AI value: Haiku 5.5 wins per dollar, Sonnet 5.5 at the top

AI model prices and capabilities keep shifting in autumn 2026, so we put two simple questions to the latest data. First: which model gives you the most intelligence per dollar? Second: if you want one of the best, which of the five smartest models is the cheapest?

The short answer: on intelligence per dollar, Claude Haiku 5.5 wins (217 index points per dollar). Among the frontier models, Claude Sonnet 5.5 is the cheapest (56.0 points, $4.00 blended per 1M tokens).

How we calculated

We measure intelligence with the Artificial Analysis Intelligence Index v4.3.2. It is an independent composite of ten evaluations, including Humanity's Last Exam, Terminal-Bench 4.0 and GDPval-AA. For each model we took its highest-scoring setting, usually max reasoning.

Prices come from the vendors' official API pricing pages (Anthropic, OpenAI, Google, xAI, DeepSeek, Z.ai). Prices for Meta, Xiaomi, Alibaba, StepFun, Moonshot and Mistral are taken from the Artificial Analysis listing. Blended price = 0.75 × input price + 0.25 × output price per 1M tokens, i.e. a 3:1 traffic mix. Intelligence per dollar is index points divided by blended price. We covered 24 current models from 12 vendors, with data retrieved on October 18, 2026.

Source: Artificial Analysis, vendor API pricing pages. Chart: Neural Daily.

The red dashed line is the Pareto frontier. For every model off it, another model is smarter and no more expensive. The six on it: MiMo-V2.6-Flash, Claude Haiku 5.5, MiMo-V2.6-Pro, Muse Spark 1.3, Claude Sonnet 5.5, Claude Opus 5.5.

Category 1: most intelligence per dollar

Claude Haiku 5.5 in max reasoning mode scores 43.4 on the index. It costs $0.10 per 1M input tokens and $0.50 per 1M output tokens, a blended $0.20. That works out to 217.0 points per dollar.

It is a close race. Xiaomi's open-weight MiMo-V2.6-Flash is effectively tied at 216.6 because it is even cheaper ($0.175), but it scores 5.5 points lower (37.9). GPT-6 Luna is third (190.5), followed by GLM-5.3-Flash (176.0) and Qwen3.8-Flash-Next (173.0).

Source: Artificial Analysis, vendor API pricing pages. Chart: Neural Daily.

One important footnote: Haiku 5.5's low price only applies to prompts up to 100K tokens. Above that the rate is $0.50 input and $2.50 output, five times higher. For long-context work its score per dollar drops to about 43, and it falls out of the lead.

Category 2: the cheapest of the five smartest

The ranking again uses the Artificial Analysis Intelligence Index v4.3.2, with each model's best setting. The five highest-scoring models and their blended prices:

The winner is Claude Sonnet 5.5: 56.0 points at $2 input and $10 output per 1M tokens, a blended $4.00. The leader, Claude Opus 5.5, scores 1.6 points higher but costs twice as much. Claude Fable 5.1 and GPT-6 Astra cost five times as much and still score lower. Meta's Muse Wonder Full ranks third but has no public API, so you cannot buy it right now.

Source: Artificial Analysis, vendor API pricing pages. Chart: Neural Daily.

Just outside the top 5 is Gemini 4 Argon (52.6 points), also at a blended $4.00. That is an introductory price: according to Google's pricing page it doubles to $4 input and $20 output on December 1. Notably, the same model leads the human-preference LMArena text leaderboard (1524), where Sonnet 5.5 sits only 36th.

What list prices don't show

• Reasoning tokens inflate real costs. Artificial Analysis also measures the cost per task including all tokens used. Sonnet 5.5 in max mode costs $5.46 per task. Opus 5.5 in xhigh mode reaches the same 56.0 points for $3.46, and GPT-6 Astra max costs $3.26. For agentic work, Opus 5.5 xhigh may be the better buy.
• The budget tier reshuffles too. Per task, MiMo-V2.6-Flash ($0.062) and GPT-6 Luna ($0.068) are the cheapest. MiMo-V2.6-Pro scores 46.3, higher than Haiku 5.5, yet costs less per task ($0.13 vs $0.21).
• Promotional pricing: Gemini 4 Argon runs at half price until November 30 and Gemini 3.8 Flash until the end of the year. DeepSeek also halves its rates off-peak.
• Benchmark limits: the index is an independent measurement rather than vendor self-reporting, but it is one organization's view. Test-set contamination cannot be fully ruled out, and gaps of one or two points are not meaningful differences.
• We did not verify prices for Meta, Xiaomi, Alibaba, StepFun, Moonshot and Mistral on the vendor pages; they come from the Artificial Analysis listing.

What to pick

For high-volume, short-prompt work (classification, extraction, routing), Claude Haiku 5.5 gives the most for the money. If you need open weights or something even cheaper, MiMo-V2.6-Flash and MiMo-V2.6-Pro are the realistic alternatives.

If top quality is the goal, Sonnet 5.5 is the cheapest ticket into the frontier tier. For long agentic runs, also try Opus 5.5 in xhigh mode, since it uses fewer tokens. Your own workload has the final say: a week of real traffic on your token bill beats any leaderboard.