Codex vs OpenCode vs Crush: 2.9x cost gap

Model: Kimi K3.1 Pro · ✓ Verified on Codex CLI 0.154.0, OpenCode 1.19.5 and Crush 0.98.1 on the same Wallaby endpoint · 2026-11-16

Codex CLI, OpenCode and Crush each fixed the same bug on Kimi K3.1 Pro, with the same five unit tests passing. The bills were $0.034, $0.070, and $0.097: a 2.9x gap between the lowest and highest run.

The model and task stayed fixed. The request count did not: Codex made 3 requests, OpenCode made 6, and Crush made 7. Input tokens ranged from 29,032 to 57,415.

This one-task pilot is not a tool ranking. It shows why the price per million tokens cannot predict a coding-session bill. The tool decides how often to call the model and how much repository context to send.

Disclosure: Wallaby sells API access to Kimi K3.1 Pro. These runs used our endpoint, and the costs below come from our gateway logs. We first published the run table in our full comparison of Cursor alternatives with BYOK support; this article examines what the tool differences mean for a buying decision.

The test

We gave each tool a small Python repository containing a three-line function with a deliberate bug and five unit tests. The instruction was the same each time:

Run the tests, fix the bug without touching the test file, run them again.

Each agent ran on 16 November 2026 with Kimi K3.1 Pro, a fresh configuration directory, and no human edits during the run. Success meant all five tests passed after the agent edited the function.

Agent Version Tests after Time Requests Input / output tokens Billed
Codex CLI 0.154.0 5 of 5 pass 47 seconds 3 29,032 / 638 $0.034
OpenCode 1.19.5 5 of 5 pass 76 seconds 6 41,831 / 1,075 $0.070
Crush 0.98.1 5 of 5 pass 68 seconds 7 57,415 / 888 $0.097

The bills use Wallaby's Kimi K3.1 Pro rates: $1.80 per million fresh input tokens, $0.18 per million cached-input tokens, and $9.00 per million output tokens. Cached input is included in the input counts and billed at the lower rate.

Same model, 2.9 times the bill

Codex was fastest and least expensive in this run. OpenCode billed about 2.1 times the Codex run; Crush billed about 2.9 times. All three produced the same passing-test outcome.

Tool choice changed the number of model calls and the amount of context sent even though the model and task stayed fixed. Our aggregate logs show that difference. They do not identify which system prompt, tool definition, or retry caused it, so we are not assigning the gap to a specific implementation detail.

For a buyer, compare tools on cost per accepted task. Count the runs that needed another prompt or human repair. A tool that finishes at $0.070 can cost less in practice than one that bills $0.034 but needs a second run.

For a cross-model application of that metric, see Gemini 4 Neon vs Kimi K3.1 Pro: the public comparison changes both the model and the harness, so the cost gap belongs to the full setups.

When the client counter disagrees with the invoice

Codex's footer reported 4,199 tokens used. The gateway recorded 29,670 input plus output tokens for the same run, about 7.1 times the client counter.

Those counters serve different purposes. The coding tool can show a session estimate or its own token view; the provider's log records what it billed. When they disagree, use the provider's itemized bill for budgeting. The 7.1x ratio describes this run, not a general error rate for Codex.

This also applies to built-in dollar estimates. A client that does not know your provider's current model rates can only estimate. Our itemized-billing guide shows the input, cached-input, output, and thinking lines that appear on a Wallaby request.

The next cost question is the monthly route

This pilot compares tool-level costs on one model. If you are deciding between a Kimi Code membership and pay-as-you-go API for your whole month, our Kimi Code subscription vs API analysis turns these observed bills into monthly arithmetic and shows where the $19 subscription price lands.

Which tool fits

Codex CLI fits if you already use OpenAI's terminal workflow or want the lowest bill observed in this pilot. It took 47 seconds and made 3 requests here. That is one observed run, not a latency promise.

OpenCode fits if model choice and provider choice drive the decision. Its catalog and custom-provider support let you change the model without changing the terminal workflow. Our run cost $0.070 and took 76 seconds.

Crush fits if you prefer its terminal interface and want model prices attached to the provider configuration. Our run cost $0.097 and took 68 seconds. In this non-interactive test, Crush edited the file and ran tests without a permission prompt. Use a disposable branch when you evaluate that behavior on your own machine.

The Kimi K3 CLI tool table covers more agents and the configuration details for each. If you are choosing an editor rather than a terminal agent, our Cursor test also explains where Cursor's plan restrictions and Tab completion differ from these tools.

What one passing test cannot answer

This pilot says nothing reliable about large repositories, multi-file refactors, frontend design, or long autonomous sessions. It does not compare subscriptions, cloud agents, security controls, or team administration. One run per tool also leaves no way to separate a stable cost difference from ordinary run-to-run variation.

A useful follow-up would run the same 10 tasks through all three tools, repeat each task, and publish failures as well as passes. The primary measure should be cost per accepted task, with time and request count beside it. Until that exists, treat this table as a receipt from three runs, not a league table.

Try the same comparison on your repository

Pick one bug with tests already written. Record the expected result before the agent starts. Run the same prompt through two tools with the same model, then compare the provider's logs:

  • Did all tests pass without human edits?
  • How many requests and tokens did each run use?
  • What was the elapsed time?
  • What was the cost per successful run?

New Wallaby accounts receive $0.50 in trial credit with no card required. That covers five runs even at the highest $0.097 bill above. Use a test key with a small spending cap and a repository branch you can discard. Create a free account and run the comparison on a task you already understand.

Sources: Codex CLI 0.154.0, OpenCode 1.19.5, and Crush 0.98.1 runs against api.wallabytoken.com on 16 November 2026; Wallaby gateway logs for those runs; Wallaby pricing.json, rates verified 5 November 2026. The run table also appears in our Cursor BYOK alternatives article. Tool behavior and prices change; repeat the test on current versions before making a purchasing decision.

Like the Dispatch? Add Wallaby as a preferred source on Google, or subscribe via RSS.
Did this help?