There are many new “Flash” models now.
The name usually means lower price and faster inference. But it is not one standard class. The models have different sizes, context windows, reasoning modes, and benchmark settings.
I wanted a simple comparison, especially after GPT-6 Luna became another low-cost option.
One disclosure: I work with Token Harbor. The benchmark data below is not from Token Harbor. I used current Artificial Analysis results and linked the sources. Token Harbor access is listed separately.
Quick scorecard
These are current Artificial Analysis Intelligence Index v4.3.2 results checked on September 24, 2026.
| Model | AA Intelligence Index | AutomationBench-AA | Terminal-Bench 4.0 | SciCode | AA-LCR v1.1 |
|---|---|---|---|---|---|
| GLM 5.3 Flash | 42 | 60% | 33% | 52% | 80% |
| Qwen3.8-Flash-Next | 40 | 56% | 25% | 51% | 80% |
| DeepSeek V4.1 Flash (max) | 39 | 69% | 27% | 52% | 84% |
| GPT-6 Luna (xhigh) | 34 | 48% | 8% | 52% | 80% |
| MiMo V2.6 Flash | Not scored yet | — | — | — | — |
GLM has the highest composite score. DeepSeek has the best AutomationBench and long-context result in this group. Qwen stays close to both. Luna is lower on the composite score, but it is very cheap and fast.
MiMo needs a note. Artificial Analysis has a score of 46 for MiMo V2.6 Pro, not Flash. I could not find an independent AA Intelligence Index result for MiMo V2.6 Flash. So I left it blank instead of copying the Pro score.
A second model score
LLM Stats has a separate composite score. It is not the AA Intelligence Index, so I did not mix the two numbers.
| Model | LLM Stats Score |
|---|---|
| DeepSeek V4.1 Flash | 48.7 |
| Qwen3.8-Flash-Next | 48.6 |
| GLM 5.3 Flash | 48.2 |
| MiMo V2.6 Flash | 45.7 |
| GPT-6 Luna | 44.5 |
This score tells a similar story for the three open Flash models: DeepSeek, Qwen, and GLM are very close. MiMo and Luna follow, with Luna having the lowest input price in the group.
Price and context
| Model | Context in AA | Input / 1M | Output / 1M | Weights |
|---|---|---|---|---|
| GLM 5.3 Flash | 1M | $0.15 | $0.50 | Open |
| Qwen3.8-Flash-Next | 256K | $0.15 | $0.47 | Open |
| DeepSeek V4.1 Flash | 1M | $0.30 | $1.20 | Open |
| GPT-6 Luna | 1M | $0.10 | $0.50 | Proprietary |
| MiMo V2.6 Flash | 1M* | $0.14* | $0.28* | Open |
*MiMo values are from Xiaomi, because AA has not published an independent Flash profile yet.
The Qwen naming is also confusing. Alibaba calls the API model Qwen3.8 Flash. Artificial Analysis tests the related Qwen3.8-Flash-Next checkpoint and reports a 256K context window. A hosted route may expose a different limit, so I would check the provider page before sending a very large prompt.
What the scores mean to me
GLM 5.3 Flash: best overall score
GLM is first here with 42. It also has the best Terminal-Bench 4.0 result at 33%.
It looks like the safest general choice when I want one Flash model for mixed work. The downside is speed. Artificial Analysis measured it much slower than DeepSeek in the tested endpoints.
DeepSeek V4.1 Flash: strongest agent result
DeepSeek is third on the total score, but it leads AutomationBench-AA at 69% and AA-LCR at 84%.
That is more useful than the total score for some workloads. For an agent using tools or reading a large codebase, I would test DeepSeek early.
Qwen3.8 Flash: close and inexpensive
Qwen scores 40, only two points behind GLM. Its list price is also a little lower.
It does not win the rows in this small table, but there is no large gap. It may be a better fit when speed and cost matter more than the last benchmark point.
GPT-6 Luna: fast, cheap, but not the score leader
Luna xhigh scores 34. Its AutomationBench result is 48%, and Terminal-Bench 4.0 is 8%.
This does not make Luna useless. It is a proprietary model made for inexpensive, high-volume work. Artificial Analysis measured around 153 output tokens per second. Its measured cost per Intelligence Index task was only $0.04. I see it more as an implementation and background-task model than the first choice for difficult terminal work.
MiMo V2.6 Flash: interesting, but wait for independent scoring
Xiaomi reports several strong numbers for MiMo V2.6 Flash:
| Xiaomi-reported benchmark | Score |
|---|---|
| CyberGym | 95.1% |
| Terminal-Bench 2.1 | 87.6% |
| OSWorld-Verified | 80.8% |
| Toolathlon-Verified | 73.6% |
| DeepSWE 1.1 | 67.9% |
| AutomationBench v1.0.6 | 52.3% |
| Terminal-Bench 4.0 | 28.8% |
It also has a 1M context window and very low list price.
Those numbers are from Xiaomi, not Artificial Analysis. They are useful, but I would not put them in the same ranking table yet. In particular, AutomationBench v1.0.6 is not the same label as AutomationBench-AA. The rank labels in the collected list also need a defined comparison set, so I did not repeat them here.
A separate vendor-reported comparison
I found two benchmark names reported for all five models. This gives a wider view, but the results are not as clean as the independent AA table.
| Vendor-reported benchmark | MiMo V2.6 Flash | DeepSeek V4.1 Flash | Qwen3.8 Flash | GLM 5.3 Flash | GPT-6 Luna |
|---|---|---|---|---|---|
| DeepSWE 1.1 | 67.9% | 74% | 59% | 63% | 67% |
| Agents' Last Exam | 27.6% | 32% | 51% | 26% | 51% |
These are directional, not a strict leaderboard. Each vendor used its own harness, context limit, effort setting, and reporting setup. DeepSeek used maximum reasoning and a 1M-token context for DeepSWE. Qwen reports the best of two harnesses with 256K context. GLM used mini-swe-agent with 400K context. OpenAI reports Luna's highest result across an effort sweep.
The results still show something useful. DeepSeek has the strongest published DeepSWE number in this group. Qwen and Luna have the highest published Agents' Last Exam result. MiMo and Luna are almost tied on DeepSWE.
Other vendor benchmarks do not cover all five models. Some interesting examples are DeepSeek's 91% on Terminal-Bench 2.1, GLM's 78% on Toolathlon, Qwen's 92% on LiveCodeBench v6, MiMo's 95.1% on CyberGym, and Luna's 53% on OSWorld 2.0. I kept them out of the main scorecard because they do not form a complete comparison.
Trying the models on Token Harbor
Three models currently have free routes:
| Model | Token Harbor model ID | Access |
|---|---|---|
| MiMo V2.6 Flash | mimo-v2.6-flash:free |
Free allowance |
| DeepSeek V4.1 Flash | deepseek-v4.1-flash:free |
Free allowance |
| Qwen3.8 Flash | qwen3.8-flash:free |
Free allowance |
| GLM 5.3 Flash | glm-5.3-flash |
Agent Pass / paid route |
| GPT-6 Luna | gpt-6-luna |
New route; check current plan |
So the five models are not all free. I think this should be clear before anyone signs up.
For a quick test, I would use the same prompt, reasoning setting, tool schema, and output limit. I would record success, latency, tokens, and corrections. One leaderboard number is useful, but it cannot tell which model will work best with a specific agent harness.
My short choice
- For general quality: GLM 5.3 Flash.
- For tool use and long context: DeepSeek V4.1 Flash.
- For low-cost fast work: Qwen3.8 Flash.
- For a proprietary high-throughput route: GPT-6 Luna.
- For a new low-cost open model: MiMo V2.6 Flash, but I want independent AA data first.
The gap between GLM, Qwen, and DeepSeek is small. I would not choose only from the total score. The benchmark row closest to the real task matters more.
Top comments (0)