AI Model Power Rankings

No longer top of either scoreboard: 58 on the v5 index and #3 on Arena text. But it's still #1 on Arena WebDev, and unlike Carbon you can actually use it.

New. #1 on Arena text and #1 on the v5 index, one point clear of Opus 5.5. It would be our #1 if you could buy it. For now, only Google's selected cyber defenders can.

Up four places. 57 on the v5 index, its best showing against Anthropic yet. Still #2 on Arena WebDev.

55 on the v5 index puts it ahead of Sonnet 5.5 at the same $2 / $10. Still the best value in OpenAI's lineup.

Slips from 56 to 54 under v5 and now trails Sol. Still top four on Arena WebDev at half Opus 5.5's price.

#2 on Arena text at 1520, but 50 on the v5 index puts it level with Muse Wonder. Budget for $4 / $20 from December 1.

#5 on Arena text. Still a specialist pick at 2.5x Opus 5.5's price.

#4 on Arena text at 1507, but one of the big losers under v5: Wonder Full fell from 56 to 50.

New. 46 on the v5 index at $2 / $10, the best of any model outside the big US labs. It uses lots of tokens, so real costs run above the list price.

Still the cheapest strong small model at $0.10 / $0.50. We haven't seen a confirmed v5 score yet.

New. A $1 / $5 model built for computer use. OpenAI's 'matches Astra' claim isn't backed by independent scores, which put it below Sol.

Dropped from 45 to 32 under v5. The model didn't change, but the gap to the closed frontier looks much wider than it did a week ago.

Still #1 on Arena WebDev, around 1810. Carbon debuted at #4, so the coding crown stays with Anthropic.

Level with Astra in the WebDev top four at a fifth of Astra's price. Still the pick for well-scoped fixes.

Still the Codex default and the most coding per dollar from OpenAI.

Still #2 on WebDev and gaining on the v5 index, but Sol does most of the same work for a fifth of the price.

New. #4 on WebDev and strong on Google's security-coding tests, but those ran without guardrails and nobody outside Fairwind can check them yet.

Excellent code, but two cheaper Anthropic models do most jobs as well.

Outside the WebDev top five, and now overshadowed by Carbon.

New. #9 on WebDev and a clear step up on Kimi K3, but heavy context use makes long agent runs expensive.

Still mid-table on WebDev at Opus 5.5's price.

New in Codex. Aimed at computer use and office work rather than heavy coding. Use Sol for that.

Still a solid open-weight coding model, now joint top of the released open models on v5 with DeepSeek.

Plenty for cheap sub-agents and simple fixes.

$0.10 / $0.50 under 100K tokens. Nothing new this week beat it on price for the quality.

New. $1 / $5 for a model that sits between Luna and Sol. Artificial Analysis found its low-effort setting very cheap per task.

55 on the v5 index at $2 / $10, with cached input at $0.10. The strongest model at that price on the new index.

Down to 32 on v5, but still very cheap: $0.30 / $1.20 at peak and half that off-peak.

$2 / $10 for one more month, then $4 / $20. At 50 on v5 it now trails Sol at the same price.

54 on v5 at $2 / $10, but it uses lots of tokens, which eats into the savings.

New. $1.36 / $4.18 now the launch sale is over, with downloadable weights if you'd rather host it yourself.

Cheap at $0.15 / $0.60, but its high hallucination rate means you should check every answer.

Still good, still expensive per task next to the Chinese Flash models, and its price doubles in January.

Same price as Haiku 5.5 with weaker independent results. Terra is now the better OpenAI pick if you can pay ten times more.

Strong conversation at $1.25 / $4.25, though Terra now competes at a similar price.

$0.15 / $0.47 with good multilingual skills.

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

Joint top of the released open models on v5 at 32, tied with GLM-5.3. The MIT license and lower hosted price give it the edge.

Tied with DeepSeek at 32 on v5, down from 45. It still has the deeper Arena vote record.

Another quiet week. Level with GLM-5.3 on Arena, on fewer votes.

New. The weights are out, about a trillion parameters in FP8 and NVFP4. Check the license: hosted-service providers need a separate deal with Mistral.

Still the strongest US open-weight model. Mistral's arrival makes it a three-way choice for teams avoiding Chinese models.

Apache 2.0 and published recipes, but still well behind the leaders and with no hosted API.

Still the one on this list that runs on hardware you might own. See our guide to running AI on your own computer.

Still #1 on Arena text-to-image. Arena's free direct-mode access to the Sunburst version ends November 2.

Still #2 on text-to-image. Microsoft previewed MAI Image 2.7, but so far it only edits images, and text-to-image is 'coming soon'.

Still top five at about half Nano Banana 2's price.

Still the indie favourite for aesthetics.

Still top-tier on its low setting. xAI retires its older image-quality API model on November 2.

Powers the CapCut/Dreamina creator ecosystem.

Another quiet week for video. Google holds the top of the board.

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.

Top five in agent mode. Fast and fun.

Still the best realistic human motion for creators.

Still the tinkerer's pick for custom pipelines.

Professionals still choose it for camera control.