Kimi K3 in GitHub Copilot: Plans, Cost, Setup & When to Pick It
September 25, 2026 · kimi k3 github copilot · kimi k3 copilot · copilot ai credits · copilot model picker · kimi k3 vs claude · copilot byok · kimi k3 pricing · github copilot models
Kimi K3 is generally available in GitHub Copilot on the Pro, Pro+, Max, Business and Enterprise plans, billed per token at $3 per 1M input tokens, $0.30 per 1M cached input tokens and $15 per 1M output tokens. Individual subscribers can pick it in the model picker today. On Business and Enterprise it is off by default until an admin enables it. There is no premium request multiplier: Copilot moved to usage-based AI credits before K3 arrived.
This guide reflects GitHub’s changelog, billing and model docs as of September 25, 2026, seven weeks after launch. It covers which plans get K3, where you can actually select it (the docs disagree in one place), how to enable it, what it costs next to Claude and GPT, and when it’s the right pick.
What GitHub actually shipped
- August 6, 2026: GitHub announces that Kimi K3 is now available in GitHub Copilot, citing “frontier-level abilities on agentic coding with highly cost-effective pricing.” The model is “hosted by GitHub on Fireworks AI.”
- Same day: the rollout was “temporarily paused” during a GitHub Actions incident (DEV Community write-up), then resumed, per a later editor’s note.
- October 2, 2026 (scheduled): Kimi K2.7 Code retires from Copilot. GitHub’s supported models page names K3 as the suggested alternative.
If you use K2.7 Code today, note that its replacement costs roughly three times more per token ($3/$15 vs $0.95/$4).
The model is the same K3 from our specifications breakdown: 2.8T total parameters, 104B active, 1,048,576-token context. GitHub confirms both the 1M context and configurable reasoning work inside Copilot.
Which Copilot plans get Kimi K3
| Plan | Price | Monthly AI credits | Kimi K3 access | Default state |
|---|---|---|---|---|
| Copilot Free | $0 | Small allowance | No (Auto selection only) | n/a |
| Copilot Student | n/a | n/a | No (Auto selection only) | n/a |
| Copilot Pro | $10/mo | 1,500 (1,000 base + 500 flex) | Yes | Available in picker |
| Copilot Pro+ | $39/mo | 7,000 (3,900 base + 3,100 flex) | Yes | Available in picker |
| Copilot Max | $100/mo | 20,000 (10,000 base + 10,000 flex) | Yes | Available in picker |
| Copilot Business | $19/seat/mo | 1,900 per user, pooled | Yes | Off until admin enables |
| Copilot Enterprise | $39/seat/mo | 3,900 per user, pooled | Yes | Off until admin enables |
| Legacy annual Pro/Pro+ | Annual | Premium requests | No new models | n/a |
Sources: GitHub Copilot plans, usage-based billing for individuals, billing for organizations. One AI credit = $0.01.
The Pro plan is where K3 stands out. GitHub sorts models into Lightweight, Versatile and Powerful categories, and on the $10 Pro plan only two “Powerful” models are selectable: Kimi K3 and GPT-5.3-Codex. Every Claude Opus and Fable model, GPT-5.5, GPT-5.6 Sol, GPT-6 Sol and GPT-6 Astra need Pro+ or higher. Also note that the “flex” part of each individual allowance is designed to “adapt as the economics of AI evolve,” so those totals can change.
There is no premium request multiplier anymore
Older guides still quote Copilot models as “1x” or “3x.” That system ended on June 1, 2026, when all Copilot plans moved to usage-based billing. Premium request units were replaced by GitHub AI Credits, calculated from actual input, cached and output tokens at each model’s list price.
GitHub’s legacy multipliers page says multipliers now apply only to Pro and Pro+ subscribers who stayed on an annual plan, and that those users “will not receive access to new models and features.” K3 is not in that table. There is no Kimi K3 premium request multiplier. It is only sold per token.
Three billing details matter for K3:
- Code completions stay unlimited. Inline suggestions aren’t billed in AI credits on paid plans. K3 only draws credits in chat, agent sessions and the CLI.
- No Auto discount. Auto model selection gets a 10% discount, but K3 isn’t in the Auto pool, so you always pay full list rate.
- No long-context surcharge. GPT-5.4 through GPT-6 models double their input price above a 200K–272K input-token threshold. K3 has no long-context tier and stays at $3/M across its full 1M window.
Where you can use it: clients and agent surfaces
| Surface | K3 selectable? | Notes |
|---|---|---|
| VS Code (Chat, incl. agent sessions) | Yes | Minimum VS Code v1.131 |
| Visual Studio | Yes | Minimum version listed as TBD |
| JetBrains IDEs | Yes | Thinking Effort submenu for reasoning models |
| Xcode, Eclipse | Yes | Minimum versions TBD |
| github.com Chat | Yes | Model dropdown at the bottom of Copilot Chat |
| Copilot CLI | Yes | /model command or --model flag |
| GitHub Copilot app, GitHub Mobile | Yes (per changelog) | |
| Copilot cloud agent | Unclear | Listed in the changelog, missing from the cloud agent model list |
| Auto model selection | No | K3 not in the Auto pool |
| Copilot code review | No | Model is chosen automatically and not disclosed |
The cloud agent conflict is real. The changelog lists “GitHub Copilot cloud agent” as a surface, but GitHub’s cloud agent model documentation, checked September 25, lists 19 selectable models plus Auto, and Kimi K3 isn’t one of them. Either the doc lags or the rollout is incomplete. If you plan to assign issues to Copilot with K3, check your own picker first. Cloud agent sessions also consume GitHub Actions minutes on top of AI credits.
For interactive agent work in the IDE, support is confirmed: K3 sits in the same chat model picker that drives agent sessions, with a reasoning-level setting.
How to enable Kimi K3: step by step
Individual plans (Pro, Pro+, Max)
Individual subscribers “can use AI models directly within Copilot without configuring access or managing policies,” per GitHub. So:
- Update your client. VS Code needs v1.131 or later for K3. For accurate billing display GitHub also recommends JetBrains plugin 1.9.1+, Visual Studio 17.14.33+ (or 18.6.0+), and Copilot CLI 1.0.48+.
- Open Copilot Chat and click the model dropdown at the bottom of the chat view.
- Select Kimi K3. If you were in an Auto session, start a new chat first.
- Set the thinking effort from the submenu next to the model name (VS Code, JetBrains). Higher effort means more thinking tokens, billed as output at $15/M.
- In the CLI, run
/modelor pass--model. - Run a small test task and check your usage page before pointing K3 at a large refactor.
Business and Enterprise (admins)
This is where teams get stuck. GitHub’s default model availability doc excludes open-weight models (DeepSeek, Kimi K2.7 Code, Kimi K3) from the “Default availability for released models” policy. They stay disabled regardless of that setting until someone enables them by hand.
- Enterprise owner: enable Kimi K3, or leave it optional so organizations can decide.
- Organization owner: open the org’s Settings, then under “Code, planning, and automation” click Copilot, then Models.
- Set Kimi K3 to Enabled. A lock icon means the enterprise owner has enforced the setting.
- Check model switching. On Business, the org must also allow members to switch models.
- Set budgets first. Additional usage beyond pooled credits is enabled by default, and there is “no automatic fallback to lower-cost models when a budget is exhausted.” Use user-level budgets or disable the “AI credits paid usage” policy for a hard cap.
GitHub also asks admins to “review open-weight models against their own security, compliance, and data-governance requirements” first. The data section below covers what that review should include.
What Kimi K3 costs in Copilot vs other models
Here are the per-token list prices from GitHub’s models and pricing page, limited to the models people actually compare K3 against:
| Model | GitHub category | Input /1M | Cached input /1M | Output /1M | Plans |
|---|---|---|---|---|---|
| Kimi K2.7 Code (retiring Oct 2) | Versatile | $0.95 | $0.19 | $4.00 | Pro and up |
| Claude Sonnet 5 | Versatile | $2.00 | $0.20 | $10.00 | Pro and up |
| GPT-6 Sol | Powerful | $2.00 | $0.20 | $10.00 | Pro+ and up |
| GPT-5.6 Terra | Versatile | $2.00 | $0.20 | $12.00 | Pro and up |
| GPT-5.3-Codex | Powerful | $1.75 | $0.175 | $14.00 | Pro and up |
| Kimi K3 | Powerful | $3.00 | $0.30 | $15.00 | Pro and up |
| Claude Opus 5.5 | Powerful | $4.00 | $0.20 | $20.00 | Pro+ and up |
| Claude Opus 5 | Powerful | $5.00 | $0.50 | $25.00 | Pro+ and up |
| Claude Fable 5.1 | Powerful | $10.00 | $0.25 | $50.00 | Pro+ and up |
| GPT-6 Astra | Powerful | $10.00 | $1.00 | $50.00 | Pro+ and up |
GPT prices shown at the default tier (up to 272K input tokens). Claude and the newer GPT models also charge a separate cache-write price; K3’s Copilot listing has no cache-write column.
K3’s Copilot price matches Moonshot’s own list price on the Kimi platform pricing page, so GitHub adds no markup. That makes K3 cheaper than every Opus and Fable model but more expensive than Claude Sonnet 5 and GPT-6 Sol. Calling K3 “the cheap option in Copilot” misses that second part.
A worked example: one agent session
Assume one cache-heavy agent session: 1.5M total input tokens across many requests, 90% cache hits (typical of agentic coding), and 40K output tokens. Every request stays under 200K tokens, so no long-context tiers apply. Cache-write charges are excluded.
| Model | Uncached input | Cached input | Output | Session cost | AI credits |
|---|---|---|---|---|---|
| Kimi K2.7 Code | $0.14 | $0.26 | $0.16 | $0.56 | ~56 |
| Claude Sonnet 5 | $0.30 | $0.27 | $0.40 | $0.97 | ~97 |
| GPT-6 Sol | $0.30 | $0.27 | $0.40 | $0.97 | ~97 |
| GPT-5.6 Terra | $0.30 | $0.27 | $0.48 | $1.05 | ~105 |
| Kimi K3 | $0.45 | $0.41 | $0.60 | $1.46 | ~146 |
| Claude Opus 5.5 | $0.60 | $0.27 | $0.80 | $1.67 | ~167 |
| Claude Opus 5 | $0.75 | $0.68 | $1.00 | $2.43 | ~243 |
| Claude Fable 5.1 | $1.50 | $0.34 | $2.00 | $3.84 | ~384 |
| GPT-6 Astra | $1.50 | $1.35 | $2.00 | $4.85 | ~485 |
Two caveats. K3 is a verbose thinker, so at high reasoning effort its output tokens (the priciest line) can run well above another model’s for the same task. And a model that finishes in fewer turns can beat a cheaper-per-token model on total cost. This table shows price per token, not cost per completed task.
How far your monthly credits go on K3
| Plan | Credits/month | K3 sessions like the example | Short K3 questions (~8 credits each)* |
|---|---|---|---|
| Pro | 1,500 | ~10 | ~180 |
| Pro+ | 7,000 | ~48 | ~840 |
| Max | 20,000 | ~137 | ~2,400 |
| Business | 1,900/user (pooled) | ~13 per user | ~230 per user |
| Enterprise | 3,900/user (pooled) | ~26 per user | ~470 per user |
*Estimated at 20K uncached input and 1.5K output tokens ≈ $0.08.
On Pro, about ten heavy K3 agent sessions use up the month. Save K3 for the tasks that justify it.
BYOK: running K3 with your own key
To pay Moonshot (or another K3 host) directly, Copilot’s bring-your-own-key support offers two routes.
Local BYOK, configured per developer:
- VS Code: model picker, Manage Language Models, Add Models, Custom Endpoint. The VS Code docs say it supports Chat Completions, Responses and the Anthropic Messages API. Moonshot’s API overview offers
https://api.moonshot.ai/v1(Chat Completions, Responses) andhttps://api.moonshot.ai/anthropic(Messages); the model ID iskimi-k3. BYOK models “work without signing into a GitHub account and without a Copilot plan,” though inline suggestions and embeddings still need Copilot. - Copilot CLI: set
COPILOT_PROVIDER_BASE_URL,COPILOT_PROVIDER_TYPE=openai,COPILOT_PROVIDER_API_KEYandCOPILOT_MODEL. - JetBrains, Xcode, the Copilot app and the Copilot SDK also support local BYOK.
Enterprise BYOK (public preview) lets enterprise owners add custom models for all Business/Enterprise users, including any OpenAI-compatible provider since January 2026.
BYOK makes sense when you already have K3 capacity elsewhere, such as a Moonshot account or a US or EU host from our K3 API providers guide. Watch the caching math: Moonshot’s own API now lists separate K3 cache-write prices ($3.00/M at 5-minute TTL, $6.00/M at 1-hour TTL), while Copilot’s listing shows none. On Business/Enterprise, the “Bring Your Own Language Model Key in Select IDEs” policy must be on. If you’d rather use Moonshot’s own harness, see our Kimi Code + K3 setup guide.
Data handling and safety caveats
GitHub’s model hosting page answers the first security question: K3 “is hosted by GitHub on Fireworks AI” under a zero data retention agreement, and “prompts and completions are not sent to Moonshot AI.”
The same docs carry caveats worth reading:
- Alignment warning. GitHub calls the Moonshot models “open-weight models that may be less aligned than other Copilot models, with an elevated risk of geographic bias.” Its model comparison page adds that in pre-release testing K3 “exhibited elevated risk on certain higher-risk prompts,” and says extra safeguards are deployed.
- Different variants by plan. “Fine-tuned model variants may be included as part of the Kimi K3 (GitHub) offering on individual plans only.” A K3 you tested on a personal Pro account may behave slightly differently from your company’s.
- Training on individual plans. GitHub doesn’t train on Business or Enterprise data, but for Free, Pro, Pro+ and Max it “may use Copilot interaction data” unless you opt out. This applies to every model.
When to pick K3 over Claude or GPT in Copilot
GitHub’s model comparison sums K3 up as best for “multi-step agent tasks across large codebases.” That matches our earlier comparisons: K3 is #1 on the Frontend Code Arena (1,679 Elo) and leads Moonshot’s long-horizon SWE Marathon (42.0), while trailing Claude Fable 5 and GPT-5.6 Sol on some broader suites. Details in K3 vs Claude Opus 5 and K3 vs Claude Fable 5.
| Your situation in Copilot | Pick | Why |
|---|---|---|
| Pro plan, want a frontier-class agent model | Kimi K3 | Opus, Fable, GPT-6 Sol/Astra aren’t available on Pro |
| Frontend/UI work in agent sessions | Kimi K3 | #1 blind-voted Frontend Code Arena ranking |
| Single requests over ~272K tokens (huge repo context) | Kimi K3 | Flat $3/M input; GPT models double past their threshold |
| Everyday chat, quick fixes, explanations | Claude Sonnet 5 or GPT-6 Sol | ~33% cheaper per token than K3 at similar tiers |
| Cost-sensitive team default | Auto | 10% discount; K3 isn’t in the pool |
| Hardest debugging, pay-anything quality | Claude Opus 5.5 / Fable 5.1 or GPT-6 Astra | Higher ceilings on some suites, at 1.1x–3.3x K3’s per-token price |
| Assigning issues to the cloud agent | Check your picker | K3 missing from the cloud agent doc as of Sept 25 |
| Currently using K2.7 Code | Move to K3 or Sonnet 5 before Oct 2 | K2.7 Code retires; K3 is GitHub’s suggested replacement |
On Pro+ and Max, a sensible pattern is Sonnet 5 or GPT-6 Sol for short interactions and K3 for multi-file agent sessions, frontend builds and very large contexts.
FAQ
Which GitHub Copilot plans include Kimi K3? Kimi K3 is available on Copilot Pro ($10/month), Pro+ ($39), Max ($100), Business ($19 per seat) and Enterprise ($39 per seat). Copilot Free and Copilot Student only get Auto model selection, and K3 is not in the Auto pool, so they cannot use it. Subscribers still on a legacy annual Pro or Pro+ plan with request-based billing do not receive new models either.
What is the premium request multiplier for Kimi K3 in Copilot? There isn’t one. GitHub moved every Copilot plan to usage-based billing on June 1, 2026, and Kimi K3 launched on August 6, after the switch. K3 is billed per token at $3 per 1M input, $0.30 per 1M cached input and $15 per 1M output, converted into GitHub AI Credits at 1 credit = $0.01. Premium request multipliers now only apply to legacy annual plans, and K3 is not on that list.
How do I enable Kimi K3 in GitHub Copilot for my organization? Kimi K3 is off by default on Copilot Business and Enterprise, and it is excluded from the Default availability policy that auto-enables new models. An organization owner goes to Organization Settings, then Copilot, then Models, and sets Kimi K3 to Enabled. In an enterprise, the enterprise owner must either enable it or delegate the choice to organizations first.
Can I use Kimi K3 in Copilot agent mode and the Copilot cloud agent? In the IDEs, yes: K3 is selectable in the chat model picker in VS Code (v1.131 or later), Visual Studio, JetBrains, Xcode and Eclipse, and in Copilot CLI via /model. GitHub’s launch changelog also lists the Copilot cloud agent, but as of September 25, 2026 the cloud agent’s own model documentation does not include Kimi K3 in its picker list, so check your picker before planning around it.
Is Kimi K3 cheaper than Claude in GitHub Copilot? Cheaper than Claude Opus 5 ($5/$25), Opus 5.5 ($4/$20) and Fable 5 ($10/$50), but more expensive than Claude Sonnet 5 ($2/$10). In our worked example of a cache-heavy agent session, K3 cost about 146 AI credits versus 97 for Sonnet 5, 167 for Opus 5.5 and 243 for Opus 5. K3 is a mid-priced frontier option in Copilot, not a budget model.
Can I bring my own Moonshot API key to GitHub Copilot? Yes, through BYOK. In VS Code you can add a Custom Endpoint pointing at Moonshot’s OpenAI-compatible API (https://api.moonshot.ai/v1, model kimi-k3), and Copilot CLI accepts an OpenAI-compatible provider via environment variables. Business and Enterprise admins can disable local BYOK by policy, and enterprise-wide custom models are still in public preview.
Where is Kimi K3 hosted in GitHub Copilot, and does Moonshot see my code? GitHub hosts Kimi K3 on Fireworks AI under a zero data retention agreement. According to GitHub’s model hosting docs, prompts and completions are not sent to Moonshot AI and Fireworks does not train on them. On individual plans, GitHub itself may use interaction data for training unless you opt out; Business and Enterprise data is not used.
The bottom line
Kimi K3 in Copilot is a genuine frontier option at a mid-range price, not a bargain model. At $3/$15 with no GitHub markup, it undercuts every Claude Opus and Fable model but costs more than Sonnet 5 or GPT-6 Sol. It wins clearly on the $10 Pro plan, on frontend and long-horizon agent work, and on very large contexts. The friction is real too: admins must enable it by hand, it isn’t in Auto, the cloud agent docs don’t list it yet, and GitHub’s safety notes deserve a read. Update your IDE, run one real multi-file task, check the credit meter, then decide.
Keep reading: every K3 API provider compared, K3 vs Claude Opus 5, K3 vs Claude Fable 5, K3 vs GPT-6 Astra, the Kimi Code + K3 setup guide and the full K3 specifications.