Kimi K3 vs GPT-6 Astra vs GPT-5.4
Too close to call on our weighted score (GPT-6 Astra 66, Kimi K3 65, GPT-5.4 65). The right pick depends on what you value most.GPT-6 Astra comes out ahead, 89 to 82 and 82 on our weighted score, though GPT-5.4 is 3.6× cheaper per token.
Moonshot AI
Kimi K3
65/100- ECI158.6
- Price$3.00 / $15.00
- Context1.05M
OpenAI
GPT-6 Astra
66/100- ECI169.2
- Price$10.00 / $50.00
- Context1.05M
OpenAI
GPT-5.4
65/100- ECI157.9
- Price$2.50 / $15.00
- Context1.05M
Moonshot AI
Kimi K3
82/100- ECI158.6
- Price$3.00 / $15.00
- Context1.05M
- Our pick
OpenAI
GPT-6 Astra
89/100- ECI169.2
- Price$10.00 / $50.00
- Context1.05M
OpenAI
GPT-5.4
82/100- ECI157.9
- Price$2.50 / $15.00
- Context1.05M
Too close to callGPT-6 Astra is our pick
It is close. Our weighted score puts them within 2 points (GPT-6 Astra 66/100, Kimi K3 65/100, GPT-5.4 65/100), so choose by what matters most for your work: GPT-6 Astra for raw capability and GPT-5.4 on price. The score weighs capability 50%, price 25%, inputs & features 15%, context window 10%.GPT-6 Astra is the better all-round choice, scoring 89/100 against Kimi K3 (82) and GPT-5.4 (82). It leads on capability. The score weighs capability 67%, inputs & features 20%, context window 13%.
- CapabilityGPT-6 AstraCapabilities Index (ECI): GPT-6 Astra 169.2 · Kimi K3 158.6 · GPT-5.4 157.9
- Lowest priceGPT-5.4GPT-5.4 $5.63 · Kimi K3 $6.00 · GPT-6 Astra $20.00 per 1M tokens (3:1 blend)
- Longest contextAbout the sameGPT-6 Astra 1,050,000 · GPT-5.4 1,050,000 · Kimi K3 1,048,576 tokens
- Widest inputsSame inputsKimi K3: Text, Images, Video · GPT-6 Astra: Text, Images, PDFs · GPT-5.4: Text, Images, PDFs
- Self-hostingKimi K3Publishes downloadable weights
| Measure | Weight | Kimi K3 | GPT-6 Astra | GPT-5.4 |
|---|---|---|---|---|
| CapabilityCapabilities Index (ECI) | 50% | 86 | 96 | 85 |
| Price | 25% | 12 | 0 | 14 |
| Inputs & features | 15% | 80 | 80 | 80 |
| Context window | 10% | 65 | 65 | 65 |
| Overall | 100% | 65/100 | 66/100 | 65/100 |
| Measure | Weight | Kimi K3 | GPT-6 Astra | GPT-5.4 |
|---|---|---|---|---|
| CapabilityCapabilities Index (ECI) | 67% | 86 | 96 | 85 |
| Inputs & features | 20% | 80 | 80 | 80 |
| Context window | 13% | 65 | 65 | 65 |
| Overall | 100% | 82/100 | 89/100 | 82/100 |
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) | 158.6 | 169.2 (best) | 157.9 |
| ECI rank | #26 of 159 | #4 of 159 (best) | #29 of 159 |
| ECI confidence interval | 156.7–160.8 | 166.4–172.5 | 156.0–160.0 |
| ECI evaluation date | Jul 16, 2026 | Sep 3, 2026 | Mar 5, 2026 |
| Documented capabilities | Multimodal Kimi model with 1M context and toggleable max-effort thinking for long-horizon agent work | GPT-6 Astra is OpenAI's most capable model for complex reasoning, coding, computer use, research, and document creation. | Agent-ready GPT for coding and computer-use workflows at a lower cost |
| Model parameters | Not published | Not published | Not published |
| Output modalities | Text | Text | Text |
| GPQA DiamondGraduate-level science questions | 93.1%max | 95.8%max (best) | 93.3%xhigh |
| FrontierMath Tiers 1–3Research-level mathematics | 72.2%max | 93.7%max (best) | 78.6%xhigh |
| OTIS Mock AIME 2024–2025Competition mathematics | 97.2%max | 100%max (best) | 97.8%high |
| SWE-bench VerifiedFixing real GitHub issues | Not published | Not published | 76.9%high |
| SimpleQA VerifiedShort factual questions | 50.6%max | 75.6%max (best) | 45.1%xhigh |
| Publisher-reported benchmarks | |||
| DeepSWE 1.1max effort | 67.5resolve rateReported resultJul 16, 2026 | Not published | Not published |
| Terminal-Bench 2.1max effort | 88.3accuracyReported resultJul 16, 2026 | Not published | Not published |
| FrontierSWEmax effort | 81.2dominance scoreReported resultJul 16, 2026 | Not published | Not published |
| Program Benchmax effort | 77.8scoreReported resultJul 16, 2026 | Not published | Not published |
| SWE Marathon 1.1max effort | 42resolve rateReported resultJul 16, 2026 | Not published | Not published |
| GDPval-AA v2max effort | 1668EloReported resultJul 16, 2026 | Not published | Not published |
| AA-Briefcasemax effort | 1548EloReported resultJul 16, 2026 | Not published | Not published |
| AutomationBench · 600-task public subsetmax effort | 30.8success rateReported resultJul 16, 2026 | Not published | Not published |
| JobBenchmax effort | 52.9scoreReported resultJul 16, 2026 | Not published | Not published |
| SpreadsheetBench 2max effort | 34.8scoreReported resultJul 16, 2026 | Not published | Not published |
| BrowseCompmax effort, context compaction | 91.2accuracyReported resultJul 16, 2026 | Not published | Not published |
| CharXiv Reasoningmax effort, with tools | 91.3accuracyReported resultJul 16, 2026 | Not published | Not published |
| ZeroBenchmax effort, with tools | 41pass@5Reported resultJul 16, 2026 | Not published | Not published |
| Agents' Last ExamPublisher-reported result | Not published | 59.3Reported resultSep 3, 2026 | Not published |
| OSWorld 2.0 · V2-Offline (v2026.08.08)Publisher-reported result | Not published | 72.6partial scoreReported resultSep 3, 2026 | Not published |
| ScreenSpot-Prono tools | Not published | 92.7accuracyReported resultSep 3, 2026 | Not published |
| AutomationBenchPublisher-reported result | Not published | 41.4success rateReported resultSep 3, 2026 | Not published |
| BenchCADwith tools | Not published | 95.9geometric overlapReported resultSep 3, 2026 | Not published |
| BrowseCompPublisher-reported result | Not published | 91.5accuracyReported resultSep 3, 2026 | 82.7accuracyReported resultApr 23, 2026 |
| Terminal-Bench 4.0Publisher-reported result | Not published | 57.9success rateReported resultSep 3, 2026 | Not published |
| DeepSWE 1.1Publisher-reported result | Not published | 74.1resolve rateReported resultSep 3, 2026 | Not published |
| FrontierCode 1.1 · ExtendedPublisher-reported result | Not published | 64.5scoreReported resultSep 3, 2026 | Not published |
| Terminal-Bench Science 0.1Publisher-reported result | Not published | 64.6success rateReported resultSep 3, 2026 | Not published |
| FrontierMath v2 · Tier 4Publisher-reported result | Not published | 97.6accuracyReported resultSep 3, 2026 | Not published |
| GPQA DiamondPublisher-reported result | Not published | 96accuracyReported resultSep 3, 2026 | 92.8accuracyReported resultApr 23, 2026 |
| Humanity's Last Examwith tools | Not published | 57.2accuracyReported resultSep 3, 2026 | 52.1accuracyReported resultApr 23, 2026 |
| ExploitBenchwithout production safeguards | Not published | 100success rateReported resultSep 3, 2026 | Not published |
| SRE-Benchsingle attempt | Not published | 88success rateReported resultSep 3, 2026 | Not published |
| ARC-AGI-3Publisher-reported result | Not published | 99.9RHAEReported resultSep 3, 2026 | Not published |
| Artificial Analysis Intelligence Index 4.1.1max | Not published | 61index scoreReported resultSep 3, 2026 | Not published |
| Artificial Analysis Coding Agent Index 1.4max | Not published | 67index scoreReported resultSep 3, 2026 | Not published |
| SWE-Bench Pro · publicPublisher-reported result | Not published | Not published | 59.1resolve rateReported result |
| SWE-Atlas Codebase QnAPublisher-reported result | Not published | Not published | 40.8scoreReported result |
| SWE-Atlas RefactoringPublisher-reported result | Not published | Not published | 44.29scoreReported result |
| SWE-Atlas Test WritingPublisher-reported result | Not published | Not published | 44.36scoreReported result |
| Artificial Analysis Coding Agent Indexmedium | Not published | Not published | 53.6average pass@1Reported result |
| SWE-Atlas Codebase QnAmedium | Not published | Not published | 72.4pass@1Reported result |
| SWE-Bench Pro · hard-aamedium | Not published | Not published | 18.4pass@1Reported result |
| Terminal-Bench 2.1medium | Not published | Not published | 69.8pass@1Reported result |
| Terminal-Bench 2.0Publisher-reported result | Not published | Not published | 75.1success rateReported resultApr 23, 2026 |
| Humanity's Last Examno tools | Not published | Not published | 39.8accuracyReported resultApr 23, 2026 |
| OSWorld-VerifiedPublisher-reported result | Not published | Not published | 75success rateReported resultApr 23, 2026 |
| GDPvalPublisher-reported result | Not published | Not published | 83wins or tiesReported resultApr 23, 2026 |
| ARC-AGI-2Verified | Not published | Not published | 73.3accuracyReported resultApr 23, 2026 |
| FrontierMath · Tier 1-3Publisher-reported result | Not published | Not published | 47.6accuracyReported resultApr 23, 2026 |
| FrontierMath · Tier 4Publisher-reported result | Not published | Not published | 27.1accuracyReported resultApr 23, 2026 |
| MMMU Prono tools | Not published | Not published | 81.2accuracyReported resultApr 23, 2026 |
| Price per million tokens | |||
| Input | $3.00 | $10.00 | $2.50 (best) |
| Output | $15.00 (best) | $50.00 | $15.00 (best) |
| Cached input | $0.30 | $1.00 | $0.25 (best) |
| Blended (3:1) | $6.00 | $20.00 | $5.63 (best) |
| Long-context rate | Same rate | Over 272K: $20.00 / $75.00 | Over 272K: $5.00 / $22.50 |
| Price source | Official Moonshot AI API | Official OpenAI API | Official OpenAI API |
| Limits | |||
| Context window | 1,048,576 tokens | 1,050,000 tokens (best) | 1,050,000 tokens (best) |
| Max input | Not published | 922,000 tokens | 922,000 tokens |
| Max output | 1,048,576 tokens (best) | 128,000 tokens | 128,000 tokens |
| Inputs and features | |||
| Text | Yes | Yes | Yes |
| Images | Yes | Yes | Yes |
| PDFs | No | Yes | Yes |
| Audio | No | No | No |
| Video | Yes | No | No |
| Reasoning | Yeslow · high · max | Yeslow · medium · high · xhigh · max | Yeslow · medium · high · xhigh |
| Tool calling | Yes | Yes | Yes |
| Structured output | Yes | Yes | Yes |
| File attachments | Yes | Yes | Yes |
| Temperature control | No | No | Yes |
| Reasoning control | effort | effort | effort |
| Availability | |||
| Weights | Open | Proprietary | Proprietary |
| API model ID | kimi-k3 | gpt-6-astra | gpt-5.4 |
| API providers | 93 (best) | 56 | 50 |
| Released | Jul 16, 2026 | Sep 4, 2026 | Mar 5, 2026 |
| Knowledge cutoff | Not published | Apr 30, 2026 | Aug 31, 2025 |
| Publisher documentation | Publisher documentation · Model weights | Publisher documentation | Publisher documentation |
| Source metadata updated | Jul 16, 2026 | Sep 4, 2026 | Mar 5, 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.
Kimi K3$60.00
GPT-6 Astra$200.00
GPT-5.4$55.00
Which should you choose?
Which is better: Kimi K3, GPT-6 Astra or GPT-5.4?
It is close. Our weighted score puts them within 2 points (GPT-6 Astra 66/100, Kimi K3 65/100, GPT-5.4 65/100), so choose by what matters most for your work: GPT-6 Astra for raw capability and GPT-5.4 on price. The score weighs capability 50%, price 25%, inputs & features 15%, context window 10%.GPT-6 Astra is the better all-round choice, scoring 89/100 against Kimi K3 (82) and GPT-5.4 (82). It leads on capability. The score weighs capability 67%, inputs & features 20%, context window 13%.
Which is cheaper, Kimi K3, GPT-6 Astra or GPT-5.4?
GPT-5.4 is cheaper at $2.50 input / $15.00 output per million tokens (official OpenAI API price). Kimi K3 costs $3.00 input / $15.00 output per million tokens (official Moonshot AI API price); GPT-6 Astra costs $10.00 input / $50.00 output per million tokens (official OpenAI API price). At a typical mix of three input tokens to one output token, that is $5.63 per million tokens for GPT-5.4 versus $6.00 for Kimi K3 (1.1× as much) and $20.00 for GPT-6 Astra (3.6× as much).
Which scores higher on benchmarks?
GPT-6 Astra scores higher on the Capabilities Index (ECI): GPT-6 Astra 169.2 (#4 of 159), Kimi K3 158.6 (#26 of 159) and GPT-5.4 157.9 (#29 of 159). Their confidence ranges do not overlap (166.4–172.5 vs 156.7–160.8), so the gap is a real one. On individual benchmarks: GPQA Diamond — GPT-6 Astra 95.8%, GPT-5.4 93.3%, Kimi K3 93.1%; FrontierMath Tiers 1–3 — GPT-6 Astra 93.7%, GPT-5.4 78.6%, Kimi K3 72.2%; OTIS Mock AIME 2024–2025 — GPT-6 Astra 100%, GPT-5.4 97.8%, Kimi K3 97.2%; SimpleQA Verified — GPT-6 Astra 75.6%, Kimi K3 50.6%, GPT-5.4 45.1%.
Which is better for coding?
There are no published SWE-bench Verified results for Kimi K3 and GPT-6 Astra yet, so there is no like-for-like coding score. On overall capability, GPT-6 Astra 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?
Their context windows are effectively the same size: Kimi K3 1,048,576, GPT-6 Astra 1,050,000 and GPT-5.4 1,050,000 tokens. Maximum output per response: Kimi K3 up to 1,048,576, GPT-6 Astra up to 128,000, GPT-5.4 up to 128,000 tokens.
Which can read images, PDFs, audio or video?
Kimi K3 accepts text, images and video; GPT-6 Astra accepts text, images and PDFs; GPT-5.4 accepts text, images and PDFs. They handle the same number of input types.
Are any of these open source?
Kimi K3 publishes its weights and can be self-hosted; GPT-6 Astra and GPT-5.4 is proprietary.
Which is newer?
GPT-6 Astra is the newest, released Sep 4, 2026. Kimi K3 came out Jul 16, 2026; GPT-5.4 came out Mar 5, 2026. Knowledge cutoff: GPT-6 Astra Apr 30, 2026, GPT-5.4 Aug 31, 2025.
How do you decide the winner?
This rating uses capability (50%, independent benchmark results); price (25%, blended price per million tokens (3 input : 1 output), log scale); inputs & features (15%, image, PDF, audio and video input, tool calling, structured output and reasoning); context window (10%, 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 (67%, independent benchmark results); inputs & features (20%, image, PDF, audio and video input, tool calling, structured output and reasoning); context window (13%, 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.