The guide · comparison

The three-model stack.

Three frontier models dropped in one week: Claude Opus 4.7, Kimi K2.6, and GPT-5.5. Most people pick one and move on. That's the wrong move. The actual win is routing each task to the right model — five seconds of routing per task cuts cost without losing quality on most workflows. This doc strips out the hype and gives you the practical setup, plus three system prompts to install.

The full comparison + system prompts

🔒 Drop your email to reveal the full setup

Get the full setup, free.

Drop your email and the setup unlocks instantly, plus every other guide and resource in the library. One email a week, unsubscribe anytime.

Setup

Three frontier models dropped in one week:

April 16: Claude Opus 4.7 (Anthropic)
April 20: Kimi K2.6 (Moonshot AI, open weights)
April 23: GPT-5.5 (OpenAI)

Most people pick one and move on. That's the wrong move. The actual win is routing each task to the right model.

My honest take upfront: the "300 parallel agents, 4,000 coordinated steps, replace a team of four" framing is mostly marketing. The real value is more boring and more useful — each model is genuinely best at different things, and switching between them cuts cost without losing quality on most workflows. That's worth setting up. The agentic swarm stuff is real but fragile; treat it as bonus, not core.

The only rule that matters

Task	Model	Why
Bulk coding, scaffolding, overnight runs, batch generation	Kimi K2.6	Cheapest by a lot, open weights, strong on long-horizon coding.
Production code, vision, legal/enterprise docs, anything precision-sensitive	Claude Opus 4.7	Leads on SWE-Bench Pro and most real-world code quality tests.
Math, deep web research, computer use / GUI navigation	GPT-5.5	Leads on FrontierMath, Terminal-Bench 2.0, OSWorld-Verified.

Five seconds of routing per task. That's the whole strategy.

What each model actually is

Kimi K2.6 (the cheap workhorse)

Released: April 20, 2026, open weights under modified MIT.
Pricing: ~$0.60 input / $2.50 output per 1M tokens (roughly 8x cheaper than Opus).
Architecture: 1T total parameters, 32B active per token, 256k context.
Strengths: long-horizon coding, agent swarm coordination, front-end generation from sketches, native video input.
Weaknesses: lags on single-turn high-stakes reasoning, weaker on math (GPQA, AIME) than GPT-5.5, vendor jurisdiction matters for some compliance contexts.

Claude Opus 4.7 (the senior engineer)

Released: April 16, 2026.
Pricing: $5 input / $25 output per 1M tokens.
Strengths: leads SWE-Bench Pro at 64.3% (~6 pts ahead of Kimi and GPT-5.5), strong vision, best-in-class on legal and enterprise documents, fewer errors than its predecessor on source-grounded work.
Weaknesses: expensive, slightly weaker on web research vs GPT-5.5.

GPT-5.5 (the researcher and operator)

Released: April 23, 2026.
Pricing: $5 input / $30 output per 1M tokens (Pro: $30/$180).
Context: 1M tokens in API.
Strengths: leads on Terminal-Bench 2.0 (82.7% vs Opus 69.4%), OSWorld-Verified for computer use (78.7%), FrontierMath, long-context retrieval.
Weaknesses: per Artificial Analysis, hallucination rate is significantly higher than Opus 4.7 on AA-Omniscience. Loses to Opus on real-world code quality.

How to actually set this up

Three options, ranked by effort.

Option 1: Manual routing (free, works today)

Three questions before every task:

Bulk coding or autonomous work? → Kimi
Production-perfect, vision, or legal? → Opus 4.7
Math, web research, or computer use? → GPT-5.5

That's it. Five seconds per task. Most people stop here and that's fine.

Option 2: Claude Code Router

Run Claude Code's interface but route requests to whichever model fits.

Repo: github.com/musistudio/claude-code-router
Lets you keep the Claude Code agent loop and swap the brain underneath.
Routes through OpenRouter so you can hit Kimi, GPT-5.5, or any model with one config.

Option 3: Auto-routing service

coderouter.io automatically picks the model per API call. No configuration. Useful if you don't want to think about routing at all and you trust someone else's heuristics.

Repos worth bookmarking

Cutting the original list down to what actually matters.

For Kimi K2.6:

github.com/moonshotai/Kimi-K2 official repo, weights, deployment guides
github.com/chongdashu/cc-kimi-k2-thinking-prompts use Kimi through Claude Code's CLI by swapping one env var

For Opus 4.7:

github.com/Piebald-AI/claude-code-system-prompts full Claude Code system prompt + 24 built-in tool descriptions

For routing across all three:

github.com/musistudio/claude-code-router single interface, three brains
github.com/asgeirtj/system_prompts_leaks leaked system prompts for all three models, useful for understanding how each company shapes behaviour

The other repos in the original article are mostly noise. These five cover 90% of what you'll actually use.

Three system prompts to install

One per model. Save somewhere accessible. Paste at session start or set as a persistent system prompt.

For Kimi K2.6 (bulk work and agents)

You are a senior engineer focused on implementation speed and correctness.
Your job: build exactly what is asked, nothing more, nothing less.
Rules:
- Read the full context before writing a single line
- Make surgical changes only, touch nothing adjacent to the task
- If you see a better approach, say so before building
- Validate your changes against existing logic before responding
- Every changed line must trace directly to the task
- For long-horizon tasks, state your plan and verify each step before moving to the next
When running as an agent:
- Report progress every 30 steps
- Flag blockers immediately instead of working around them silently
- If a subtask fails, pause and surface it rather than continuing
Success: the change works, nothing else broke, every step is traceable.

You are a senior engineer and architect working on production systems where correctness matters more than speed.
Your job: produce work that is right on the first pass, not fast on the first draft.
Rules:
- Identify what is actually being asked, not just what was literally said
- If multiple interpretations exist, name them and ask before proceeding
- Apply the simplest solution that fully solves the problem
- Flag assumptions explicitly before building on them
- If the approach is wrong, say so before building it
- Touch only what the task requires, no drive-by improvements
- Verify your output against the existing logic before responding
For documents and legal content:
- Flag any claim that requires a specific source to be trustworthy
- Distinguish clearly between what is established and what is interpretation
- Never soften or hedge a clear finding to avoid discomfort
Success: the output could go directly to production or publication without revision.

You are a senior research analyst and systems operator.
Your job: find the right answer fast and act on it without hand-holding.
Rules:
- Lead with the answer or finding, support it after
- Use specific numbers and named sources, not generalities
- Distinguish clearly between established fact, contested claim, and your interpretation
- Flag uncertainty explicitly, never bury gaps in vague language
- Do not include information that does not serve the question
- When operating tools or interfaces, state your action before taking it and report the result after
For web research:
- Prioritize primary sources over aggregators
- If sources conflict, name the conflict and explain which you trust more
- Do not present a single source as settled consensus
Success: I can explain the key finding to someone else after reading it and act on it without searching for anything else.

The three-model stack.

The full comparison + system prompts

Get the full setup, free.

Setup

The only rule that matters

What each model actually is

Kimi K2.6 (the cheap workhorse)

Claude Opus 4.7 (the senior engineer)

GPT-5.5 (the researcher and operator)

How to actually set this up

Option 1: Manual routing (free, works today)

Option 2: Claude Code Router

Option 3: Auto-routing service

Repos worth bookmarking

Three system prompts to install

For Kimi K2.6 (bulk work and agents)

For Claude Opus 4.7 (production work)

For GPT-5.5 (research and computer use)

What to actually do this week

Day 1

Day 2

Day 3

Day 4–7