Model comparison
GPT-6 Astra vs Kimi K3
OpenAI against Moonshot AI, compared on context, price, and attributed benchmark results.
Recorded profile verification October 1, 2026Individual provider fields may changeEvidence confidence: High — see receipts below
OpenAI
GPT-6 Astra
Frontier
Moonshot AI
Kimi K3
Frontier · Open weights
| Specification | GPT-6 Astra | Kimi K3 |
|---|---|---|
| Provider | ||
| Provider | OpenAI | Moonshot AI |
| Tier | ||
| Tier | Frontier | Frontier |
| Context window | ||
| Context window | Winner: 1.05M | 1.05M |
| Max output | ||
| Max output | 128K | Winner: 1.05M |
| Input / 1M tokens | ||
| Input / 1M tokens | $10 | Winner: $3 |
| Output / 1M tokens | ||
| Output / 1M tokens | $50 | Winner: $15 |
| Weights | ||
| Weights | Closed | Open |
| Parameters | ||
| Parameters | Not verifiedUnverified | 2.8T total / 104B active (MoE) |
| Reasoning levels | ||
| Reasoning levels | low, medium, high, xhigh, max | low, high, max |
| Modalities | ||
| Modalities | text, image | text, image, video |
| License | ||
| License | Not verifiedUnverified | Kimi K3 License (custom) |
| API model id | ||
| API model id | gpt-6-astra | kimi-k3 |
| Released | ||
| Released | September 3, 2026 | July 16, 2026 |
| Artificial Analysis Intelligence Index [max] · source date GPT-6 Astra: 2026-10-01 | ||
| Artificial Analysis Intelligence Index [max] · source date GPT-6 Astra: 2026-10-01 | 53 | Not verifiedUnverified |
| Artificial Analysis Intelligence Index [max] · source date Kimi K3: 2026-10-01 | ||
| Artificial Analysis Intelligence Index [max] · source date Kimi K3: 2026-10-01 | Not verifiedUnverified | 44 |
| DeepSWE 1.1 · source date GPT-6 Astra: 2026-09-03 | ||
| DeepSWE 1.1 · source date GPT-6 Astra: 2026-09-03 | 74.1 | Not verifiedUnverified |
| OSWorld 2.0 (offline) · source date GPT-6 Astra: 2026-09-03 | ||
| OSWorld 2.0 (offline) · source date GPT-6 Astra: 2026-09-03 | 72.6 | Not verifiedUnverified |
| GPQA Diamond · source date GPT-6 Astra: 2026-09-03 | ||
| GPQA Diamond · source date GPT-6 Astra: 2026-09-03 | 96 | Not verifiedUnverified |
| GPQA Diamond [max] · source date Kimi K3: 2026-10-01 | ||
| GPQA Diamond [max] · source date Kimi K3: 2026-10-01 | Not verifiedUnverified | 93.5 |
| Humanity's Last Exam · source date GPT-6 Astra: 2026-09-03 | ||
| Humanity's Last Exam · source date GPT-6 Astra: 2026-09-03 | 57.2 | Not verifiedUnverified |
| Humanity's Last Exam [max] · source date Kimi K3: 2026-10-01 | ||
| Humanity's Last Exam [max] · source date Kimi K3: 2026-10-01 | Not verifiedUnverified | 56 |
| Terminal-Bench 2.1 [max] · source date Kimi K3: 2026-10-01 | ||
| Terminal-Bench 2.1 [max] · source date Kimi K3: 2026-10-01 | Not verifiedUnverified | 88.3 |
| SWE-bench Verified · source date Kimi K3: 2026-10-01 | ||
| SWE-bench Verified · source date Kimi K3: 2026-10-01 | Not verifiedUnverified | 93.4 |
Dates beside benchmark scores refer to the linked source’s publication or capture date unless that source explicitly states when it measured the result. A shared URL alone does not establish comparable runs; unverified setups are marked. Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.
Benchmark receipts
Where each score comes from, and how far it can be compared across models.
- Source date2026-10-01Vals.ai exact Kimi K3 93.40% key-takeaway result; independent bash-only harness, 500-task Verified subset; experiment date unverified
Real GitHub issue resolution: does the model's patch pass the hidden tests.
Comparable with caveatPost-audit vendor claims and pre-audit scores sit on different task trust levels; scaffolding (agent harness, compute budget) also dominates results. Never aggregate across scaffolds.
- Source date2026-10-01Kimi K3 official model-card table (vendor-reported) (HLE-Full with general tools; temperature 1.0; 43.5% separately without tools; exact run date unverified)
Frontier-knowledge ceiling: extremely hard multi-domain questions written to be near-impossible without deep expertise.
Comparable with caveatSubset/tool configurations (text-only vs with-tools) differ between vendors and materially change scores; must match configuration to compare.
- Source date2026-10-01Kimi K3 official model-card table (vendor-reported) (Single-step, temperature 1.0, top-p 0.95; exact run date unverified)
Graduate-level science reasoning: multiple-choice questions written by domain PhDs.
Comparable with caveatDo not treat a fixed two-point band as universal measurement uncertainty; it depends on run count and evaluation setup. Chain-of-thought, prompt, and tool policy must match. The small pool limits fine-grained conclusions, and frontier scores can be near ceiling; Artificial Analysis dropped it from its Intelligence Index in v4.3.
- Source date2026-10-01Kimi K3 official model-card table (vendor-reported) (Kimi Code harness; temperature 1.0, top-p 1.0; exact run date unverified)
Agentic terminal work: multi-step tasks executed in a sandboxed shell environment.
Comparable with caveatNot comparable with Terminal-Bench 3.0 or 4.0 (different task sets) or v1; harness configuration (container, time limits) also shifts results.
- Source date2026-10-01Artificial Analysis (Kimi K3 (max))
Composite index blending reasoning, knowledge, and coding evals into one 0–100 score.
Comparable with caveatOnly same-version scores are comparable. v4.3.2 was rebaselined (the top score fell from 66 on v4.1.1 to ~58), so v4.1.1 figures must not be compared with v4.3.2 figures — check the cited source and its date.
- Source date2026-10-01Artificial Analysis (GPT-6 Astra (max))
Composite index blending reasoning, knowledge, and coding evals into one 0–100 score.
Comparable with caveatOnly same-version scores are comparable. v4.3.2 was rebaselined (the top score fell from 66 on v4.1.1 to ~58), so v4.1.1 figures must not be compared with v4.3.2 figures — check the cited source and its date.
- Source date2026-09-03OpenAI GPT-6 Astra evaluation table (vendor-reported, with tools; maximum across evaluated efforts)
Frontier-knowledge ceiling: extremely hard multi-domain questions written to be near-impossible without deep expertise.
Comparable with caveatSubset/tool configurations (text-only vs with-tools) differ between vendors and materially change scores; must match configuration to compare.
- Source date2026-09-03OpenAI GPT-6 Astra evaluation table (vendor-reported; maximum across evaluated efforts)
Graduate-level science reasoning: multiple-choice questions written by domain PhDs.
Comparable with caveatDo not treat a fixed two-point band as universal measurement uncertainty; it depends on run count and evaluation setup. Chain-of-thought, prompt, and tool policy must match. The small pool limits fine-grained conclusions, and frontier scores can be near ceiling; Artificial Analysis dropped it from its Intelligence Index in v4.3.
- Source date2026-09-03OpenAI (vendor-reported) (v2026.08.08 offline set, partial reward; vendor research/API evaluation, maximum across evaluated efforts)
Computer-use: completing real OS tasks (files, apps, settings) in a sandboxed desktop.
Comparable with caveatOffline vs live interaction modes are NOT comparable; step budgets and screen resolution must match.
- Source date2026-09-03OpenAI (xhigh, mini-swe-agent; matches Datacurve's independent board (74.1% at $4.43/task, rank 1; board updated 2026-09-22))
Agentic software engineering: pass@1 on real repo tasks with reported cost per completed task.
Directly comparable
Pricing tiers
GPT-6 Astra: Standard $10/$50 per MTok (≤272K input); above 272K the full request bills $20/$75. Cached input $1/$2 and cache writes $12.50/$25 for short/long context. Batch/Flex are half Standard; Fast is 2x. Ultrafast is 6x: $60/$300 short, $120/$450 long, through Responses with global/US processing; no EU or other regional inference residency. Regional/FedRAMP processing adds 10% where eligible. Knowledge cutoff Apr 30 2026. Enterprise model access is disabled by default.
Kimi K3: $3/$15 per million input/output tokens; cached input $0.30. Cache writes add $3/MTok for 5-minute TTL or $6/MTok for 1-hour TTL; hits refresh TTL without another write charge. API access requires a successful top-up of at least $1. max_completion_tokens defaults to 131,072 and can be set up to 1,048,576. Kimi K3 uses a custom license; self-hosting has compute costs.
GPT-6 Astra
OpenAI describes GPT-6 Astra as its most capable model for complex reasoning, coding and computer use, and the first it gates at the Preparedness Framework's 'Critical' cybersecurity threshold.
Best for
- Computer-use agents
- Agentic coding
- Hard reasoning
Watch out
OpenAI places Astra at its Critical cybersecurity threshold and disables it by default for Enterprise. Tool calling requires Responses; none reasoning and custom temperature/top_p/logprobs are unsupported. Long prompts above 272K increase the full-request rate. GPT-6.1 Sol has lower base token prices; compare results on your own workload.
Kimi K3
Kimi K3 is Moonshot AI's open-weight multimodal 2.8T/104B MoE model with a 1,048,576-token context.
Best for
- Open-weight frontier work
- Long-context
- Multimodal
Watch out
Thinking is always on; preserve complete assistant reasoning/tool history. Vision input requires base64 or ms:// rather than public image URLs, and temperature/top_p/n/penalties are fixed. Current docs discourage production web-search use. The custom licence requires an agreement for Model-as-a-Service businesses above $20M aggregate licensee/affiliate revenue over consecutive 12 months; commercial products above 100M monthly active users or $20M monthly revenue require Kimi K3 UI attribution. Internal use and official/certified serving routes have separate exemptions. July 16 is Kimi Code's release date, distinct from weights publication.
Benchmark
DeepSWE 1.1 in context
Only GPT-6 Astra has a published DeepSWE 1.1 result. It is shown against the wider field, with cost per completed task alongside the score.
Local leader
GPT-6 Astra [xhigh]
74.1%
Rows shown
28
Best source-matched Pass@1 per model
Snapshot date
2026-10-01
Mirrored from deepswe.datacurve.ai
Better is toward the top-right (higher pass rate, lower cost). X-axis is reversed to match DeepSWE’s public chart. v1.1 uses average cost, output tokens and steps. Legacy v1 transcriptions are excluded here until an exact primary receipt is retained.
| # | Model | Pass@1 | Cost / task | Tokens / task | Steps / task |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra [xhigh] | 74.1% | $4.43 | 30k | 29 |
| 2 | Gemini 3.8 Flash [high] | 73.8% | $2.36 | 143k | 166 |
| 3 | Claude Opus 5 [max] | 73.6% | $11.84 | 118k | 99 |
| 4 | GPT-5.6 Sol [max] | 72.7% | $8.39 | 60k | 61 |
| 5 | Claude Fable 5 [xhigh] | 69.9% | $13.41 | 80k | 68 |
| 6 | GPT-5.6 Terra [max] | 69.6% | $4.95 | 72k | 76 |
| 7 | GLM 5.3 [max] | 69.0% | $3.99 | 80k | 124 |
| 8 | Kimi K3 [max] | 68.5% | $4.65 | 82k | 98 |
| 9 | Grok 4.6 [medium] | 67.5% | $3.45 | 50k | 70 |
| 10 | GPT-5.6 Luna [max] | 67.2% | $3.03 | 73k | 102 |
| 11 | GPT-5.5 [xhigh] | 67.0% | $7.23 | 46k | 82 |
| 12 | Gemini 3.7 Flash [medium] | 65.5% | $2.03 | 94k | 117 |
| 13 | GLM 5.3 Flash [max] | 63.4% | $0.48 | 73k | 123 |
| 14 | DeepSeek V4-Pro [max] | 62.8% | $0.24 | 106k | 155 |
| 15 | Claude Opus 4.8 [max] | 59.0% | $13.22 | 135k | 120 |
| 16 | Qwen3.8-Max [xhigh] | 57.5% | $3.73 | 95k | 111 |
| 17 | Muse Spark 1.2 [xhigh] | 54.9% | $3.70 | 99k | 101 |
| 18 | Claude Sonnet 5 [max] | 53.8% | $26.40 | 214k | 268 |
| 19 | Grok 4.5 [high] | 53.8% | $2.42 | 36k | 61 |
| 20 | DeepSeek V4-Flash [max] | 53.3% | $0.10 | 108k | 153 |
| 21 | Muse Spark 1.1 [xhigh] | 53.3% | $2.36 | 74k | 96 |
| 22 | GPT-5.4 [xhigh] | 51.8% | $5.65 | 71k | 70 |
| 23 | Gemini 3.6 Flash [high] | 46.7% | $4.42 | 96k | 117 |
| 24 | GLM 5.2 [max] | 43.8% | $3.92 | 78k | 129 |
| 25 | Gemini 3.5 Flash [high] | 36.1% | $3.45 | 76k | 105 |
| 26 | Kimi K2.7 Code | 30.5% | $2.82 | 59k | 149 |
| 27 | Claude Sonnet 4.6 [high] | 29.9% | $5.52 | 76k | 134 |
| 28 | Gemini 3.1 Pro [high] | 11.7% | $2.14 | 28k | 76 |
“Best per model” shows each model’s highest source-matched Pass@1 in this capture; switch to all effort levels to see every retained configuration. Efforts differ across models, so this view is not a comparison at matched effort. The source was generated September 22 and captured October 1. Small gaps may not be statistically meaningful — confirm on deepswe.datacurve.ai.
When the cheaper one wins
Kimi K3 is cheaper on output at $15 per million tokens against $50 for GPT-6 Astra — about 3.3×. Use the cheaper tier for classification, extraction, summarisation, and any task where the expensive model’s extra score does not change the accepted output. The expensive one only pays if your hardest task actually fails on the cheap tier. On DeepSWE 1.1, Kimi K3 is 68.5% Pass@1 at $4.65/task versus GPT-6 Astra at 74.1% / $4.43/task. This compares standard short-context token rates, not total cost per accepted task. Include input, cache, context tiers, schedules, retries and review in your pilot.
Run the model pickerEvidence confidence: High
How strong and complete the evidence behind this comparison is — not a prediction of which model is better.
- Pricing verified on both sides — Input and output rates are verified for both models.
- 4/5 core specs verified on both sides — Not published for at least one side: parameter count.
- 1 benchmark setup aligned with differing scores — Matched source, source date, units, effort and recorded run details: Artificial Analysis Intelligence Index.
- Verified within the last 90 days — Older of the two catalog checks was 25 days ago.
- Both models carry source citations — Each side has at least two catalog sources on record.
Source receipts
Each catalog figure was checked against the provider or an independent second source on the date shown.
- GPT-6 Astra: OpenAI model card — whole-profile source check October 1 (accessed 2026-10-01)
- GPT-6 Astra: OpenAI current API pricing and processing tiers (accessed 2026-10-01)
- GPT-6 Astra: OpenAI API changelog — September 3 release; checked October 1 (accessed 2026-10-01)
- GPT-6 Astra: OpenAI — Introducing GPT-6 Astra (accessed 2026-09-03)
- GPT-6 Astra: OpenAI API pricing (gpt-6-astra $10/$50) (accessed 2026-09-03)
- GPT-6 Astra: CNBC — OpenAI launches GPT-6 Astra (accessed 2026-09-03)
- GPT-6 Astra: OpenAI model docs — gpt-6-astra (128K max output, effort levels) (accessed 2026-09-05)
- GPT-6 Astra: OpenAI API pricing (gpt-6-astra ≤272K $10/$50, >272K $20/$75) (accessed 2026-09-26)
- Kimi K3: Whole-profile primary source check October 1 (accessed 2026-10-01)
- Kimi K3: Whole-profile primary source check October 1 (accessed 2026-10-01)
- Kimi K3: Kimi Code dated release history (K3 July 16; distinct from weights publication) (accessed 2026-10-01)
- Kimi K3: Kimi K3 API pricing — native cache-write TTL charges (accessed 2026-10-01)
- Kimi K3: Exact Kimi K3 custom license (accessed 2026-10-01)
- Kimi K3: HuggingFace — Kimi K3 (accessed 2026-08-29)
- Kimi K3: Moonshot — Kimi K3 guide (accessed 2026-08-29)
- Kimi K3: Kimi K3 quickstart (max_completion_tokens up to 1,048,576) (accessed 2026-09-26)
Related comparisons
- Claude Fable 5.1 vs GPT-6 Astra
- Claude Fable 5.1 vs Kimi K3
- Claude Fable 5 vs GPT-6 Astra
- Claude Fable 5 vs Kimi K3
- Claude Mythos 5.1 vs GPT-6 Astra
- Claude Mythos 5.1 vs Kimi K3
Diving deeper on one model? GPT-6 Astra · Kimi K3
Common questions
GPT-6 Astra vs Kimi K3
Answered from the verified figures on this page rather than general guidance.
Is GPT-6 Astra or Kimi K3 cheaper for input?
Is GPT-6 Astra or Kimi K3 cheaper for output?
Which has the larger context window, GPT-6 Astra or Kimi K3?
Do GPT-6 Astra and Kimi K3 support the same reasoning levels?
Should I use GPT-6 Astra or Kimi K3?
Can I self-host GPT-6 Astra or Kimi K3?
Next step
Choosing between them
The verified figures that separate this pair, computed from the catalog rather than restated boilerplate.
- Input price: Kimi K3 costs $3 per 1M tokens versus $10 for GPT-6 Astra — a 3.3x difference at the headline tier.
- Context: GPT-6 Astra takes 1.05M against 1.05M for Kimi K3 — only decisive if your prompts approach the smaller window.
- Published Artificial Analysis Intelligence Index results differ: GPT-6 Astra scores 53 to 44 on the same listed benchmark, source attribution, source date, effort, and run variant (2026-10-01).
- Deployment: Kimi K3 publishes weights you can self-host; the other is API-only.
Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Windsurf, or Lovable? Compare agentic harnesses · Latest releases.