Home › Issues › Issue #001

Issue #001 · Oct 20, 2026

BenchSmoke #001: OpenCode 2.0 builds a to-do app, twice

Oct 20, 2026 · AI tools, actually tested. Every Tuesday and Thursday.

Researched and drafted by AI agents, approved by a human before sending. How we test.

Welcome to issue #1. BenchSmoke is a newsletter for builders, and it comes out every Tuesday and Thursday. In each issue we use AI tools on a real small task and tell you what happened, with one verdict per tool: Use it, Watch it, or Skip it. See how we test.

This first issue is short. We planned to test 3 tools and finished 1, so you get 1 full review. The other 2 are cut until we can test them properly.

Issue #001 scorecard. Use it: 1 (OpenCode 2.0). Watch it: 0. Skip it: 0. We planned to test 3 tools and finished 1; the other 2 were cut until we can test them.

The Short Version

Best find: OpenCode 2.0: built a working to-do app for free, with no account or API key, on both of our runs.
Biggest letdown: None this issue. We only finished one test, so there's nothing to compare it against.
Try this: Give a coding agent a small task, then check whether it left a server running on your port after it says it's done.

Tested This Issue

Task: our standard AI coding task. Build a small to-do app with add, complete, delete, and saved state, and run it locally.

OpenCode 2.0: Use it

What it does: OpenCode is a free, open-source (MIT) AI coding agent with a command-line tool. Version 2.0's upgrade path went out Oct 11, 2026.
What we built: The standard to-do app, run twice from the same prompt in fresh folders. We used OpenCode v2.0.29 with the free "Big Pickle" model (opencode/big-pickle) from OpenCode Zen, its pay-as-you-go model service. The docs now call it Console. No login, no API key. Tested Oct 15, 2026.

Animated screen recording of the to-do app OpenCode built in run 2, driven by a script in headless Chrome. A task, Buy milk, is typed and added to the list. It is checked off and shows struck through. The page is reloaded and the task is still there. Then it is deleted. — The app OpenCode built in run 2, driven by a script in headless Chrome: add a task, check it off, reload the page, delete it. We typed the sample task.

What we tested, and what we didn't. We only tested the opencode run command with a free model, on one standard to-do-app task, run twice. We used --auto (auto-approve tool permissions) because run is non-interactive. We did not test the 2.0 headline features: the background service across terminal, web and desktop, /pair, Code Mode, background subagents, plugins, or the desktop app. We also didn't test moving from v1 to v2, paid models, the Go plan, or how easy it is to install (it was already installed on our test machine).

What worked (what we saw):

The free, keyless model path worked on the first try. No account needed.
Both runs finished the task with 0 retries: 23 seconds for run 1 and 121 seconds for run 2.
The agent wrote its code, ran it, and tested it on its own.
We checked both apps ourselves. Adding, completing and deleting tasks worked. Tasks were still there after a server restart and after a page reload.
In run 2, it noticed a port conflict and fixed it on its own (more in Vibe Coding Corner).

Bar chart of OpenCode 2.0 time to a working to-do app, same prompt, fresh folders, 0 retries. Run 1: 23 seconds. Run 2: 121 seconds. Measured with a timer around opencode run on Oct 15, 2026.

Terminal output from our saved log of run 1's own check. A task is marked complete, the server is killed and restarted, and the list after restart still shows Buy milk with done true and Walk the dog with done false. Deleting the first task returns ok true, and the index page returns 200. Long lines trimmed. — Run 1's own check, from our saved log: the completed task was still there after the server restarted. Long lines trimmed.

What didn't (what we saw):

In run 1, the agent said it had cleaned up, but it left its test server running on port 3000. It had deleted the data file, not stopped the process. We stopped it by hand. Minor, but it caused a problem in run 2.
opencode models failed with a "port already in use" error because an earlier OpenCode background service was running on our shared test machine. Adding --standalone fixed it. This may be caused by our setup, not the product. We haven't confirmed which.
The 2 runs produced different code (different ID styles, and run 2 added a README). Both worked, but don't expect the same output twice.
Big Pickle is labeled "free on OpenCode for a limited time." OpenCode's models page also says data collected during its free period may be used to improve the model. Keep that in mind before you paste in private code.
We didn't hit a rate limit, but 2 runs isn't a rate-limit test.

Price (checked Oct 15, 2026; we only used the free tier, so the paid prices below come from OpenCode's pages, not from our testing):

OpenCode itself: free and open source (MIT license), per FounderMag.
Models via Zen/Console: pay-as-you-go, priced per 1M tokens, with free models listed as Free, per OpenCode's pricing table. Card fees are passed through at 4.4% + $0.30 per transaction. Auto-reload adds $20 when your balance drops below $5, and you can turn it off.
OpenCode Go: $10/mo subscription; Go Plus $40/mo with higher limits. We did not try either.

Animated verdict reveal. OpenCode 2.0 gets Use it, not Watch it or Skip it, with medium confidence: a narrow slice of one command, one free model, one small task. Tested Oct 15, 2026, 2 runs, 0 retries.

Verdict (our opinion): Use it, with medium confidence. It finished the task twice, for free, and the apps worked when we checked them ourselves. Our confidence is medium because we tested a narrow slice: one command, one free model, one small task. Good for builders who want a free coding agent and are fine working from the terminal.

Disclosure: The maker did not give us free access. We used the public free tier.

Vibe Coding Corner

This issue's build: A small to-do app, built by sending one prompt to OpenCode 2.0. These are the same 2 runs as the review above.

Tool(s) used: OpenCode v2.0.29, free Big Pickle model, opencode run --standalone --auto.
Prompt (exact, used for both runs):
Build a small to-do app in this directory. It must let me add tasks, mark tasks complete, and delete tasks, and it must save state so tasks are still there after a restart. It must run locally with a single command (tell me the command). Keep it simple, no paid services or API keys.
Time taken: 23 seconds (run 1) and 121 seconds (run 2), measured with a timer around opencode run. Checking each result ourselves took a few minutes.
What it made: Run 1 produced a Node server with no dependencies (server.js, 121 lines) and a single HTML page (112 lines). It told us to run node server.js, which serves the app on port 3000.

Screenshot of run 1's to-do app in a browser at localhost:3111: a plain white page with a To-Do heading, a text box with an Add button, and two tasks. Buy milk is checked and struck through. Walk the dog is unchecked. Each row has an x button to delete it. — Run 1

Screenshot of run 2's to-do app in a browser at localhost:3222: a light grey page with a To-Do heading, a note that tasks are saved to disk, a card with a text box and a blue Add button, and two tasks. Buy milk is checked and struck through. Walk the dog is unchecked. Each row has a Delete button, and a counter sits at the bottom. — Run 2

Same prompt, two different designs. Screenshots of each run's app on our test machine; we typed in the two tasks.

Where it broke: Run 1 left its test server running on port 3000. When run 2 started, its first test hit that old server instead of the new one. The agent noticed ("a different process is already listening on port 3000"). It then added an automatic switch to another port, safer file saving (write to a temp file, then rename), a 400 error for empty tasks, and a 404 for unknown task IDs.

Terminal output from our saved log of run 2's first test. The new server crashed with Error: listen EADDRINUSE: address already in use :::3000, and a check of who is on port 3000 shows another node process already listening. Some lines trimmed. — Run 2's first test, from our saved log: port 3000 was already taken by the server run 1 left running. Some lines trimmed.

Result: Both apps passed our own checks for add, complete, delete and saved state.
Takeaway: The build was quick and it worked. After an agent says "done," check what it left running.

Agent Watch

Bubble launches Agent 2 and Bubble MCP

What changed: Bubble launched Agent 2 and Bubble MCP (beta) on Oct 6, 2026, alongside AI credits. Per Bubble's post, the free plan includes 500 AI credits per month, and the web-only Starter plan rises to $39/mo ($32/mo billed annually) starting Oct 20.
Should you care? Not yet, unless you already build on Bubble. In that case, check the Starter price change. Bubble says current Starter Web plans change at the next renewal date.
Reporting news only. We did not test it, because it needs an account.

Quick Hits

News, not tests. We haven't used any of these.

Pi 1.0: Earendil shipped its open-source coding agent (MIT) on Oct 1. v1.1.0 followed on Oct 7 and v1.2.0 on Oct 12. We couldn't finish a test in time, so no verdict yet. Pi 1.0 post
Buildful: A background agent that opens tested GitHub PRs. The vendor lists a free plan with 3 tasks and Pro at $10/mo for 1,000 tasks. Buildful launch post
Buildra v2.0: An AI app builder, posted to Show HN on Oct 7. The free tier is listed as 5 credits/day, up to 25/mo and 3 projects, with no card needed. Buildra's site
Coderblock: Launched on Product Hunt Oct 7. It builds full-stack apps from prompts and pitches an "AI team in a 3D office." Coderblock launch post
Cascade: An AI filmmaking platform, out of beta as of Oct 8. New accounts start with 3,500 free credits. Cascade announcement
Moyai: LiteLLM open-sourced it on Oct 7. It needs Modal plus a LiteLLM gateway. Moyai on the LiteLLM blog
Agent37: Sandboxes for agents, posted as a Launch HN on Oct 12. Agent37 on Hacker News

Reader Question

This is issue #1, so we don't have any reader questions yet.

Your turn: What's one AI tool you want us to test, or one that wasted your time? Hit reply. We read every reply, and reader picks go to the top of our test list.

Reply with a tool you want tested

Read how we test. Affiliate links are labeled and only go to tools rated "Use it." This issue has no affiliate links and no sponsor. Sponsors never affect verdicts. BenchSmoke is researched and drafted by AI agents and approved by a human before sending.