Head to head · October 2026

Claude Haiku 5.5 vs Mistral Small 4

Claude Haiku 5.5 scores higher on the Models at Work leaderboard (80 vs 42 out of 100). Claude Haiku 5.5 is about 1.3× cheaper per blended million tokens ($0.20 vs $0.26). On Artificial Analysis' independent Intelligence Index, Claude Haiku 5.5 leads 43 to 11.

Updated Oct 11, 2026 · every number links to who measured it

	Claude Haiku 5.5 · Anthropic	Mistral Small 4 · Mistral AI
Leaderboard rank	#8	#24
Score (0–100)	80	42
Tier	Fast and cheap	Open weights
Price per 1M tokens	$0.10 in · $0.50 out	$0.15 in · $0.60 out
Context window	1M	256K
Open weights	No	Yes
Artificial Analysis Intelligence Index	43	11
Best for	Classification and extraction, Routing tiers, Latency-sensitive paths	Self-hosted multimodal chat on one GPU, Fine-tuning base for EU-regulated teams, Cheap fast tasks with a reasoning_effort dial

Choose Claude Haiku 5.5 if…

you need classification and extraction
you need routing tiers
you need latency-sensitive paths

Ten cents a million and, per the people using it, smarter than the other ten-cent model; mind the 100K cliff.

Caveats: Price jumps five-fold above 100K input tokens, which commenters call an absurdly low cutoff for agent runs.

Choose Mistral Small 4 if…

you need self-hosted multimodal chat on one gpu
you need fine-tuning base for eu-regulated teams
you need cheap fast tasks with a reasoning_effort dial

One Apache-2.0 model replacing Magistral, Pixtral and Devstral: 119B MoE with 6B active, so it fits one H100 at 4-bit and runs at ~165 tok/s. $0.15/$0.60 is hard to beat for a self-hostable multimodal model, but it is not clever.

Caveats: AA 11; HN testers rate it 'okay, nothing exceptional' and 'not the best' for complex agentic work. Not listed on Arena; limited independent evaluation. Mistral Medium 3.5 (Modified MIT) is the step up if you need more intelligence from the same vendor.

What practitioners say

Claude Haiku 5.5: The thread had over a thousand points in a day; the recurring reports are near-perfect structured classification in production and 'noticeably smarter than GPT-6 Luna', against complaints about the 100K pricing tier.

Mistral Small 4: Modest HN reception (126 points). Fans liked the economics: '$0.60/1M output is a steal' versus Qwen models that 'waste tokens on reasoning', and that ~120B 'fits onto a single H100 with 4 bit quant'. Others were blunt: 'okay, nothing exceptional', 'I'd use it for some basic tasks but not actual complex tasks', and one tester called it 'worse than glm air 4.5'. A later comment found the 4-bit quant roughly comparable to Qwen 3.6 27B.

Questions people ask

Which is better, Claude Haiku 5.5 or Mistral Small 4?

When should I choose Claude Haiku 5.5 over Mistral Small 4?

Choose Claude Haiku 5.5 for classification and extraction, routing tiers, latency-sensitive paths. Ten cents a million and, per the people using it, smarter than the other ten-cent model; mind the 100K cliff.

When should I choose Mistral Small 4 over Claude Haiku 5.5?

Choose Mistral Small 4 for self-hosted multimodal chat on one gpu, fine-tuning base for eu-regulated teams, cheap fast tasks with a reasoning_effort dial. One Apache-2.0 model replacing Magistral, Pixtral and Devstral: 119B MoE with 6B active, so it fits one H100 at 4-bit and runs at ~165 tok/s. $0.15/$0.60 is hard to beat for a self-hostable multimodal model, but it is not clever.

Which is cheaper, Claude Haiku 5.5 or Mistral Small 4?

Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens on Anthropic's list price (Claude pricing, read Oct 11, 2026). Mistral Small 4 costs $0.15 per million input tokens and $0.60 per million output tokens on Mistral AI's list price (Mistral Small 4 announcement, read Oct 11, 2026).

Scores come from the Models at Work leaderboard: Score is 0 to 100 and computed, not typed: 55% independent evals (Artificial Analysis, Arena, Scale SEAL, Epoch, each scaled against its natural floor and the best score in this table, then averaged), 15% blended price on a fixed log scale ($0.05 per million is 100, $60 is 0), 15% practitioner sentiment from Signals and community threads, 15% operational fit (context window, open weights). A missing component drops out and the rest are reweighted. A model with no independent eval yet is provisional and ranks below every measured one. Within three points is a tie. Vendor-published figures are labelled and never counted. Prices are vendor list prices where published; blended figures assume three input tokens per output token.