Free tool
Model routing planner for coding agents.
Pick every AI subscription you pay for and the job. Get a suggested model and effort for planning, building, reviewing and testing, with a reason for each choice.
FreeNo signupRuns in your browser
Let the canvas do the routing.
Starkomand routes handoffs using your settings, available models and, when enabled, live usage meters. Its choices can differ from this planner. Claude Code, Codex, Grok, Cursor and the rest run side by side on one Mac canvas.
v1.0.7 · Apple silicon · free for 3 days · no account
How the planner decides.
Six heuristics inspired by Starkomand’s Auto routing. No model is asked, and no benchmark or live quota is measured.
Strong models think, workhorses build.
Balanced and Best quality reserve the strongest tier for planning and review. Lower usage steps these down for ordinary work. A lower tier is a starting point: escalate if it struggles.
Spread the work across vendors.
With two plans of one size, one plans while the other builds and tests. A third vendor reviews when you have one, so the plan, the build and the review come from three different models.
Bigger plans lead, $20 plans step down.
A $100 or $200 plan plans and builds, and a smaller plan beside it only reviews. Claude Pro never runs Fable, ChatGPT Plus never runs GPT-6 Astra, and X Plus runs Grok at medium effort.
The job sets the build.
A small fix skips planning. Routine and big builds start at balanced and thorough tiers. Your preference moves the build one step; Lower usage also reduces planning, review and testing tiers for ordinary work.
Risk goes up, never down.
A change to auth, payments, data or security is built by your strongest subscription at higher effort and reviewed by another one.
Flat plans build small, credits buy second opinions.
GLM’s Coding Plan builds small fixes, and routine ones when you lean toward lower usage. Copilot spends credits per request, so it reviews with another vendor’s model instead of doing the heavy steps.
Published token rates and plan access.
Rates are relative to each vendor’s workhorse. They are not estimates of subscription usage or total task cost.
| Model | Price per token | Where you get it |
|---|---|---|
| Fable 5.1 | 5× Sonnet 5.5 | Claude Max, up to 50% of the weekly limit; usage credits only on Pro; Copilot |
| Opus 5.5 | 2× Sonnet 5.5 | Claude Pro and Max, where it is the default model; Copilot |
| Sonnet 5.5 | 1× Sonnet 5.5 (the workhorse) | Claude Pro and Max; Copilot |
| Haiku 5.5 | 1/20× Sonnet 5.5 | Claude Pro and Max; Copilot |
| GPT-6 Astra | 5× GPT-6.1 Sol | ChatGPT Pro, with no 5-hour limit; 5–45 messages per 5 hours on Plus; Copilot |
| GPT-6.1 Sol | 1× GPT-6.1 Sol (the workhorse) | ChatGPT Plus, 15–160 messages per 5 hours, and Pro; Copilot |
| GPT-6 Luna | 1/20× GPT-6.1 Sol | ChatGPT Plus, 350–3,000 messages per 5 hours, and Pro; Copilot |
| Gemini 4 Argon | 2.7× Gemini 3.8 Flash | Google AI Pro and Ultra, in Antigravity; Copilot |
| Gemini 3.8 Flash | 1× Gemini 3.8 Flash (the workhorse) | Google AI Pro and Ultra, in Antigravity; Copilot |
| Grok 4.7 | 1× Grok 4.7 (the workhorse) | The X Subscription’s pool, in Grok Build and Cursor, medium effort on Plus; Copilot |
| Composer 2.5 | 0.4× Grok 4.7 | The X Subscription’s pool, in Cursor |
| Kimi K3 | 3.8× Kimi K2.7 Code | Kimi Code plans; Copilot |
| Kimi K2.7 Code | 1× Kimi K2.7 Code (the workhorse) | Kimi Code plans; Copilot |
| GLM-5.3 | Flat monthly plan | GLM Coding Plan, Lite from $18 a month |
Read on 16 October 2026: Claude Code model configuration, Fable on your Claude plan, Anthropic API pricing, Codex pricing and limits, Google AI plans, Gemini 4 Argon in Antigravity, Gemini CLI quotas, Cursor plans and the X Subscription, Cursor model prices, Grok Build settings, GitHub Copilot plans, Kimi membership, GLM Coding Plan. Rates compare output prices per token; input prices sit in the same ratio except Composer 2.5 (1/4× Grok 4.7) and Kimi K3 (3.2× K2.7 Code). Gemini, Grok, Composer and Kimi prices are from Cursor’s list; message ranges are OpenAI’s estimates. Product names belong to their owners; Starkomand is not affiliated with them. Spotted a change? Tell us.
Questions.
Will Starkomand choose these exact models?
Not necessarily. This standalone planner uses a fixed model catalog and rules of thumb. The installed app has its own routing table and also considers available models, your settings and live usage when enabled. App versions may name older models. Check the receiving model before starting a handoff.
What is model routing for coding agents?
Running each step of a task on the model that suits it. A frontier model such as Fable 5.1, GPT-6 Astra or Gemini 4 Argon plans and reviews, where judgement matters most; a cheaper model such as GPT-6.1 Sol, Opus 5.5 or Grok 4.7 writes the code. These are rules of thumb, not measured quality or savings guarantees. Check the results and escalate when needed.
Which subscriptions does the planner cover?
Claude, ChatGPT for Codex, Google AI for Gemini in Antigravity, The X Subscription for Grok Build and Cursor, GitHub Copilot, Kimi Code and GLM’s Coding Plan. On a legacy Cursor plan, pick the X tier Cursor maps it to: Pro to Premium, Pro+ to Super, Ultra to Ultra.
Who plans when I have several subscriptions?
The biggest plan. Between plans of one size, Claude leads by default, then ChatGPT, Google AI, X and Kimi; the button under the answer hands planning to the next one. The plan that plans also reviews, unless a third vendor can, so the review always comes from a second model.
Why does Copilot only review?
Copilot spends your monthly AI credits on every request, while Claude, ChatGPT, Google AI, X and Kimi limits reset on their own. So the planner keeps Copilot for a second opinion from another vendor’s model, such as Gemini 4 Argon checking a Claude build, and gives it every step only when it is your one subscription.
Does Claude Pro include Fable 5.1?
No. Anthropic says Fable 5 and Fable 5.1 are outside Pro’s usage limits and run on usage credits. On Max they are included, and up to 50% of the weekly limit can go to them. So on Pro the planner thinks with Opus 5.5 instead.
How many GPT-6 Astra messages does ChatGPT Plus get?
OpenAI estimates 5–45 local Codex messages per five hours on Plus, against 15–160 for GPT-6.1 Sol and 350–3,000 for GPT-6 Luna, with weekly limits on top. Pro has no five-hour limit. That is why Plus thinks on GPT-6.1 Sol at high effort here.
Where is Gemini CLI?
Google replaced Gemini CLI with Antigravity CLI, agy, on 18 June 2026, and Gemini 4 Argon reached Antigravity for Google AI Pro and Ultra subscribers on 13 October. Antigravity, Kimi Code and Copilot CLI do not publish their model names, so for them the planner gives the command and the model to pick with /model.
Is this the same as Claude Code’s opusplan?
Same idea, wider reach. opusplan uses Opus in plan mode and Sonnet for the rest of one Claude Code session. The planner splits a task across steps and across subscriptions, so the review can come from another vendor’s model.
Does the effort level change token use?
Higher effort can increase reasoning tokens. Lower usage reduces the recommended tiers across ordinary workflows, while risky work keeps its checks. Actual usage depends on context, output, retries and handoffs, so a cheaper model does not guarantee fewer tokens.
What does the rate on each step mean?
The model’s output price per token against its family’s workhorse, such as Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash or Grok 4.7, from the vendors’ API prices, OpenAI’s Codex credit rates and Cursor’s price list. A 5× rate means five times the published price per token, not five times the subscription usage. These ratios do not compare vendors directly or predict your remaining quota.
Is what I pick sent anywhere?
No. The planner runs in your browser and calls no model. Your choices live only in the page address, so a result can be shared as a link.