Our models.Against every flagship.
The four open-weight text models on ccsio.ai against the last five flagship generations from OpenAI, Anthropic and Google — plus six vision, speech and embedding capabilities included in every plan. Published scores only, each linked to its source.
Chat & reasoning
Alibaba QwenChatAvailable now
40Intelligence Index
Total parameters180B
Active per token6B
ArchitectureMoE
Context1M
Max output128K (256K thinking)
Inputtext
Released2026-08
Model card ↗GoogleChatAvailable now
17Intelligence Index
Total parameters25.2B
Active per token3.8B
ArchitectureMoE
Context256K
Max output—
Inputtext + image
Released2026-04
Model card ↗DeepSeekReasoningComing soon
39Intelligence Index
Total parameters552B
Active per token8B / 16B
ArchitectureMoE
Context1M
Max output384K
Inputtext + image
Released2026-09
Model card ↗Z.aiReasoningComing soon
45Intelligence Index
Total parameters744B
Active per token40B
ArchitectureMoE
Context1M
Max output128K
Inputtext
Released2026-08
Model card ↗Vision & audio
FLUX.2 [klein] 4B
Black Forest LabsImagesAvailable now
—Intelligence Index
Total parameters4B
Active per token—
ArchitectureRectified flow transformer
Context1024×1024 px
Max output—
Inputtext → image
Released2026-06
Model card ↗Whisper large v3
OpenAITranscriptionAvailable now
—Intelligence Index
Total parameters1.55B
Active per token—
ArchitectureEncoder–decoder ASR
Context30s per segment
Max output—
Inputaudio → text
Released2023-11
Model card ↗Kokoro
HexgradSpeechAvailable now
—Intelligence Index
Total parameters82M
Active per token—
ArchitectureStyleTTS 2 · ISTFTNet
Context—
Max output—
Inputtext → audio
Released2025-01
Model card ↗Retrieval & decisions
Clef
CloudflareDecisionsAvailable now
—Intelligence Index
Total parameters27B
Active per token—
ArchitectureMultimodal decision model
Context16K
Max output—
Inputtext + image + video
Released2026-10
Model card ↗mxbai-embed-large v1
Mixedbread AIEmbeddingsAvailable now
—Intelligence Index
Total parameters0.3B
Active per token—
ArchitectureBERT-large embedder (AnglE)
Context512
Max output—
Inputtext
Released2024-03
Model card ↗Qwen3 Embedding 0.6B
Alibaba QwenEmbeddingsAvailable now
—Intelligence Index
Total parameters0.6B
Active per token—
ArchitectureQwen3 dense embedder
Context32K
Max output—
Inputtext
Released2025-06
Model card ↗Intelligence Index · Artificial Analysis · independent, one index version for every model (v4.3.2)
Higher is better
- 1Claude Opus 5.558
- 2GPT-6 Astra53
- 3Gemini 4 Argon53
- 4Claude Opus 551
- 5GPT-5.6 Sol47
- 6GLM 5.3Coming soon45
- 7Claude Opus 4.842
- 8Claude Opus 4.741
- 9Qwen 3.8 Flash Next40
- 10DeepSeek 4.1 FlashComing soon39
- 11GPT-5.439
- 12GPT-5.538
- 13Claude Opus 4.632
- 14GPT-5.230
- 15Gemini 3.1 Pro30
- 16Gemini 3 Pro28
- 17Gemma 417
- 18Gemini 2.5 Pro16
- 19Gemini 2.0 Pro Experimental9
Qwen 3.8 Flash Next: 8 of 15 flagships already score lower
Measured with the same independent test for every model: the Artificial Analysis Intelligence Index. Orange: our model scores the same or higher. Grey: the flagship scores higher.
OpenAI
- GPT-6 Astra2026-09 · AA 53Ahead of us−13 pts
- GPT-5.6 Sol2026-07 · AA 47Ahead of us−7 pts
- GPT-5.52026-04 · AA 38✓ Beaten+2 pts
- GPT-5.42026-03 · AA 39✓ Beaten+1 pts
- GPT-5.22025-12 · AA 30✓ Beaten+10 pts
Anthropic
- Claude Opus 5.52026-09 · AA 58Ahead of us−18 pts
- Claude Opus 52026-07 · AA 51Ahead of us−11 pts
- Claude Opus 4.82026-05 · AA 42Ahead of us−2 pts
- Claude Opus 4.72026-04 · AA 41Ahead of us−1 pts
- Claude Opus 4.62026-02 · AA 32✓ Beaten+8 pts
- Gemini 4 Argon2026-09 · AA 53Ahead of us−13 pts
- Gemini 3.1 Pro2026-02 · AA 30✓ Beaten+10 pts
- Gemini 3 Pro2025-11 · AA 28✓ Beaten+12 pts
- Gemini 2.5 Pro2025-06 · AA 16✓ Beaten+24 pts
- Gemini 2.0 Pro Experimental2025-02 · AA 9✓ Beaten+31 pts
Qwen 3.8 Flash Next · AA 40 · Artificial Analysis v4.3.2
Every published score, with its source
Click any number to open its source. A dash means no published number for that exact version; we do not estimate.
| Model | Released | Intelligence Index | Humanity's Last Exam | GPQA Diamond | DeepSWE | SWE-bench Verified | LiveCodeBench | MMLU-Pro | AIME 2025 |
|---|---|---|---|---|---|---|---|---|---|
| Qwen 3.8 Flash NextChat | 2026-08 | 40 | 35.9 | 91.7 | 58.7 | — | 91.9 | — | — |
| Gemma 4Chat | 2026-04 | 17 | 8.7 | 82.3 | — | — | 77.1 | 82.6 | — |
| DeepSeek 4.1 FlashReasoningComing soon | 2026-09 | 39 | 36.8 | 90.9 | 74.2 | — | — | — | — |
| GLM 5.3ReasoningComing soon | 2026-08 | 45 | 42 | 88.1 | 66.9 | — | 53.5* | 87.8* | 99.3* |
| ClefDecisions | 2026-10 | — | — | — | — | — | — | — | — |
| FLUX.2 [klein] 4BImages | 2026-06 | — | — | — | — | — | — | — | — |
| Whisper large v3Transcription | 2023-11 | — | — | — | — | — | — | — | — |
| mxbai-embed-large v1Embeddings | 2024-03 | — | — | — | — | — | — | — | — |
| Qwen3 Embedding 0.6BEmbeddings | 2025-06 | — | — | — | — | — | — | — | — |
| KokoroSpeech | 2025-01 | — | — | — | — | — | — | — | — |
| GPT-6 Astra | 2026-09 | 53 | 59.1 | 96 | 74.1 | — | — | — | — |
| GPT-5.6 Sol | 2026-07 | 47 | 55.5 | 94.6 | 72.7 | — | — | — | — |
| GPT-5.5 | 2026-04 | 38 | 54.9 | 93.6 | 67* | — | — | — | — |
| GPT-5.4 | 2026-03 | 39 | 39.8 | 93 | 51.8* | — | — | — | — |
| GPT-5.2 | 2025-12 | 30 | 34.5 | 92.4 | — | 80 | — | — | 100 |
| Claude Opus 5.5 | 2026-09 | 58 | 64.4 | — | 74.2 | — | — | — | — |
| Claude Opus 5 | 2026-07 | 51 | 56.3 | — | 73.6* | 96* | — | — | — |
| Claude Opus 4.8 | 2026-05 | 42 | 49.8 | — | 59* | 88.6* | — | — | — |
| Claude Opus 4.7 | 2026-04 | 41 | 46.9 | — | — | 87.6* | — | — | — |
| Claude Opus 4.6 | 2026-02 | 32 | 40 | 91.3 | — | 80.8 | — | — | — |
| Gemini 4 Argon | 2026-09 | 53 | — | — | 77.9 | — | — | — | — |
| Gemini 3.1 Pro | 2026-02 | 30 | 44.4 | 94.3 | 11.7* | 80.6 | — | — | — |
| Gemini 3 Pro | 2025-11 | 28 | 37.5 | 91.9 | — | 76.2 | — | — | — |
| Gemini 2.5 Pro | 2025-06 | 16 | 21.6 | 86.4 | — | 59.6 | 74.2 | — | 88 |
| Gemini 2.0 Pro Experimental | 2025-02 | 9 | — | 65.7 | — | — | 34.7 | 80.5 | — |
* independent reproduction (GLM 5.3) · SWE-bench leaderboard (SWE-bench: Claude Opus 4.7–5) · DeepSWE leaderboard (Datacurve) (DeepSWE: GPT-5.4, GPT-5.5, Opus 4.8, Opus 5, Gemini 3.1 Pro)
Read this before you quote a number
- Vendors report their own scores, usually at maximum reasoning effort, and settings differ. The Intelligence Index is the only column measured by one independent party for every model.
- All scores and specs of our models come from their official model cards. Qwen 3.8 Flash Next: 262K native context, 1M with YaRN. Qwen 3.8 Flash Next HLE is judged by GPT-4o; GLM 5.3 MMLU-Pro, AIME and LiveCodeBench are NVIDIA reproductions. DeepSeek 4.1 Flash and GLM 5.3 arrive on ccsio.ai in late Q4 2026.
- Several vendors no longer report SWE-bench Verified because of contamination concerns, so that column is incomplete on purpose. DeepSWE v1.1 is the agentic coding benchmark that most current models publish.
- Gemini 4 Argon is in a limited rollout. Gemini 2.0 Pro was an experimental release.
- Clef, FLUX.2 [klein], Whisper large v3, mxbai-embed-large v1, Qwen3 Embedding 0.6B and Kokoro are capability models (vision, image, speech, embeddings): text-generation benchmarks do not apply to them, so their rows in the table stay empty. We do not estimate.
Data as of 2026-10-06 · Artificial Analysis v4.3.2
Calculate your costFlagship-level answers. In your region. On infrastructure we operate.
Create your account and pick the region your data must stay in.