Independent benchmarks put Qwen3.8-Flash-Next through intelligence, speed and cost tests

What happened

Artificial Analysis, an independent outfit that measures AI models on its own benchmark suite, has published a dedicated evaluation page for Qwen3.8-Flash-Next, the newest addition to the Qwen "Flash" line. The write-up puts the model's benchmark intelligence, output speed, latency, context window, token usage and pricing side by side, and it drew enough interest to reach Hacker News's front page on 27 August 2026, according to the post's metadata.

The timing is the point: new Flash-tier releases typically arrive with vendor-selected benchmarks and headline per-token prices, while developers deciding whether to switch care about how fast the model answers and what a typical task actually costs.

A nine-part intelligence index

At the centre of the analysis is the Artificial Analysis Intelligence Index, currently at version 4.1.1, which aggregates nine evaluations: GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR. The spread is deliberately wide, covering quantitative work on spreadsheets and documents, a banking-flavoured reasoning task, coding inside a terminal environment, scientific code generation, and two well-known hard-knowledge tests. Models that reason before answering are flagged with a separate marker, so a score earned through long chains of thought is not mistaken for one achieved in a single pass.

The index charts also split open-weight and proprietary models into separate lanes, and each model carries an Openness Index score: a 0-to-100 normalised measure of how open the release is, with labels such as "Commercial Use Restricted" or "Non-commercial" applied when the licence limits commercial use.

Hallucination gets its own score

One of the nine components, AA-Omniscience, measures knowledge reliability rather than raw knowledge. It rewards correct answers, penalises hallucinated ones, and imposes no penalty for refusing to answer. Scores run from -100 to 100, where zero means the model gives as many wrong answers as right ones and a negative score means it is wrong more often than correct.

Cost per task, not just per token

The pricing section is where this kind of analysis earns its keep. Rather than stopping at list prices per million tokens, Artificial Analysis computes the weighted average cost, in dollars, of completing a single Intelligence Index task. That figure folds in input, cache-hit, cache-write, reasoning and answer token prices, divides by task count, and weights each evaluation by its share of the index. The page also reports output tokens consumed per task and the total cost of running the entire index end to end.

This framing matters because reasoning models can spend a large hidden budget of thinking tokens before producing any answer, so two models with identical output prices can differ enormously in what a real job costs. Cache-hit pricing is reported separately from standard input pricing, with the caveat that cache writes and storage are billed separately and vary by provider.

Speed, latency and context

On the performance side, the page tracks output speed in tokens per second, measured on a model's first-party API where one exists or the median across providers where it does not. Latency is measured as the time to the first answer token, which for reasoning models includes the "thinking" phase before any answer text appears. A separate end-to-end figure estimates the seconds needed to emit a 500-token response, combining first-token latency, thinking time and generation speed.

There is also a per-task decode time, calculated by dividing output tokens per task by output speed and weighting across the index, which translates raw throughput into something closer to user-perceived responsiveness. The context window limit is listed too, a number that matters for retrieval-heavy workflows that stuff large document sets into the prompt.

Why it matters

Flash-tier models are the workhorses of production AI: high call volume, tight latency budgets, thin margins. For that class of model, the deciding questions are rarely about which one tops a vendor's favourite benchmark, but which one answers fast enough and costs least per completed job. By converting pricing into per-task dollars and throughput into per-task minutes, and by running the same evaluations independently rather than accepting vendor numbers, this style of analysis lets developers rank candidates on measured behaviour. The Hacker News pickup suggests demand for exactly that kind of evidence is keeping pace with the flood of new model releases.