Moonshot launches Kimi K3.1 Pro with 1M-token context

• Kimi K3.1 Pro ranks fourth on Artificial Analysis’s Intelligence Index: 44 → 55 versus K3, +11 points, according to Artificial Analysis’s launch results.
• Standard API rates fall 33% to $2/M input, $10/M output and $0.20/M cached input tokens, as Artificial Analysis’s pricing comparison records.
• Hosted access is live across Kimi’s products and API. Downloadable weights are promised by November 4, 2026, per Moonshot’s announcement.
• Cyber training filters cut an internal exploitation score from 90.5% to 81%, according to the published ablation. K3.1 Pro nevertheless leads publicly accessible models on Artificial Analysis’s defensive Cyber Index.

A mistyped Kimi Code model ID silently routes requests to K2.8 Preview without thinking. Moonshot’s technical report describes an agent correcting 36 shoulder-muscle calculation errors in a walking simulator. Plotly’s launch-day analytics exam found all five hidden data patterns, at nearly three times K3’s exam cost.

• Surfaces: Kimi web, mobile, Kimi Work, Kimi Code and the Kimi API are live, per Moonshot’s release thread.
• Standard API: kimi-k3.1-pro costs $2/M input, $10/M output and $0.20/M cache reads in Moonshot’s pricing table.
• Highspeed: kimi-k3.1-pro-highspeed promises the same capability at 2–3× output speed, with $4/M input and $20/M output, per OpenRouter’s announcement; Moonshot lists unchanged $0.20/M cache reads.
• Context and modalities: Up to 1,048,576 tokens, unchanged from K3, with image and video input documented in the Kimi Code model table.
• Coding plans: Standard access starts at Plus/Moderato; 1M context and Highspeed require Pro/Allegretto or above, according to the release-note screenshots.
• API access: The API quickstart requires a successful top-up of at least $1; cumulative top-ups determine rate-limit tiers.
• Enterprise: Organization access is off by default at launch; administrators can enable it, according to the availability announcement.
• Weights: Release is promised by November 4; Plotly reports that the license has not yet been announced.

First-party

• Kimi Code Bench 3.0 Long-horizon: K3 55.95% → K3.1 Pro 61.2%, +5.25 points, per OpenRouter’s summary of Moonshot’s results.
• SWE Marathon: Opus 5.5 46.3% → K3.1 Pro 51.7%, +5.4 points head-to-head, in Moonshot’s comparison.
• OfficeQA Pro v2: Opus 5.5 62.2 → K3.1 Pro 64.4, +2.2 points head-to-head, in Moonshot’s finance results.
• Internal Financial Modeling: K3 50.4 → K3.1 Pro 57.6, +7.2 points, in the internal finance chart.
• Internal Financial Research: K3 58.8 → K3.1 Pro 70.4, +11.6 points, in the same finance evaluation.
• Internal Business Operation: K3 8.9 → K3.1 Pro 37.0, +28.1 points, in Moonshot’s simulated-business results.
• BenchCAD: K3.1 Pro without tools 72.8% → with tools 92.7% voxel IoU, +19.9 points from tool access, in the BenchCAD chart.

Third-party evaluators

• Terminal-Bench 4.0: K3 13% → K3.1 Pro 67%, +54 points, in Artificial Analysis’s independent runs.
• GDPval-AA: K3 1,534 → K3.1 Pro 1,846 Elo, +312 Elo, per Artificial Analysis’s results.
• AutomationBench-AA: K3 58% → K3.1 Pro 73%, +15 points, per Artificial Analysis’s SaaS-workflow evaluation.
• AA-Omniscience: hallucinations 53% → 44%, −9 points; accuracy 48% → 51%, +3 points, per Artificial Analysis’s factuality tests.
• WebDev Arena: K3 1,655 → K3.1 Pro 1,692, +37 Arena points, moving from model rank 17 to 8 in Arena’s launch results.
• Cyber Index: K3 41% → K3.1 Pro 64%, +23 points, per Artificial Analysis’s defensive evaluation.

Customer-reported

• PlotlyDataAnalyticsBench: K3 35/43 → K3.1 Pro 42/43, +16.3 percentage points, in Plotly’s product-team evaluation.

• Word grouping: K3 93.6% → K3.1 Pro high 93.4%, −0.2 points, on the 940-puzzle Extended NYT Connections benchmark; it used fewer reasoning tokens, according to LechMazur’s results.
• Offensive cyber ablation: The unfiltered checkpoint’s 90.5% → the release model’s 81.0% pass@3, −9.5 points, after training-data removal, per Moonshot’s published comparison.
• Defensive task cost: K3 $2.45 → K3.1 Pro $2.81, +15%, despite lower token rates, because input-token consumption roughly doubled in Artificial Analysis’s Cyber Index runs.
• Analytics cost and latency: K3’s $5.01 → $13.47 per exam, +$8.46; median question time 43 → 150 seconds, +107 seconds, in Plotly’s test. Its comparison used OpenRouter pricing, including K3 input at $0.24/M, rather than Moonshot’s $3/M first-party baseline.

Frontier gaps remain visible in Moonshot’s own results:

• Terminal-Bench Science: K3.1 Pro 38.1% versus Astra 68.1%, a 30-point gap, in the launch-results breakdown.
• BenchCAD with tools: K3.1 Pro 92.7% versus Opus 5.5 96.2%, a 3.5-point gap, in the CAD reconstruction results.

A RuneScape comparison shared by teortaxesTex put Kimi at 5.7 XP/min and $4.70 per run, against DeepSeek V4.1 Flash 1017 at 5.8 XP/min and $0.09. The table explicitly labels the skill results “Best of 1” and asks readers to allow a wide error margin.

The same MoE, more post-training

K3.1 Pro retains K3’s 2.8T total and 104B active parameters, as Artificial Analysis’s parameter comparison confirms. Moonshot attributes the gains to substantially more post-training compute and joint reinforcement learning across hundreds of domains in its technical report.

The lab names six infrastructure changes:

1. RL framework.
2. Long-context inference.
3. Training efficiency.
4. Sandbox infrastructure.
5. Rollout scheduling.
6. Shared environment-quality checks.

Thinking-only requests and silent routing

Kimi Code changes several integration behaviors, according to its model configuration documentation:

• Reasoning supports low, high and max. The default changes from K3’s high to K3.1 Pro’s max.
• Disabling thinking returns HTTP 400.
• Coding-plan IDs are k3.1-pro and k3.1-pro-highspeed, distinct from the inference API’s kimi- prefixed IDs.
• Mistyped coding-plan IDs silently fall back to K2.8 Preview without thinking.
• Switching models or effort invalidates the context cache, triggering another prefill. Moonshot also warns that mixing reasoning histories can degrade output quality.
• Highspeed accelerates model output; tool execution and scripts run at their existing speed.

The inference API guide additionally requires complete assistant messages to be returned in subsequent tool-call turns, including more than just content. Streaming separates reasoning_content from final-answer content; strict JSON Schema constrains the final content field.

Claude Code and mixed comparison baselines

Moonshot evaluated K3.1 Pro with Claude Code on its principal coding benchmarks, according to the evaluation footnotes. Several comparisons combine those runs with competitors’ best published scores.

• Moonshot’s Terminal-Bench score is 65.7%, 0.7 points behind its listed Opus 5.5 result. Artificial Analysis’s separate run produces the 67% figure reported above.
• PostTrainBench uses H20 rather than H100 GPUs and a DeepSeek judge substituted for the official GPT judge.
• ProgramBench uses a binary threshold: a task scores one only when its underlying score reaches at least 95%.
• DeepSeek’s Terminal-Bench result comes from its release blog in the main table, but from Artificial Analysis in the cost chart, as teortaxesTex’s comparison documents.

Executable 3D workflows

K3.1 Pro writes and executes Blender and CAD scripts without requiring MCP connectors, according to Moonshot’s 3D description. Its BenchCAD evaluation uses a 1,000-case subset of the 17,900-case Vision2Code dataset, with Claude Code supplying tool access.

The accompanying 3D Make interface supports:

• Browser-based preview and editing.
• Downloads for 3D printing.
• Video and Live Photo exports.

Hallucination figures differ across launch surfaces

Moonshot’s launch prose reports claim-level errors falling from 23.4% to 8.3%, −15.1 points, on Kimi Hallu Bench Hard. It separately says responses containing any factual error fell to 42.4% of K3’s rate, and cautions that the deliberately difficult prompts do not represent everyday error rates.

Other launch surfaces carry different figures: the chart shared by ZhihuFrontier lists 5.9% for Hallu Bench Hard with tools, while the API quickstart reports 6.85% against a 23.35% K3 baseline. The official text does not reconcile those numbers.

Cyber filtering targets training data

Moonshot removed cyber-related material from both mid-training and RL while retaining general coding tasks, its cybersecurity section states. Its internal exploitation evaluation supplies source code, a bug hint and a binary, then scores pass@3 without hard timeouts or output-token limits.

Artificial Analysis measures defensive API-accessible workflows instead. It reports no safety refusals from K3.1 Pro; proprietary competitors can lose points to refusals in the Cyber Index results. Commenters on Hacker News debated whether removing cyber examples could also remove useful general security knowledge.

Token savings depend on the workload

K3.1 Pro generates 55K output tokens per Intelligence Index task versus K3’s 48K, +14%, according to Artificial Analysis’s token accounting. Lower first-party rates leave measured task cost almost unchanged: $2.00 → $1.98, −$0.02.

• Multi-agent 3D builds: crystalsssup’s robotic-hand website reportedly separates into 2,317 pieces. The run took about two hours with multiple agents, according to the timing follow-up.

• Blender to web: chetaslua reported 90 minutes and $12.50 for a folding-phone model and playable webpage, including its folding animation.

• Effort changes detail: cedric_chee found that high missed a shelter in a pagoda garden that max included, according to the effort comparison.
• Compiler work: QuixiAI reported that K3.1 Pro did not match Fable 5.1 or Opus 5.5 for productivity while developing a compiler in a hands-on follow-up.
• Visual comparisons: Reviewer Peter Gostev tested 3D scenes, HTML animations, websites, UI builds and research-driven pages in Arena’s first-impressions video.
• 3D ranking remains pending: Artificial Analysis says early AA-3D preference votes point to a lower placement than its agentic-work and coding results, with voting still underway in its mountain-bike scene demonstration.

• Vercel AI Gateway: Both variants are available as moonshotai/kimi-k3.1-pro and moonshotai/kimi-k3.1-pro-highspeed, per Vercel’s announcement.
• OpenRouter: Standard and Highspeed hosting are live, with Highspeed selectable through the provider list in OpenRouter’s release thread.
• OpenCode Go: K3.1 Pro is available in the coding subscription, according to OpenCode’s announcement.
• Quantization: Unsloth offered to investigate GGUF quantizations despite the model’s size in its response to the launch.

Moonshot also introduced Kimi Hosted Agents, a managed harness with a continuously available execution environment. Its product overview documents capabilities beyond model inference:

• Complete session history persisted as events.
• Platform-managed execution retries, scheduling and failure recovery.
• Built-in tools executed in cloud sandboxes; custom tools can remain client-side.
• Customer-hosted sandbox execution for enterprise-verified accounts, while Kimi retains model calls, orchestration and session management.
• An enterprise-only Inference Hook that screens task inputs and tool results before inference.

Hosted Agents bills separately for model tokens, sandbox runtime and plugin invocations. The initial catalog lists 140-plus skills, 30-plus plugins and 12 agent templates spanning finance, games, SRE incident response and community operations.