Week 42, 2026Published October 8, 2026
AI Model Power Rankings
Haiku 5.5 turns the $0.10 tier into a real fight. A week for cheaper models, not bigger ones. Anthropic's Claude Haiku 5.5 arrived at exactly GPT-6 Luna's $0.10 / $0.50, the same day OpenAI made Luna the default for free ChatGPT users. Google halved its image prices with Nano Banana 2.1, and Mistral put a trillion-parameter model up for testing. Opus 5.5 keeps #1, and Gemini 4 Argon is still out of reach for nearly everyone.
Debuts at #8 overall and #1 for value. Artificial Analysis gives it 43 on its index at max effort, up from 17 for Haiku 4.5, at a quarter of the old price.
Drops out of the overall top 12. Haiku 5.5 now costs exactly the same, and Anthropic's table has it far ahead on Terminal-Bench 4.0 (39.2% to 16.4%). Being ChatGPT's free default doesn't make Luna the best $0.10 model.
01Power Rankings
The best general-purpose AI models right now, with quality, price and momentum rolled into one score.
| # | Movement± | Model | Price / 1M tokens | Score |
|---|---|---|---|---|
| 1 | — | Claude Opus 5.5Anthropic Still #1 on Arena WebDev and still the model to beat on hard coding work. Argon has the text board, but you can't buy Argon. |
$4 in / $20 out | 97 |
| 2 | — | Claude Sonnet 5.5Anthropic Its second week confirms the debut: top four on WebDev at half Opus 5.5's price. Watch your bills, though, because Artificial Analysis found it unusually token-hungry. |
$2 in / $10 out | 95 |
| 3 | — | Gemini 4 ArgonGoogle Still #1 with Arena text voters, and still unavailable outside Google's security partners. It moves up the day Google opens the API. |
$2 in / $10 out | 94 |
| 4 | — | GPT-6.1 SolOpenAI #5 in Arena's agent rankings and the default in Codex. OpenAI's best value case near the frontier. |
$2 in / $10 out | 93 |
| 5 | ▲1 | Claude Fable 5.1Anthropic Still superb on the hardest problems. Anthropic's own cheaper models now match it on most work. |
$10 in / $50 out | 91 |
| 6 | ▼1 | GPT-6 AstraOpenAI Still #2 on WebDev, but with 6.1 shelved, OpenAI's own $2 Sol is the one it's pushing. |
$10 in / $50 out | 90 |
| 7 | — | Gemini 3.8 FlashGoogle A text top-10 model at Flash prices. Its price doubles on January 1, and Haiku 5.5 now undercuts it. |
$0.75 in / $3.75 out | 89 |
| 8 | NEW | Claude Haiku 5.5Anthropic Scores 43 on the Artificial Analysis index at max effort, at $0.10 / $0.50. The small-model story of the month. |
$0.1 in / $0.5 out | 88 |
| 9 | ▼1 | Muse Spark 1.3Meta Pleasant to talk to, but slipping as the 5.5 and 6.1 releases pile up above it. |
$1.25 in / $4.25 out | 86 |
| 10 | ▼1 | Kimi K3Moonshot AI Still the strongest model from outside the US frontier labs. A quiet week. |
Not published | 85 |
| 11 | ▼1 | GLM-5.3Z.ai · open Still the best-proven open-weight model. Mistral Large 4's weights, due by the end of October, are its next real challenge. |
Not published | 84 |
| 12 | ▼1 | Grok 4.7xAI No new model from xAI, and the field keeps moving up around it. |
$2 in / $6 out | 82 |
02Coding
For developers and 'vibe coders': writing, fixing and shipping real software.
| # | Movement± | Model | Price / 1M tokens | Score |
|---|---|---|---|---|
| 1 | — | Claude Opus 5.5Anthropic Still #1 on Arena WebDev, ahead of GPT-6 Astra. The default pick for open-ended coding. |
$4 in / $20 out | 98 |
| 2 | — | Claude Sonnet 5.5Anthropic Third on WebDev and ahead of Opus 5.5 on Terminal-Bench 4.0 by Anthropic's numbers. The pick for well-scoped fixes at half the price. |
$2 in / $10 out | 94 |
| 3 | — | GPT-6.1 SolOpenAI The default in Codex. Most of Astra's coding for a fifth of the price. |
$2 in / $10 out | 93 |
| 4 | — | GPT-6 AstraOpenAI #2 on WebDev and #1 on Image-to-WebDev. Still strong when the job means driving a terminal for hours. |
$10 in / $50 out | 92 |
| 5 | — | Claude Fable 5.1Anthropic #2 on Image-to-WebDev. Excellent code, but two cheaper Anthropic models do most jobs as well. |
$10 in / $50 out | 90 |
| 6 | — | Gemini 4 ArgonGoogle Google's own table puts it behind Opus 5.5 on Terminal-Bench 4.0, and it sits outside the WebDev top five. Coding isn't where it wins. |
$2 in / $10 out | 88 |
| 7 | — | Kimi K3Moonshot AI Still a popular pick for cheap coding agents. |
Not published | 84 |
| 8 | — | GLM-5.3Z.ai · open The open-weight coding pick, level with MiMo-V2.6-Pro on WebDev. |
Not published | 83 |
| 9 | NEW | Claude Haiku 5.5Anthropic Anthropic says 39.2% on Terminal-Bench 4.0. Independent runs land nearer 33–35%, which is still plenty for cheap sub-agents. |
$0.1 in / $0.5 out | 81 |
03Best Value
The most capability per dollar. If you're building an app or watching a budget, start here.
| # | Movement± | Model | Price / 1M tokens | Score |
|---|---|---|---|---|
| 1 | NEW | Claude Haiku 5.5Anthropic $0.10 / $0.50 under 100K tokens, and Artificial Analysis ranks it second of 182 models in its price class. Mind the 5x jump above 100K. |
$0.1 in / $0.5 out | 95 |
| 2 | ▼1 | Gemini 3.8 FlashGoogle Top-10 text quality at Flash prices until January. Then it doubles. |
$0.75 in / $3.75 out | 94 |
| 3 | ▼1 | GPT-6.1 SolOpenAI Near-Astra coding at $2 / $10, with cached input at $0.10. |
$2 in / $10 out | 93 |
| 4 | ▼1 | Claude Sonnet 5.5Anthropic Near-Opus coding at $2 / $10, but its heavy token use eats into the savings. |
$2 in / $10 out | 91 |
| 5 | ▼1 | GPT-6 LunaOpenAI Same $0.10 / $0.50 as Haiku 5.5, but the independent evidence favours Anthropic's model. Its new Decisions API is worth a look for extraction jobs. |
$0.1 in / $0.5 out | 89 |
| 6 | ▼1 | Muse Spark 1.3Meta Strong conversation at $1.25 / $4.25. |
$1.25 in / $4.25 out | 87 |
| 7 | ▼1 | Qwen3.8-FlashAlibaba $0.15 / $0.47 with good multilingual skills. Qwen 4 is still 'very soon'. |
$0.15 in / $0.47 out | 85 |
| 8 | ▼1 | Mercury 2.5Inception Very fast diffusion text generation at $0.20 / $0.75. |
$0.2 in / $0.75 out | 79 |
04Open-Weight Watch
Models you can download, run privately and customise.
| # | Movement± | Model | Price / 1M tokens | Score |
|---|---|---|---|---|
| 1 | — | GLM-5.3Z.ai · open Still the strongest open-weight model on the most evidence. Read its custom license before you build on it. |
Not published | 91 |
| 2 | — | MiMo-V2.6-ProXiaomi · open Level with GLM-5.3 on Arena, on fewer votes. Mistral Large 4's weights later this month are the next test for both. |
Not published | 88 |
| 3 | — | Ternary Bonsai 2 27BPrismML · open Apache 2.0 and small enough for real-world hardware. |
Not published | 78 |
05Image Generation
Making pictures from text, based on LMArena text-to-image votes plus our hands-on notes.
| # | Movement± | Model | Price / 1M tokens | Score |
|---|---|---|---|---|
| 1 | — | GPT Image 2.5OpenAI Still leads the text-to-image board, and Nano Banana 2.1 didn't close the gap. |
— | 97 |
| 2 | — | MAI-Image 2.6Microsoft AI Microsoft's in-house model stays just ahead of Google's new release. |
— | 88 |
| 3 | NEW | Nano Banana 2.1Google Debuts at #5 on text-to-image and #4 for multi-image editing, at about half Nano Banana 2's price. The best pick for editing in chat. |
— | 87 |
| 4 | ▼1 | Reve 2.1Reve Pushed down a place by Google's newcomer. Still the indie favourite for aesthetics. |
— | 85 |
| 5 | ▼1 | Grok Imagine Image 2.0xAI Still top-tier on its low setting. |
— | 83 |
| 6 | — | Seedream 5.0 ProByteDance Powers the CapCut/Dreamina creator ecosystem. |
— | 80 |
06Video Generation
Text-to-video, and what to use now that Sora is gone.
| # | Movement± | Model | Price / 1M tokens | Score |
|---|---|---|---|---|
| 1 | — | Gemini Omni 1.1 FlashGoogle Google still holds the top two text-to-video spots with Omni Flash and its 1.1 update. |
— | 95 |
| 2 | — | Veo 3.1Google The safest all-rounder, with native audio and 4K. |
— | 92 |
| 3 | — | FLUX 3 VideoBlack Forest Labs Holds #3, still on a small vote count. |
— | 90 |
| 4 | — | Kling 3.0Kuaishou The value king at about 11–14 cents per second, plus storyboard mode. |
— | 89 |
| 5 | — | Grok Imagine Video 1.5xAI Top five in agent mode. Fast and fun. |
— | 87 |
| 6 | — | Seedance 2.5ByteDance Near the top of Artificial Analysis's video board. Still the best realistic human motion for creators. |
— | 86 |
| 7 | — | Wan 3.0Alibaba Leads Artificial Analysis's text-to-video board, even though Arena voters rank it lower. Still the tinkerer's pick. |
— | 84 |
| 8 | — | Runway Gen-4.5Runway Professionals still choose it for camera control. |
— | 82 |
The Wire · what happened this week
Mistral opens a public API preview of Mistral Large 4, a mixture-of-experts model of about a trillion parameters. The open weights are promised by the end of October, but the license isn't published yet.
Scores are our editorial blend of public leaderboards (LMArena, Artificial Analysis and published benchmarks), price and hands-on use. Read the methodology →