MorseSign inOpen Morse

The best models for coding agents, and what each one costs

The Morse teamPublished 4 min read

In November 2026, the best models for coding agents on Morse are Kimi K2.7 Code for fast edits, Kimi K3.1 Pro when the agent needs to reason or read screenshots, and GLM 5.3 Flash for cheap, quick changes. All of them take tool calls and long contexts, and they're among 80+ open-weight models behind one API. Morse speaks OpenAI's Chat Completions, Anthropic's Messages API, and OpenAI's Responses API, so Claude Code and Codex run on these models when you change one URL, and Morse keeps no prompt and no reply on any tier.

Every price, context window, and model name on this page is read from Morse's model catalog when the page builds, so it changes when the catalog does.

What a coding agent needs from a model

A coding agent calls tools on almost every turn: it reads a file, runs a test, writes a patch, and reads the output. It also sends a lot of code with each request. Three things decide which model suits it:

  • Tool calls on every request. A model that drops a tool call stalls the agent.
  • A long context window. Big files, long diffs, and test output all have to fit.
  • A low price on input tokens. An agent sends far more than it gets back, so the input price sets most of the bill.

Each model below takes tools on every endpoint a key reaches without a privacy floor.

The models, side by side

ModelContextIn / out per 1MBest for
Kimi K2.7 Code, moonshotai/kimi-k2.7-code262.1K$0.95 / $4Kimi's coding model. It doesn't reason, so it starts writing sooner.
Kimi K3.1 Pro, moonshotai/kimi-k3.1-pro1M$2 / $10Reasons, and reads images.
GLM 5.3, z-ai/glm-5.31M$1.40 / $4.40Reasons, and takes tools at TEE too.
DeepSeek V4 Pro, deepseek/deepseek-v4-pro-08131M$1.32 / $3.96Reasons, at a lower price.
GLM 5.3 Flash, z-ai/glm-5.3-flash1M$0.15 / $0.50Cheap and fast for small edits, and takes tools at TEE too.

Prices are in dollars per million tokens, on each model's default endpoint. Morse's models page lists every model, with its price on each privacy tier.

What a session costs

Take an agent session that sends 2M input tokens and gets 200K back. Priced on each model's default endpoint, it costs:

Model2M in, 200K out
Kimi K2.7 Code$2.70
Kimi K3.1 Pro$6.00
GLM 5.3$3.68
DeepSeek V4 Pro$3.43
GLM 5.3 Flash$0.40

Each figure is the input price times two plus the output price times 0.2. No subscription. Your first top-up gets 50% extra.

Set up your agent

Put your key in MORSE_API_KEY, and point the tool at Morse. The examples use Kimi K3.1 Pro. Any model id from the table works in its place.

Claude Code

export ANTHROPIC_BASE_URL=https://inference.trymorse.ai
export ANTHROPIC_AUTH_TOKEN="$MORSE_API_KEY"
export ANTHROPIC_MODEL=moonshotai/kimi-k3.1-pro
claude

To check it works, ask Claude Code to change one line in a file. It reads the file and edits it with its tools, so a finished edit shows that the request and its tool calls went through. The Claude Code page in Morse's docs has the steps in full.

Codex

Add this to ~/.codex/config.toml, then run codex in your project's folder:

model = "moonshotai/kimi-k3.1-pro"
model_provider = "morse"
web_search = "disabled"

[model_providers.morse]
name = "Morse"
base_url = "https://inference.trymorse.ai/v1"
env_key = "MORSE_API_KEY"
wire_api = "responses"

To check it works, ask Codex to change one line in a file. It writes the change with apply_patch, so a finished edit means the tool call came back as Codex expects. Leaving web search off keeps Codex to the tools it runs on your machine. The Codex page has the rest.

OpenCode and Cline

Both take an OpenAI-compatible provider. Give it the base URL https://inference.trymorse.ai/v1, the key from MORSE_API_KEY, and a model id from the table, such as moonshotai/kimi-k3.1-pro. OpenCode reads them from opencode.json, and Cline from its settings under OpenAI Compatible. The OpenCode and Cline pages have whole configs, with each model's limits filled in.

To check it works, ask either one to fix a typo in your README. A fixed typo means the agent read the file and wrote the change through Morse.

Give each tool its own key with a spend limit, so one runaway agent can't spend all your credits. Limit a key shows how.

Keep an agent's code private

Morse keeps no prompt and no reply on any tier. To bill you, it records the model, the endpoint that served it, the token counts, the cost, and when the request ran, never what was in it.

To keep an agent on attested hardware, set a TEE floor on its key in the console, and pick GLM 5.3 or GLM 5.3 Flash, whose TEE endpoints take tools. Each request the key sends then runs inside a hardware enclave Morse checks first, or Morse refuses it. Morse never sends it to a lower tier. Tools on TEE and E2EE explains which tiers take tools.

Anyone can check the enclave Morse runs with one command, npx @askmorse/verify inference.trymorse.ai, before an agent sends it a line of code. The quickstart makes your first key, and the Morse docs cover streaming, tool calls, and errors.

Common questions

What is the best model for a coding agent?

On Morse, Kimi K2.7 Code for fast edits, Kimi K3.1 Pro when the agent needs to reason or read screenshots, and GLM 5.3 Flash for cheap, quick changes. All three take tool calls and long contexts.

Can I use these models in Claude Code?

Yes. Set ANTHROPIC_BASE_URL to https://inference.trymorse.ai, put your Morse key in ANTHROPIC_AUTH_TOKEN, and set ANTHROPIC_MODEL to a model id from the table. Morse speaks Anthropic's Messages API.

Can I use them in Codex?

Yes. Add Morse as a provider in ~/.codex/config.toml with wire_api set to responses and your key in MORSE_API_KEY. Morse speaks OpenAI's Responses API.

Which of these models take tools on attested hardware?

GLM 5.3 and GLM 5.3 Flash. Set a TEE floor on the agent's key, and every request it sends runs on attested hardware or is refused.

How much does an agent session cost?

It depends on the model and on how much code the agent reads. The worked example on this page prices a session that sends 2M tokens and gets 200K back on each model.

Does Morse keep my code?

No. Morse keeps no prompt and no reply on any tier. To bill you, it records the model, the token counts, the cost, and when the request ran.

How do I stop an agent from spending too much?

Give it its own key with a spend limit in the console. Once the key reaches its limit, Morse refuses the key's requests.

How do I pay?

Per token, at each model's listed price, from credits. No subscription. Your first top-up gets 50% extra.

Ask what you actually want to ask.

80+ models, characters, and images, with every chat encrypted and kept on your device. Free to start.

Start a private chat

Read next