The Best-Value AI Models Right Now (Data, Not Vibes)

The cheapest model is rarely the best value, and the smartest model is rarely the cheapest. This page settles the argument with math. We rank the best value AI models by capability-per-dollar, so you can see where a strong Arena score meets a low token price.

Data from our live pipeline, updated July 8, 2026. Prices sync every 6 hours.

No lab tests here, no vibes. Our authority is the pipeline: prices synced every 6 hours from the live market, capability from public LMArena scores. We combine the two into one number and sort. That is it.

How the Value Score works

The formula is deliberately plain:

Value Score = (Arena Score − 1200) ÷ blended price per 1M tokens, where blended = 0.75×input + 0.25×output.

Two choices worth explaining. We subtract 1200 from the Arena score because a raw score is not "capability from zero" — 1200 is roughly the floor of usefulness on the board, so the numerator measures capability above that floor. And we weight the blend 75/25 toward input because most real workloads read far more tokens than they write (retrieval context, system prompts, documents) than they generate.

Higher is better. A high score means you are buying a lot of measured capability per dollar. It does not mean the model is the smartest one available; it means it is the most efficient use of a budget. Keep that distinction in mind as you read.

Arena Scores © LMArena, licensed CC-BY-4.0, as of 2026-07-02. Prices are from our own live market sync.

The current top 10

|---|-------|-------|---------------------|---------|-------------|

| 1 | [gpt-oss-120b](/models/openai-gpt-oss-120b) | 1366 | $0.03 / $0.15 | $0.06 | 2758.3 |

| 2 | [Qwen3 30B A3B Instruct 2507](/models/qwen-qwen3-30b-a3b-instruct-2507) | 1384 | $0.0482 / $0.193 | $0.08 | 2175.9 |

| 3 | [DeepSeek V4 Flash](/models/deepseek-deepseek-v4-flash) | 1431 | $0.09 / $0.18 | $0.11 | 2056.9 |

| 4 | [Gemma 4 26B A4B](/models/google-gemma-4-26b-a4b-it) | 1435 | $0.06 / $0.33 | $0.13 | 1839.2 |

| 5 | [Gemma 3 12B](/models/google-gemma-3-12b-it) | 1334 | $0.05 / $0.15 | $0.08 | 1789.3 |

| 6 | [Qwen3.5-Flash](/models/qwen-qwen3-5-flash-02-23) | 1399 | $0.065 / $0.26 | $0.11 | 1746.8 |

| 7 | [Gemma 3 27B](/models/google-gemma-3-27b-it) | 1358 | $0.08 / $0.16 | $0.10 | 1582.0 |

| 8 | [gpt-oss-20b](/models/openai-gpt-oss-20b) | 1288 | $0.029 / $0.14 | $0.06 | 1545.4 |

| 9 | [Gemma 3 4B](/models/google-gemma-3-4b-it) | 1291 | $0.05 / $0.10 | $0.06 | 1452.8 |

| 10 | [Gemma 3n 4B](/models/google-gemma-3n-e4b-it) | 1306 | $0.06 / $0.12 | $0.08 | 1416.0 |

The full, always-current version lives on the [value leaderboard](/leaderboard/value). Prices move; the ranking moves with them.

Reading the leaderboard

#1: gpt-oss-120b sets the pace

gpt-oss-120b leads with a Value Score of 2758.3, and the gap to second place is wide: it beats #2 by more than 580 points. It does not have the top Arena score on this list, but at a blended $0.06 it is the cheapest of the capable tier, and price is what does the work here. An Arena of 1366 for six cents blended is the best deal our sync can find right now.

#2 and #3: the open-source middle

[Qwen3 30B A3B Instruct 2507](/models/qwen-qwen3-30b-a3b-instruct-2507) posts a higher Arena score (1384) than the leader, but its blended $0.08 pulls the Value Score to 2175.9. [DeepSeek V4 Flash](/models/deepseek-deepseek-v4-flash) is the interesting one: at Arena 1431 it is the second-most-capable model on the whole list, and its blended $0.11 still lands it at 2056.9 and third overall. If you want the most raw capability among the value picks, this is the line to watch. See how it stacks against a mainstream option in our [DeepSeek vs ChatGPT](/vs/deepseek-vs-chatgpt) comparison.

#4: highest capability, mid value

[Gemma 4 26B A4B](/models/google-gemma-4-26b-a4b-it) tops the list on Arena at 1435. Why is it only #4 by value? Its output price of $0.33 is the steepest here, which lifts the blended cost to $0.13. That is the whole story of this leaderboard in one row: capability alone does not win: the price tag decides placement.

The Gemma cluster

Google's Gemma family occupies five of the ten slots. [Gemma 3 12B](/models/google-gemma-3-12b-it) at #5 is a clean pick — Arena 1334 at a blended $0.08. Below it, [Gemma 3 27B](/models/google-gemma-3-27b-it), [Gemma 3 4B](/models/google-gemma-3-4b-it) and [Gemma 3n 4B](/models/google-gemma-3n-e4b-it) trade size for price. The 4B models sit at a blended $0.06 with Arena scores near 1291 and 1306; they are the floor-price options that still clear the 1200 usefulness line.

The cheapest capable option

[gpt-oss-20b](/models/openai-gpt-oss-20b) is tied for the lowest blended price on the board at $0.06, matching its bigger sibling. Its Arena of 1288 is lower, which is why it lands at #8 rather than the top. But if your workload is high-volume and forgiving, the 20B and the 120B share the same blended cost, so the 120B is the obvious upgrade at no extra price per token.

What each buyer should take from this

Hobbyists and side projects

Start with the cheapest capable tier. gpt-oss-20b and the Gemma 4B models all blend to $0.06, so cost is a rounding error at hobby volume. Pick on quality: gpt-oss-120b gives you a materially higher Arena score (1366 vs 1288) for the same blended $0.06. There is no reason to pay attention to anything more expensive until a real workload demands it. Browse the full field on [/models](/models).

Startups shipping a product

You want headroom without overspending. gpt-oss-120b is the default: #1 value, Arena 1366, and the widest lead on the board. If your product leans on reasoning or long context and you can absorb $0.11 blended, DeepSeek V4 Flash buys you the highest capability in the value tier (Arena 1431). Wire two into a fallback chain and compare behavior in your own traffic. Our [stacks](/stacks) and [coding](/coding) pages show how teams combine models rather than betting on one.

Heavy API users

At scale, the blended price is your margin. The spread between $0.06 (gpt-oss-120b) and $0.13 (Gemma 4 26B A4B) is more than double per token — on billions of tokens that is a real line item. Anchor on the $0.06 tier and only step up when a measurable quality gap justifies the cost. Watch the [value leaderboard](/leaderboard/value) closely, because a single input-price cut can reshuffle the top five, and check [/pricing](/pricing) before you commit a quarter's spend.

Caveats worth stating

Arena scores measure general preference, not your specific task. A model that ranks well overall can underperform on your domain, and one further down can surprise you. Treat this leaderboard as a shortlist generator, not a verdict. Prices also move every 6 hours; the ordering here is a snapshot, and the [live page](/leaderboard/value) is the source of truth. For head-to-head context on the mainstream names buyers ask about most, see [ChatGPT vs Gemini](/vs/chatgpt-vs-gemini) and [Claude vs ChatGPT](/vs/claude-vs-chatgpt), and track what is climbing on [/trending](/trending).

FAQ

Does the highest Value Score mean the best model?

No. It means the most measured capability per dollar. Gemma 4 26B A4B has the top Arena score (1435) but sits at #4 because its blended $0.13 costs more than double the leader's $0.06. Value and raw capability are different questions.

Why weight input over output in the blend?

Most production workloads read far more than they write — long prompts, retrieved context, documents. The 75/25 input weighting reflects that, so the score tracks real bills rather than best-case generation-only pricing.

How current are these numbers?

Prices sync from the live market every 6 hours. Arena scores are from LMArena as of 2026-07-02. If a provider cuts input price, the ranking can shift, so check the [value leaderboard](/leaderboard/value) for the latest order.

Ezra, Scout AI Team