DeepSeek V4 Flash 0731 vs GPT-6 Luna vs Muse Spark 1.3
Too close to call on our weighted score (GPT-6 Luna 80, DeepSeek V4 Flash 0731 80, Muse Spark 1.3 68). The right pick depends on what you value most.Too close to call on our weighted score (Muse Spark 1.3 82, GPT-6 Luna 81, DeepSeek V4 Flash 0731 79). The right pick depends on what you value most.
DeepSeek
DeepSeek V4 Flash 0731
80/100- ECI154.4
- Price$0.14 / $0.30
- Context1M
OpenAI
GPT-6 Luna
80/100- ECI156.5
- Price$0.10 / $0.50
- Context1.05M
Meta
Muse Spark 1.3
68/100- ECI158.2
- Price$1.25 / $4.25
- Context1.05M
DeepSeek
DeepSeek V4 Flash 0731
79/100- ECI154.4
- Price$0.14 / $0.30
- Context1M
OpenAI
GPT-6 Luna
81/100- ECI156.5
- Price$0.10 / $0.50
- Context1.05M
Meta
Muse Spark 1.3
82/100- ECI158.2
- Price$1.25 / $4.25
- Context1.05M
Too close to call
It is close. Our weighted score puts them within a point (GPT-6 Luna 80/100, DeepSeek V4 Flash 0731 80/100, Muse Spark 1.3 68/100), so choose by what matters most for your work: Muse Spark 1.3 for raw capability and DeepSeek V4 Flash 0731 on price. The score weighs capability 59%, price 29%, context window 12%.It is close. Our weighted score puts them within 1 points (Muse Spark 1.3 82/100, GPT-6 Luna 81/100, DeepSeek V4 Flash 0731 79/100), so choose by what matters most for your work: Muse Spark 1.3 for raw capability and DeepSeek V4 Flash 0731 on price. The score weighs capability 83%, context window 17%.
- CapabilityMuse Spark 1.3Capabilities Index (ECI): Muse Spark 1.3 158.2 · GPT-6 Luna 156.5 · DeepSeek V4 Flash 0731 154.4
- Lowest priceDeepSeek V4 Flash 0731DeepSeek V4 Flash 0731 $0.18 · GPT-6 Luna $0.20 · Muse Spark 1.3 $2.00 per 1M tokens (3:1 blend)
- Longest contextGPT-6 Luna and Muse Spark 1.3GPT-6 Luna 1,050,000 · Muse Spark 1.3 1,048,576 · DeepSeek V4 Flash 0731 1,000,000 tokens
- Widest inputsMuse Spark 1.3DeepSeek V4 Flash 0731: Text · GPT-6 Luna: Text, Images, PDFs · Muse Spark 1.3: Text, Images, PDFs, Audio, Video
- Self-hostingDeepSeek V4 Flash 0731Publishes downloadable weights (MIT)
| Measure | Weight | DeepSeek V4 Flash 0731 | GPT-6 Luna | Muse Spark 1.3 |
|---|---|---|---|---|
| CapabilityCapabilities Index (ECI) | 59% | 82 | 84 | 86 |
| Price | 29% | 81 | 79 | 34 |
| Context window | 12% | 64 | 65 | 65 |
| Overall | 100% | 80/100 | 80/100 | 68/100 |
Left out because at least one model lacks the data: inputs & features. The remaining weights were rescaled.
| Measure | Weight | DeepSeek V4 Flash 0731 | GPT-6 Luna | Muse Spark 1.3 |
|---|---|---|---|---|
| CapabilityCapabilities Index (ECI) | 83% | 82 | 84 | 86 |
| Context window | 17% | 64 | 65 | 65 |
| Overall | 100% | 79/100 | 81/100 | 82/100 |
Left out because at least one model lacks the data: inputs & features. The remaining weights were rescaled.
Every spec in one table
Highlighted cells lead their row. Unavailable specifications are marked “Not published”. Publisher results include their evaluation settings and source.
| Specification | |||
|---|---|---|---|
| Capability | |||
| Capabilities Index (ECI) | 154.4 | 156.5 | 158.2 (best) |
| ECI rank | #53 of 159 | #38 of 159 | #28 of 159 (best) |
| ECI confidence interval | 152.6–156.4 | 154.6–158.5 | 156.2–160.5 |
| ECI evaluation date | Jul 31, 2026 | Sep 22, 2026 | Sep 2, 2026 |
| Documented capabilities | Official DeepSeek V4 Flash release with enhanced agentic capabilities and integrated DSpark speculative decoding | OpenAI's most efficient model for focused, high-volume tasks | Muse Spark 1.3 is a multimodal reasoning model from Meta for long-running agentic, multi-agent, and coding workflows. It improves long-horizon agent collaboration, instruction following, and coding efficiency relative to Muse Spark 1.2. |
| Model parameters | Not published | Not published | Not published |
| Output modalities | Text | Text | Text |
| GPQA DiamondGraduate-level science questions | 91.0%max (best) | 90.5%max | Not published |
| FrontierMath Tiers 1–3Research-level mathematics | 57.5%max | 79.0%max (best) | 74.4%xhigh |
| OTIS Mock AIME 2024–2025Competition mathematics | 94.4%max | 98.9%max | 99.2%xhigh (best) |
| SimpleQA VerifiedShort factual questions | 33.6%max | 41.4%max (best) | Not published |
| Publisher-reported benchmarks | |||
| Terminal-Bench 2.1max | 82.7pass@1Reported result | Not published | Not published |
| NL2Repomax effort | 54.2resolve rateReported resultJul 31, 2026 | Not published | Not published |
| CyberGymmax effort | 76.7scoreReported resultJul 31, 2026 | Not published | Not published |
| DeepSWEmax effort | 54.4resolve rateReported resultJul 31, 2026 | Not published | Not published |
| Toolathlon-Verifiedmax effort | 70.3scoreReported resultJul 31, 2026 | Not published | Not published |
| Agents' Last Exammax effort | 25.2scoreReported resultJul 31, 2026 | Not published | Not published |
| AutomationBench · publicmax effort | 25.1success rateReported resultJul 31, 2026 | Not published | Not published |
| DSBench-FullStack · internalmax effort | 68.7scoreReported resultJul 31, 2026 | Not published | Not published |
| DSBench-Hard · internalmax effort | 59.6scoreReported resultJul 31, 2026 | Not published | Not published |
| DeepSWE 1.1max effort | Not published | 66.6scoreReported resultSep 22, 2026 | Not published |
| GDPval-AA 2max effort | Not published | Not published | 1754EloReported resultSep 2, 2026 |
| JobBenchmax effort | Not published | Not published | 64.9mean rubric scoreReported resultSep 2, 2026 |
| OSWorld 2.0 · 08.08max effort | Not published | Not published | 66.9mean partial scoreReported resultSep 2, 2026 |
| OSWorld 2.0 · 08.08max effort | Not published | Not published | 32binary completion rateReported resultSep 2, 2026 |
| DeepSearchQAmax effort | Not published | Not published | 90.3F1Reported resultSep 2, 2026 |
| Agentic IF Index · Meta internalmax effort | Not published | Not published | 57.8scoreReported resultSep 2, 2026 |
| AutomationBench · public v3 task setmax effort | Not published | Not published | 49.6pass@1Reported resultSep 2, 2026 |
| MRCR 2 · 8-needle, 256K-512Kmax effort | Not published | Not published | 98.5mean sequence-match ratioReported resultSep 2, 2026 |
| MRCR 2 · 8-needle, 512K-1Mmax effort | Not published | Not published | 98.1mean sequence-match ratioReported resultSep 2, 2026 |
| DeepSWE 1.1max effort | Not published | Not published | 75.4pass@1Reported resultSep 2, 2026 |
| SWE-Atlas Codebase QnAmax effort | Not published | Not published | 59.4pass@1Reported resultSep 2, 2026 |
| Terminal-Bench 2.1max effort | Not published | Not published | 88.8pass@1Reported resultSep 2, 2026 |
| Price per million tokens | |||
| Input | $0.14 | $0.10 (best) | $1.25 |
| Output | $0.30 (best) | $0.50 | $4.25 |
| Cached input | Not published | $0.01 (best) | $0.15 |
| Blended (3:1) | $0.18 (best) | $0.20 | $2.00 |
| Long-context rate | Same rate | Over 272K: $0.20 / $0.75 | Same rate |
| Price source | Median of 65 providers | Official OpenAI API | Official Meta API |
| Limits | |||
| Context window | 1,000,000 tokens | 1,050,000 tokens (best) | 1,048,576 tokens |
| Max input | Not published | 922,000 tokens | Not published |
| Max output | 384,000 tokens (best) | 128,000 tokens | 131,072 tokens |
| Inputs and features | |||
| Text | Yes | Yes | Yes |
| Images | No | Yes | Yes |
| PDFs | No | Yes | Yes |
| Audio | No | No | Yes |
| Video | No | No | Yes |
| Reasoning | Yes | Yeslow · medium · high · xhigh · max | Yesminimal · low · medium · high · xhigh · max |
| Tool calling | Yes | Yes | Yes |
| Structured output | Not published | Yes | Yes |
| File attachments | No | Yes | Yes |
| Temperature control | Yes | Yes | Yes |
| Reasoning control | always | effort | effort |
| Availability | |||
| Weights | OpenMIT | Proprietary | Proprietary |
| API model ID | Not published | gpt-6-luna | muse-spark-1.3 |
| API providers | 66 (best) | 49 | 18 |
| Released | Jul 31, 2026 | Sep 22, 2026 | Sep 2, 2026 |
| Knowledge cutoff | May 2025 | May 18, 2026 | Not published |
| Publisher documentation | Model weights | Publisher documentation | Publisher documentation |
| Source metadata updated | Jul 31, 2026 | Sep 22, 2026 | Sep 2, 2026 |
| Data sources | models.dev · Epoch AI | models.dev · Epoch AI | models.dev · Epoch AI |
| Catalog checked | Nov 22, 2026 | Nov 22, 2026 | Nov 22, 2026 |
What would a month cost?
Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.
DeepSeek V4 Flash 0731$2.00
GPT-6 Luna$2.00
Muse Spark 1.3$21.00
Which should you choose?
Which is better: DeepSeek V4 Flash 0731, GPT-6 Luna or Muse Spark 1.3?
It is close. Our weighted score puts them within a point (GPT-6 Luna 80/100, DeepSeek V4 Flash 0731 80/100, Muse Spark 1.3 68/100), so choose by what matters most for your work: Muse Spark 1.3 for raw capability and DeepSeek V4 Flash 0731 on price. The score weighs capability 59%, price 29%, context window 12%.It is close. Our weighted score puts them within 1 points (Muse Spark 1.3 82/100, GPT-6 Luna 81/100, DeepSeek V4 Flash 0731 79/100), so choose by what matters most for your work: Muse Spark 1.3 for raw capability and DeepSeek V4 Flash 0731 on price. The score weighs capability 83%, context window 17%.
Which is cheaper, DeepSeek V4 Flash 0731, GPT-6 Luna or Muse Spark 1.3?
DeepSeek V4 Flash 0731 is cheaper at $0.14 input / $0.30 output per million tokens (median across 65 API providers). GPT-6 Luna costs $0.10 input / $0.50 output per million tokens (official OpenAI API price); Muse Spark 1.3 costs $1.25 input / $4.25 output per million tokens (official Meta API price). At a typical mix of three input tokens to one output token, that is $0.18 per million tokens for DeepSeek V4 Flash 0731 versus $0.20 for GPT-6 Luna (1.1× as much) and $2.00 for Muse Spark 1.3 (11× as much).
Which scores higher on benchmarks?
Muse Spark 1.3 scores higher on the Capabilities Index (ECI): Muse Spark 1.3 158.2 (#28 of 159), GPT-6 Luna 156.5 (#38 of 159) and DeepSeek V4 Flash 0731 154.4 (#53 of 159). The confidence ranges of the top two overlap (156.2–160.5 vs 154.6–158.5), so treat the gap as small. On individual benchmarks: FrontierMath Tiers 1–3 — GPT-6 Luna 79.0%, Muse Spark 1.3 74.4%, DeepSeek V4 Flash 0731 57.5%; OTIS Mock AIME 2024–2025 — Muse Spark 1.3 99.2%, GPT-6 Luna 98.9%, DeepSeek V4 Flash 0731 94.4%.
Which is better for coding?
There are no published SWE-bench Verified results for DeepSeek V4 Flash 0731, GPT-6 Luna and Muse Spark 1.3 yet, so there is no like-for-like coding score. On overall capability, Muse Spark 1.3 leads, which tends to carry over to coding, but test on your own codebase. All three support tool calling for agent workflows.
Which has the bigger context window?
GPT-6 Luna and Muse Spark 1.3 have the largest context windows (1,050,000 and 1,048,576 tokens), against 1,000,000 for DeepSeek V4 Flash 0731. Maximum output per response: DeepSeek V4 Flash 0731 up to 384,000, GPT-6 Luna up to 128,000, Muse Spark 1.3 up to 131,072 tokens.
Which can read images, PDFs, audio or video?
DeepSeek V4 Flash 0731 accepts text; GPT-6 Luna accepts text, images and PDFs; Muse Spark 1.3 accepts text, images, PDFs, audio and video. Muse Spark 1.3 handles the widest range of inputs.
Are any of these open source?
DeepSeek V4 Flash 0731 publishes its weights (MIT) and can be self-hosted; GPT-6 Luna and Muse Spark 1.3 is proprietary.
Which is newer?
GPT-6 Luna is the newest, released Sep 22, 2026. Muse Spark 1.3 came out Sep 2, 2026; DeepSeek V4 Flash 0731 came out Jul 31, 2026. Knowledge cutoff: DeepSeek V4 Flash 0731 May 2025, GPT-6 Luna May 18, 2026.
How do you decide the winner?
This rating uses capability (59%, independent benchmark results); price (29%, blended price per million tokens (3 input : 1 output), log scale); context window (12%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.This rating uses capability (83%, independent benchmark results); context window (17%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.