OpenAI benchmarked efficiency vs Alibaba's low-cost multimodal Flash API · research snapshot July 12, 2026

GPT-5.6 Luna vs Qwen 3.6 Flash

Qwen3.6-35B-A3B is the open-weight model that Alibaba exposes through Model Studio as qwen3.6-flash. It combines 35B total parameters with only 3B active parameters, native multimodal input and a surprisingly strong official scorecard across coding, agents, reasoning and vision. Luna still has the higher published frontier scores on several shared rows, but Qwen Flash is no longer an unbenchmarked mystery: the official Qwen evaluation gives us a real, efficient open-model baseline.

Last updated July 12, 2026 · Qwen 3.6 Flash benchmark coverage is explicitly labeled

62.7%
Luna SWE-Bench Pro
49.5%
Qwen Flash SWE-Bench Pro
73.4%
Qwen Flash SWE-Bench Verified
35B / 3B
Qwen total / active parameters
Short answer: choose GPT-5.6 Luna when you need a known quantity for difficult coding, tool-heavy agents or procurement backed by public benchmark evidence. Choose Qwen 3.6 Flash when video input, very low API cost, 1M context and built-in multimodal tools matter—and you are willing to run your own acceptance tests before trusting it with high-stakes reasoning.

What each model actually is

SpecificationGPT-5.6 LunaQwen 3.6 Flash
Provider / releaseOpenAI · July 2026 GAAlibaba / Qwen · April 15, 2026
API model IDgpt-5.6-lunaqwen3.6-flash (Alibaba API alias)
Model identityGPT-5.6 cost-efficient tierQwen3.6-35B-A3B · 35B total / 3B active
Context window1.05M tokens262K native; extensible to 1.01M with YaRN; Alibaba API advertises 1M
Maximum output128K tokens64K API; Qwen research used up to 81,920 for difficult benchmark runs
Input / outputText and images / textText, images and video / text
Function callingSupportedSupported; thinking preservation is available for agents
Built-in toolsWeb, file, code, computer use, MCP and tool search through Responses APIQwen Code, Qwen-Agent, MCP and API tool calling
Structured outputSupportedSupported
Images / video limitsImage input; OpenAI limits vary by detail and endpointNative vision encoder plus image and video input
Weights / licenseClosed proprietary APIOpen-source/open weights; available on Hugging Face and ModelScope

The identity is explicit in Qwen's release post: Qwen3.6-35B-A3B is the open-weight checkpoint exposed through Alibaba Model Studio as qwen3.6-flash. They are not separate models for this comparison. Qwen3.6-Plus is a different, larger model and its scores are intentionally excluded.

Benchmark evidence: Qwen Flash now has a real scorecard

Qwen's official Qwen3.6-35B-A3B release provides a broad scorecard for the exact checkpoint exposed as qwen3.6-flash. The table below combines those Qwen-reported results with Luna's OpenAI scorecard where a useful comparison exists. Scores are not automatically apples-to-apples: Qwen's SWE series uses its internal agent scaffold and corrected tasks, while Luna's results come from OpenAI's evaluation lane.

EvaluationGPT-5.6 LunaQwen 3.6 Flash
Qwen3.6-35B-A3B
Reading
SWE-Bench Pro62.7%49.5%Luna +13.2; harnesses differ
SWE-Bench VerifiedNot published in Luna scorecard73.4%Qwen-only official result
SWE-Bench MultilingualNot published67.2%Qwen-only official result
Terminal-Bench84.7% on 2.151.5% on 2.0Version and harness mismatch
QwenClawBenchNot published52.6%Qwen internal real-user-distribution agent benchmark
QwenWebBench EloNot published1397Qwen internal front-end/artifacts benchmark
NL2RepoNot published29.4%Qwen official result
MCPMarkNot published37.0%GitHub MCP benchmark; Qwen official result
MCP-Atlas publicNot published by OpenAI62.8%Different benchmark from MCPMark
GPQA / GPQA Diamond92.3%86.0%Luna +6.3; provider setups differ
HMMT Feb 26Not published83.6%Qwen-only official result
MMMU / MMMU Pro78.4% on MMMU Pro81.7% on MMMUDifferent benchmark variants
RealWorldQANot published85.3%Qwen vision-language result
Fair-comparison rule: Qwen's numbers are official but mostly vendor-reported, and the SWE results use a corrected task set, internal scaffold and 200K context. Luna's rows use OpenAI's own harness and different benchmark versions in some cases. Read the table as directional evidence, not as a universal league table.
Selected official benchmark scores
Higher is better. Terminal-Bench and MMMU rows use different versions or protocols, so they are shown as directional context only.
SWE Pro · Luna
62.7
SWE Pro · Qwen Flash
49.5
GPQA · Luna
92.3
GPQA · Qwen Flash
86.0
SWE Verified · Qwen
73.4
MMMU · Qwen
81.7

This evidence asymmetry matters for enterprise adoption. Luna can be evaluated before production with public reference points for coding, reasoning, tool use, computer use, science and long context. Flash should be evaluated directly on your data: image and video extraction, JSON validity, tool-call selection, grounding, latency, multilingual behavior, refusals and cost per successful task.

Radar: selected official benchmark profile

SWE ProTerminalGPQAMMMU

● Luna ● Qwen3.6 Flash

Native percentages are mapped onto the same 0–100 radial scale. Terminal-Bench uses 2.1 for Luna versus 2.0 for Qwen, and MMMU versus MMMU Pro are different variants; those axes are directional, not a definitive ranking.

  • Luna leads the selected SWE Pro, Terminal and GPQA axes.
  • Qwen is competitive on multimodal MMMU and has additional official wins in vision, MCPMark, QwenWebBench and SWE-Bench Verified.
  • This radar intentionally excludes unavailable Luna scores rather than plotting missing data as zero.

Multimodal and tool workflows

Luna's workflow advantage

Luna sits inside OpenAI's GPT-5.6 tool ecosystem. The cited API docs list web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. For a software agent, that breadth can be more important than raw input modality count: the model can search, inspect, edit, execute and verify through a consistent platform.

Its public scorecard also includes computer-use, browsing, tool-use and multimodal evaluation rows, giving teams more evidence before they grant an agent permissions.

Qwen Flash's workflow advantage

Qwen 3.6 Flash is documented as a native vision-language model that accepts images and video alongside text. Alibaba's visual-understanding documentation lists 1M context, 64K maximum output, function calling, built-in tools and structured output, with generous media-count limits.

That makes Flash a natural candidate for video summarization, document intake, media search, multimodal extraction and cost-sensitive assistants. The official scorecard now backs up that feature story: 73.4% SWE-Bench Verified, 67.2% SWE-Bench Multilingual, 81.7% MMMU, 85.3% RealWorldQA and 37.0% MCPMark. Those results are valuable directional evidence, but the provider's internal harness and task corrections still make a local acceptance test essential.

Workflow questionBetter starting pointWhy
Need video inputQwen 3.6 FlashAlibaba documents native video input and video limits.
Need large text/code outputLuna128K maximum output versus 64K.
Need a public coding baselineLunaOpenAI publishes SWE Pro, Terminal-Bench and other rows.
Need built-in multimodal toolsQwen FlashAlibaba lists built-in tools and structured output for the Qwen3.6 family.
Need computer use and MCP ecosystemLunaOpenAI documents computer use, MCP and tool search.
Need the lowest token rateQwen FlashAlibaba's global base tier is far below Luna's list price.

Pricing: Qwen is cheaper by a wide margin

Alibaba Cloud's global rate table uses context tiers. The base tier is listed at $0.165 input / $0.99 output per 1M tokens; requests above 256K input tokens are listed at $0.66 input / $3.961 output. OpenRouter displays a provider route at $0.1875/$1.125, so production buyers should use the rate card for the exact endpoint they will call.

Per 1M tokensGPT-5.6 LunaQwen 3.6 Flash
Base fresh input$1.00$0.165
Base output$6.00$0.99
Long-input fresh input$2.00 above 272K$0.66 above 256K
Long-input output$9.00 above 272K$3.961 above 256K
Example: 10M base input + 1M output$16.00$2.64
Example: 10M long-tier input + 1M output$29.00$10.561
Aggregate base-tier example: 10M input + 1M output
Alibaba Cloud Global base rates versus OpenAI standard short-context rates; provider markups, tool fees and retries excluded.
Qwen 3.6 Flash
$2.64
GPT-5.6 Luna
$16.00

At base rates, Luna costs about six times as much in this mixed example. Qwen remains cheaper in the long tier, but the gap narrows because long-context input and output are billed at higher rates. The most important production metric is still cost per successful task: a cheap Flash response that needs re-prompting or human correction can lose to a more expensive Luna response that succeeds immediately.

How to interpret and validate the Qwen scorecard

The official Qwen table establishes a meaningful baseline, but it does not eliminate the need for product-level evaluation. Qwen's SWE rows use a corrected task set, internal scaffolding and a 200K context; several agent and web rows are internal benchmarks. Use the published results to choose what to test, then measure the complete agent with your own repository, tools and latency budget:

Test laneSuggested measurements
Multimodal extractionOCR accuracy, table fidelity, chart understanding, image grounding and JSON validity.
Video understandingTemporal question accuracy, event localization, long-video recall and media-token cost.
CodingUse Qwen's 49.5% SWE-Bench Pro and 73.4% Verified rows as baselines, then reproduce the same repository, tools, timeout and acceptance tests for Luna.
Agent toolsCorrect tool selection, argument validity, recovery after errors, steps per success and human interventions.
Production economicsTime to first token, end-to-end latency, cache hit rate, retries, output tokens and cost per accepted result.
Safety and complianceData residency, retention, refusal behavior, prompt-injection resistance and provider contract terms.
Recommended experiment: send 100 representative tasks to both models, stratified across text, image, video, coding and tools. Record pass/fail with a deterministic rubric, then calculate cost per successful task. This will tell you more than assuming that a “Flash” label means weak reasoning or that a 1M window means perfect long-context recall.

Who should choose which?

Choose GPT-5.6 Luna if…

  • You need public benchmark evidence for coding, reasoning, tool use and computer use.
  • Your agent must work across OpenAI-hosted web, file, code, shell, MCP or computer-use tools.
  • You need 128K output and a newer knowledge cutoff.
  • You are optimizing for first-pass success rather than the lowest raw token price.

Choose Qwen 3.6 Flash if…

  • Video understanding and multimodal input are central to the product.
  • You process very high volumes and need a low-cost global API route.
  • Your tasks fit a 64K output ceiling and you can run acceptance tests.
  • You want Alibaba's built-in multimodal tools and 1M context without paying frontier-model rates.

Validate the scorecard on your workload

Qwen Flash now has a substantial official benchmark profile. A CodingFleet side-by-side test still matters because provider harnesses, media inputs and tool schemas can change the production outcome. Compare it against Luna on your own prompts, repositories and tool calls.

Test both in CodingFleet →

Final verdict

GPT-5.6 Luna is the higher-capability managed choice. It leads the selected SWE Pro, Terminal and GPQA comparisons, has a broader hosted tool ecosystem and offers a larger output ceiling. Its benchmark evidence is also more directly aligned with OpenAI's production stack.

Qwen 3.6 Flash is the value-and-modality choice. Alibaba documents a 1M context, image and video input, built-in tools, structured output and a remarkably low rate card. That combination is compelling for media-rich and high-volume workloads.

The honest result is not a fabricated benchmark winner. Luna is the safer model to choose when quality must be predicted before deployment. Qwen Flash is the model to test when video and unit economics matter enough to justify doing the evaluation yourself.

Sources and methodology