.

Good conversationalist at a fair price. Three new $2 / $10 models above it make it look less special.

Still the strongest model from outside the US frontier labs, but this week's launches widened the gap.

The best-proven open-weight model. A Flash version is reportedly out, but we haven't seen independent results.

Good value, and xAI's agent products are growing. The model itself is a step behind the frontier.

Very cheap and fast. We still want independent coding numbers before moving it up.

Holds #1 on Arena WebDev and leads SWE-Bench Pro clearly over Sonnet 5.5. Still the default for serious coding.

Arena WebDev top four on debut, and it beats Opus 5.5 on Terminal-Bench 4.0 in both Anthropic's and Artificial Analysis's runs. Results vary by harness, though.

The new default in Codex, with an Ultrafast option. OpenAI says it matches Astra on DeepSWE at about a fifth of the cost per task.

Still strong at driving a terminal, but Sol 6.1 now does most of the same work for much less.

Excellent code, but two cheaper Anthropic models are now at least as good for most jobs.

Google claims 77.9% on DeepSWE, but it's only #8 on Arena WebDev, and you can't use it yet.

Still a popular pick for cheap coding agents.

The open-weight coding pick, level with MiMo-V2.6-Pro on WebDev.

Fine for quick edits and autocomplete-style work. Not the one to run an agent on.

Top-tier quality at Flash prices, until January.

Near-Astra coding at $2 / $10, and cached input is now just $0.10 per million tokens.

Same price as Sol, higher on the independent index, and Anthropic says it uses fewer tokens per task than Sonnet 5.

$0.10 / $0.50. Still the cheapest serious option, though most of the evidence is OpenAI's.

Strong conversation at $1.25 / $4.25.

$0.15 / $0.47 with good multilingual skills. Qwen 4 is due 'very soon', according to Alibaba.

Very fast diffusion text generation at $0.20 / $0.75.

Its $6 output tokens are still among the cheapest at this quality level.

Still the strongest open-weight model on the most evidence. Read its custom license.

Level with GLM-5.3 on Arena text and WebDev as its vote count grows.

Apache 2.0 and small enough for real-world hardware. DeepSeek V4.1 Pro and Qwen 4 are both announced but not out.

The Sunburst version still leads the text-to-image board. No serious challenger this week.

Still the best non-OpenAI option.

Holding fifth. The indie favourite for aesthetics.

Steady in the top six on its low setting.

Not the best at raw generation, but still the best for editing an image through conversation.

Powers the CapCut/Dreamina ecosystem, level with Alibaba's Qwen-Image 3.0 Pro.

Google still holds the top two text-to-video spots with Omni Flash and its 1.1 update.

The safest all-rounder, with native audio and 4K.

Holds #3, still on a small vote count.

The value king at about 11–14 cents per second, plus storyboard mode.

Holding #4 in agent mode, close behind Flux.

Still the best realistic human motion for creators.

Outside the top five, but still the tinkerer's pick.

Professionals still choose it for camera control.