Kimi K3.1 Pro vs Opus 5.5, Gemini 4 Argon and GPT-6 Astra
Compare Kimi K3.1 Pro with Opus 5.5, Gemini 4 Argon and GPT-6 Astra on historical benchmarks and current API prices, with a November 11 Carbon availability update.
Published · Updated
Kimi K3.1 Pro is released. Moonshot shipped it on October 23, 2026 on kimi.ai, Kimi Work, Kimi Code and the Kimi API under the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed; its launch statement promised full weights by November 4, 2026. This update checks Carbon availability and route prices, not that later weight release. K3.1 Pro is a 3T-class MoE model with a 1,048,576-token context and native visual input, positioned by Moonshot at 2.8T total / 104B active parameters. On Moonshot's own historical Terminal-Bench 4.0 run it scores 65.7%, just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2% under that run's settings.
Official price is the headline: kimi-k3.1-pro costs $2 input / $10 output per 1M tokens, half of Opus 5.5's $4 / $20 and a third of GPT-6 Astra's $10 / $50. The Highspeed variant is $4 / $20, i.e. twice the price for roughly 2-3x the speed.
APIMaster public route prices checked 2026-11-11 are below. Only active status=1 routes without client-exclusive labels are counted. USD per million input / output tokens; the live marketplace card is authoritative. Historical launch benchmarks elsewhere on this page retain their original dates.
| Model ID | Official / reference price | Current APIMaster price | Active routes | Discount | Model card |
|---|---|---|---|---|---|
kimi-k3.1-pro |
$2 / $10 | From $0.80 / $4 | 5 | Up to ~60% off | Compare routes |
claude-opus-5-5 |
$4 / $20 | From $1.340784 / $6.70392 | 4 | Up to ~66.5% off | Compare routes |
gemini-4-carbon |
$2 / $10 through 2026-11-30 | From $0.3588624 / $1.794312 | 4 | Up to ~82.1% off | Compare Carbon routes |
gemini-4-argon |
$2 / $10 through 2026-11-30 | From $0.8341424 / $4.170712 | 4 | Up to ~58.3% off | Compare routes |
gpt-6-astra |
$10 / $50 reference | From $0.448485 / $2.242425 | 27 | Up to ~95.5% off | Compare routes |
kimi-k3 |
$3 / $15 | From $1.20 / $6 | 9 | Up to ~60% off | Compare routes |
GPT Trial(GPT-6 Astra, GPT-6 Sol, GPT-5.6, GPT-5.5, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-6.1-sol, gpt-6-astra, gpt-6-sol, and gpt-6-luna available)
Sign up and claim a $20 GPT trial
Kimi K3.1 Pro specifications
| Item | Kimi K3.1 Pro | Notes |
|---|---|---|
| Release | October 23, 2026 | kimi.ai, Kimi Work, Kimi Code, Kimi API |
| API model IDs | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
Highspeed is 2-3x faster |
| Architecture | 3T-class MoE, 2.8T total / 104B active (per Moonshot's launch post) | Post-trained from Kimi K3 with larger post-training compute |
| Context | 1,048,576 tokens | Same window as K3 |
| Input / output | Text, images and video in; text out | Native visual understanding |
| Weights | Launch promise: full weights by November 4, 2026 | Later release status not rechecked in this Carbon update |
| Pricing | $2 / $10 per 1M tokens; Highspeed $4 / $20 | Cache hit $0.20, 5m write $2.00, 1h write $4.00 |
The "3T-class" wording is Moonshot's. The 2.8T/104B split comes from the launch post, not from the API documentation, so treat it as vendor-reported. Kimi's pricing table lists a $0.20 cache-hit rate and separate cache-write charges, which matters for long agent runs.
Benchmarks: where K3.1 Pro actually lands
These are vendor and leaderboard numbers, not our measurements. Harnesses differ, so compare rows only where the same benchmark and effort setting are named.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Argon |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% | 57.4% |
| BenchCAD, with tools | 92.7% | 96.2% | 95.9% | — |
| BenchCAD, no tools | 72.8% | 73.0% | — | — |
| Hallucination, claim-level, no tools (lower is better) | 5.4% | 6.8% | 4.2% | — |
| Hallucination, claim-level, with tools (lower is better) | 8.3% | 15.0% | 5.5% | — |
| SWE-Marathon v1.1, public leaderboard | Not listed | 46.3% | 42.5% | 43.8% |
Reading the table honestly:
- Terminal-Bench 4.0 is the cleanest cross-model row. Moonshot evaluated K3.1 Pro with the Claude Code harness and took the other scores from each vendor's blog or the official leaderboard. The gap to Opus is 0.7 points, which is inside the noise of most agent benchmarks.
- 3D/CAD is close but not equal. K3.1 Pro lands 3.5 points behind Opus with tools and 0.2 points behind without tools. For parametric parts and scripted Blender/CAD work, that is competitive; for the hardest reconstructions, Opus still leads.
- Hallucination is where the gap shows. On Moonshot's own stress test, Astra is best with and without tools. K3.1 Pro beats Opus, but it is not the most factually reliable model in this group.
- SWE-Marathon is the caution. Kimi publishes an internal long-horizon run, but K3.1 Pro is not on the public v1.1 leaderboard, where K3, Opus, Argon and Astra all have entries. Do not mix Kimi's harness number with those columns.
- Gemini 4 Argon is not a coding leader. Google's own model card puts Argon at 57.4% on Terminal-Bench 4.0 and 55.0% on FrontierSWE v2, behind Astra's 65.5%. Argon's strength is knowledge work: 68.9% on the Vals Index, 65.4% on Vals Finance Agent v2 and 19.6% on Harvey's Legal Agent — all first in that table. If your workload is finance or legal analysis rather than terminal agents, Argon is the relevant comparison, not Kimi.
Gemini 4 Carbon is now released
Carbon reached general API availability on 2026-11-10 as gemini-4-carbon. The earlier October 22 observation that no Carbon endpoint existed is superseded. Its current specification includes a 1,048,576-token unified context and 65,536-token default synchronous output; Google Standard is $2 / $10 through November 30, then $4 / $20 from December 1.
At the 2026-11-11 marketplace check, APIMaster has 4 public active Carbon routes from $0.3588624 / $1.794312 per million input / output tokens, up to ~82.1% off current Google Standard. The live Carbon card is authoritative; native features vary by route.
The older Kimi benchmark table above is historical and does not provide a controlled Kimi-vs-Carbon measurement. Carbon's separate launch evaluation cannot be inserted as if tools and settings matched. See the Carbon overview, cheap Carbon API guide and Carbon vs Argon.
Early user tests: cheaper per task, not always cheaper per token
Independent results are one day old, so treat them as signals rather than verdicts.
- Pac-Man coding bench: a Hacker News commenter reports K3.1 Pro tying Claude Sonnet 5.5 at 93 points, while costing $0.68 and 11 minutes versus Sonnet's $1.61 and 14 minutes. Same score, roughly 58% cheaper and faster.
- Speed: the same thread reports K3.1 Pro being "almost twice as fast" as K3 on a private benchmark, consistent with Moonshot's claim that the post-training rework cut wasted thinking.
- Artificial Analysis: the index places K3.1 around 55, roughly six months behind the frontier on AA-Omniscience, and on the Pareto line for intelligence versus cost — strong for an open-weight model, not a new ceiling.
- The token-cost catch: another commenter notes that at
xhigh, Opus 5.5 emits about 100M tokens and stays only 1 point ahead, while athighit emits 53M and comes out cheaper per task. A cheap per-token price does not automatically mean a cheap finished task.
The honest summary from the thread: this is the first open-weight model in a long time that trades blows with the frontier — and the first in a while where the "open" part is not a handicap.
Which model should you actually use?
Long-horizon coding agents: test K3.1 Pro and Opus 5.5 on your own repository. They are within a point in the historical Terminal-Bench 4.0 table and the official price difference is 2x. APIMaster's November 11 check lists 5 public Kimi routes from $0.80 / $4, below Opus's $1.340784 / $6.70392 minimum.
3D, CAD and game-engine work: K3.1 Pro is genuinely close to Opus on BenchCAD and much cheaper officially. If your pipeline is scripted Blender/CAD generation, it is worth a trial this week.
Financial document analysis: OfficeQA Pro v2 is Moonshot's strongest claim (64.4, vendor-reported). Validate it on your own filings before switching anything material.
Factual reliability: GPT-6 Astra has the lowest hallucination rate in this group on Kimi's own stress test. If a wrong number is expensive, that row matters more than the coding gap.
Carbon and Argon are both available now. Evaluate Carbon on the same tasks and acceptance criteria; this article does not claim a measured Kimi-vs-Carbon winner.
To test any of these, register, create an API key, and point your client at https://apimaster.ai/v1 with kimi-k3.1-pro, claude-opus-5-5, gemini-4-argon or gpt-6-astra. Start with a representative task, check the model card's context and tool limits, and compare cost per accepted result rather than price per token.
FAQ
Is Kimi K3.1 Pro released?
Yes. Moonshot launched it on October 23, 2026 across kimi.ai, Kimi Work, Kimi Code and the Kimi API. API model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
How much does Kimi K3.1 Pro cost?
$2 per 1M input tokens and $10 per 1M output tokens on the Kimi API. Highspeed is $4 / $20. Cache hits are $0.20 per 1M, with separate cache-write charges.
Is Kimi K3.1 Pro better than Claude Opus 5.5?
On Terminal-Bench 4.0 they are within 0.7 points (65.7% vs 66.4%), and K3.1 Pro is half the official price. On BenchCAD and hallucination, Opus and Astra still lead in the published numbers. "Better" depends on whether you are paying per token or per accepted task.
Can I use Kimi K3.1 Pro on APIMaster?
Yes. At the November 11, 2026 check, kimi-k3.1-pro has 5 public active routes from $0.80 / $4 per million tokens, up to about 60% below the $2 / $10 official rate.
Is Gemini 4 Carbon available?
Yes. Google marks Carbon generally available on November 10, 2026. APIMaster lists 4 public active routes from $0.3588624 / $1.794312 at the November 11 check.
Is Kimi K3.1 Pro open source?
Moonshot's launch statement promised full weights by November 4, 2026. This Carbon availability update has not rechecked the later artifact release; confirm the current repository before relying on an open-weight deployment.
Sources
Checked October 24, 2026:
- Kimi K3.1 Pro technical blog: specs, benchmarks, availability and weight timeline.
- Kimi API pricing: $2 / $10, Highspeed $4 / $20, cache rates.
- Kimi model list: model IDs and context window.
- Google Gemini API model list (historical October 22 check); superseded for Carbon by the November 10 GA changelog.
- Google DeepMind Gemini 4 Argon model card: Terminal-Bench 4.0 score.
- SWE-Marathon v1.1 leaderboard: public long-horizon results.
- Hacker News discussion: Pac-Man bench, speed and token-cost observations.
- APIMaster marketplace data: active route snapshot.