Tested · October 2026
The best LLM for sentiment analysis in 2026, tested
We asked 19 models the same two sentiment questions: is this product review positive, and which of six emotions does this tweet express. Positive versus negative turned out to be solved; emotions did not. Here is the best LLM for sentiment analysis at each, what it costs, and a box to try your own text.
Run this example
2 questions · no key neededQuick answer · verified 2026-10-11
What is the best LLM for sentiment analysis?
For positive versus negative, almost any current model will do: 18 of 18 hosted models scored 95% or more on product reviews, so choose on price and on whether you need a probability. For fine-grained emotions no model is reliable yet, and a good part of the gap is the labels themselves.
| Positive or negative | 18 of 18 hosted models at 95%+ — 100 product reviews |
|---|---|
| Cheapest at 98%+ | Decider V1.1 27B, $0.0072 per 1,000 — 98% on reviews |
| Six emotions | 56% to 67% — every model, one suspect left out |
| Best chat LLM | Claude Haiku 5.5 — 99% reviews, 67% emotions |
How we tested
Both sets come from BTZSC, a public zero-shot classification benchmark, sampled with a fixed seed: 100 Amazon product reviews (positive or not) and 100 tweets, each labelled with one of six emotions. Chat LLMs ran at temperature 0 with reasoning off and had to reply with one option; decision models answered through the Decisions API; Laya ran locally. These are the same runs as the text classification test, read for the two sentiment sets only. Scripts and raw answers are in jev-measured.
Sentiment results, model by model
| Model | Type | Reviews | Emotions | Per 1,000 |
|---|---|---|---|---|
| d1 | Decision model | 100% | 64% | $0.0078 |
| Mercury Decide (free)(likely saw the emotion set in training) | Decision model | 99% | 92% | free tier |
| Claude Haiku 5.5 | Chat LLM | 99% | 67% | $0.034 |
| Qwen3.8 Flash | Chat LLM | 99% | 63% | $0.023 |
| DeepSeek V4.1 Flash | Chat LLM | 98% | 65% | $0.022 |
| Jev 1.13 | Decision model | 98% | 63% | $0.030 |
| Clef | Decision model | 98% | 61% | $0.154 |
| Decider V1.1 27B | Decision model | 98% | 60% | $0.0072 |
| Solar Decide Flash | Decision model | 98% | 59% | $0.022 |
| Tev1 4B Experimental | Decision model | 98% | 59% | $0.0072 |
| Clef Flash | Decision model | 98% | 56% | $0.058 |
| GPT-6 Luna Decisions | Decision model | 97% | 62% | $0.040 |
| GPT-6 Luna (chat) | Chat LLM | 97% | 60% | $0.024 |
| Solar Decide | Decision model | 97% | 59% | $0.022 |
| Span-01 Lite (free) | Decision model | 97% | yes/no only | free tier |
| Span-01 | Decision model | 97% | yes/no only | $0.0029 |
| Nemotron 3 Ultra (free) | Chat LLM | 95% | 59% | free tier |
| Kev 4B | Decision model | 95% | 57% | $0.011 |
| Laya (ModernBERT, local) | Local, open weights | 67% | 60% | your hardware |
Chat LLMs are shaded. With 100 texts a set, two or three points apart is noise.
Positive or negative: solved, so pay less
On whole-review polarity the hosted models sit between 95% and 100%. The misses are mostly reviews that praise and complain in the same breath (“I love it! When it works”), plus the odd sarcastic one that calls a TV show the perfect cure for sleepless nights. That makes price the deciding factor: Decider V1.1 27B scored 98% for $0.0072 per 1,000 reviews, a fraction of what the chat models cost here. If your reviews are mixed as often as not, ask for a probability and treat the middle as its own class.
Emotions: where every LLM struggles, and why
Six-way emotion is a different story: apart from one model that looks like it trained on this dataset, every model landed between 56% and 67%. We counted the mistakes across all of them. The most common was a text labelled sadness answered as joy, about one error in five, then anger answered as sadness and joy answered as love.
Reading those items explains most of it. “Im much more peaceful and happy when the house is clean” is labelled sadness in the dataset; every model that answered joy was right in any sense a person would accept. Short first-person posts also carry two feelings at once. Treat an emotion benchmark score under about 70% as a ceiling set partly by the labels, and test on texts you labelled yourself before choosing a model for this.
Best small, local and free options for sentiment
Small does not have to mean worse here: Tev1 4B and Kev 4B scored 98% and 95% on reviews, and Kev publishes its weights. Nemotron 3 Ultra on OpenRouter's free tier reached 95%. The local general-purpose model we ran, Laya, managed only 67% on reviews, and its score moved between 56% and 71% with the wording of the question, so a model trained for sentiment is the safer local choice. More on Laya in Jev vs Laya.
Beyond one label: sentiment per aspect
A whole-review label hides what customers actually like and dislike. For that, ask one question per aspect (sound, battery, support) and count the answers across reviews: the AI review analysis tool does this and reports complaints by aspect, and the content moderation pattern covers toxicity, the other sentiment-like task in our test.
Best LLM for sentiment analysis: questions
Is a fine-tuned model better than an LLM for sentiment?
For positive versus negative, the hosted models here already reach 95% to 100% without training, so fine-tuning buys little. It earns its keep for a domain with its own vocabulary, or when you need to run millions of texts locally at almost no cost.
What is the best open-source LLM for sentiment analysis?
Of the models we ran, three publish their weights: Clef and Clef Flash scored 98% on product reviews and Kev 4B 95%. Laya, which we ran on a laptop, reached only 67%. None of the open models did better than 61% on the six emotions.
Can an LLM tell how sure it is about sentiment?
A chat reply does not say. Decision models return a probability with every answer, so you can act on clear cases and send mixed reviews, which sit near 0.5, to a person or a closer read.
What about sentiment per product feature?
That is aspect-based sentiment: a review can praise the sound and complain about the battery. The AI review analysis tool on this site asks one question per aspect and counts praise and complaints across a set of reviews.