Llama 4 Scout vs Qwen3 235B A22B Instruct 2507
Llama 4 Scout (Meta, released April 2025) and Qwen3 235B A22B Instruct 2507 (Qwen, released July 2025) are both mixture-of-experts models priced under $0.35/M tokens, but they optimize for different things. Scout activates 17B of 109B total params and leans on a 10M-token context window; Qwen3 activates 22B params and leans on newer training data and stronger Design Arena scores. This comparison looks at real cost, benchmark gaps, and context tradeoffs to pick between them.
Spec vs spec
| Spec | Llama 4 Scout | Qwen3 235B A22B Instruct 2507 |
|---|---|---|
| Context window | 10M | 262K |
| Max output | 16K | 16K |
| Input modalities | text, image | text |
| Output modalities | text | text |
| Knowledge cutoff | Aug 31, 2024 | Jun 30, 2025 |
| Released | Apr 5, 2025 | Jul 21, 2025 |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
meta-llama/llama-4-scout
Input · 1M tokens
$0.100 + 3%$0.103
Output · 1M tokens
$0.300 + 3%$0.309
FEE — FLAT, EVERY MODEL3%
qwen/qwen3-235b-a22b-2507
Input · 1M tokens
$0.090 + 3%$0.093
Output · 1M tokens
$0.100 + 3%$0.103
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
meta-llama/llama-4-scout
$1.65
$1.60 provider + 3%
qwen/qwen3-235b-a22b-2507Cheaper
$1.13
$1.10 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| Llama 4 Scout | Qwen3 235B A22B Instruct 2507 | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| Code | 839 | #106 | 1088 | #81 |
| Data viz | 940 | #96 | 1102 | #77 |
| Game dev | 838 | #105 | 1018 | #94 |
| UI components | 824 | #100 | 1022 | #86 |
| Websites | 793 | #112 | 1101 | #83 |
Head-to-head preference voting. How we filter and rank
Pricing math
Provider list price: Llama 4 Scout is $0.10/M input, $0.30/M output. Qwen3 235B A22B Instruct 2507 is $0.09/M input, $0.10/M output. On OpenKey, both get the flat 3% fee applied (price × 1.03): Scout lands at $0.103/M input and $0.309/M output; Qwen3 at $0.0927/M input and $0.103/M output. Run the numbers on a realistic 10M-input / 2M-output workload and Scout costs $1.60 total versus Qwen3's $1.10 — Qwen3 is the cheaper model here even though its output-token price advantage ($0.10 vs $0.30 provider-side, a 3x gap) matters more than the smaller input price ratio of 1.11x. If your workload is output-heavy (long generations, verbose code), that output price gap compounds fast.
Coding and agentic performance
Qwen3 beats Llama 4 Scout on every Design Arena category both models report. In codecategories, Qwen3 scores 1088 elo (rank 81) vs Scout's 839 elo (rank 106). In dataviz, Qwen3 hits 1102 elo (rank 77) vs Scout's 940 elo (rank 96). In gamedev, Qwen3 is 1018 elo (rank 94) vs Scout's 838 elo (rank 105). In uicomponent, Qwen3 is 1022 elo (rank 86) vs Scout's 824 elo (rank 100). In website, Qwen3 is 1101 elo (rank 83) vs Scout's 793 elo (rank 112). Scout also reports an Artificial Analysis coding index of 8.2 and agentic index of 1.1 — low agentic index suggests it's not built for multi-step tool-use loops. Qwen3 doesn't report Artificial Analysis numbers in this data, but its Design Arena wins are consistent and large across every shared category.
Context and long-document work
This is Scout's one clear lane. Its context window is 10,000,000 tokens against Qwen3's 262,144 — a 38.15x ratio. Both cap max output at 16,384 tokens per call. If you're doing single-pass analysis over massive inputs (huge log files, full repos, multi-book corpora) where you can't chunk, Scout is the only one of the two that fits the whole thing in one context window. For anything under a quarter-million tokens, Qwen3's window is plenty and its per-token cost and benchmark scores make it the better default.
Modality and freshness
Llama 4 Scout takes text and image input and outputs text — useful if your pipeline needs to reason over screenshots or diagrams alongside text. Qwen3 is text-only in and out. On freshness, Qwen3's knowledge cutoff is 2025-06-30 versus Scout's 2024-08-31, roughly a 10-month gap, and Qwen3 was released three and a half months later (created 2025-07-21 vs 2025-04-05). If your use case needs recent world knowledge baked into the weights, Qwen3 has the edge; if it needs image understanding, Scout is the only option of the two.
When to pick each
Pick Qwen3 235B A22B Instruct 2507 by default: it's cheaper on a real workload, wins every shared Design Arena benchmark, and has a more recent knowledge cutoff. Pick Llama 4 Scout only when you specifically need its 10M-token context window or its image input support — those are the two things Qwen3 can't do. Both models are available on OpenKey through one API key, with OpenKey pricing set at the provider's list price plus a flat 3% fee.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| General coding assistant | Qwen3 235B A22B Instruct 2507 | Higher elo in codecategories (1088 vs 839) at lower cost |
| Analyzing a full repo or huge log file in one pass | Llama 4 Scout | 10M-token context vs Qwen3's 262,144 |
| Image + text input tasks | Llama 4 Scout | Only Scout supports image input modality |
| UI component generation | Qwen3 235B A22B Instruct 2507 | 1022 elo vs Scout's 824 in uicomponent category |
| Cost-sensitive high-output-volume jobs | Qwen3 235B A22B Instruct 2507 | $0.10/M output vs Scout's $0.30/M output, provider price |
| Multi-step agentic tool use | Qwen3 235B A22B Instruct 2507 | Scout's Artificial Analysis agentic index is just 1.1 |
Questions
- Which model is cheaper for a typical workload?
- Qwen3 235B A22B Instruct 2507 costs $1.10 for 10M input + 2M output tokens versus $1.60 for Llama 4 Scout, using OpenKey pricing (provider price plus flat 3% fee). The gap comes mostly from output pricing: $0.10/M vs $0.30/M provider-side.
- Does Llama 4 Scout beat Qwen3 on any benchmark?
- Not in the Design Arena categories both models report. Qwen3 leads in all five shared categories, including codecategories (1088 vs 839 elo) and website (1101 vs 793 elo). Scout's advantage is architectural — context length and image input, not benchmark scores.
- How much bigger is Llama 4 Scout's context window?
- Scout supports 10,000,000 tokens of context versus Qwen3's 262,144 — a 38.15x ratio. Both models cap output at 16,384 tokens per response, so the difference only matters for input size, not generation length.
- Can either model handle images?
- Only Llama 4 Scout. Its input modalities are text and image; Qwen3 235B A22B Instruct 2507 accepts text only. Both output text only, with max completion length capped at 16,384 tokens.