Product case study · August 26, 2026

Ox Alpha free preview ended: what it became and what to use now

Ox Alpha's free preview has ended and was revealed as Z.ai GLM-5.3-Flash. This archived guide explains the former routes and their data-handling differences.

Reading time
17 min
Checked
Nov 20, 2026
An ochre cartridge beside a charcoal routing switch with muted blue paths leading to three paper evidence cards
A temporary model preview is useful only when its route, limits, and end state are documented
Bottom line

The free Ox Alpha preview is over. OpenRouter now identifies the model as Z.ai GLM-5.3-Flash, and OpenCode Zen no longer lists the preview route. Keep this guide as a record of the evaluation workflow and privacy differences, but do not use its former model IDs as evidence of a current free deal.

Update, August 27, 2026: The free preview has ended. OpenRouter now identifies Ox Alpha as Z.ai GLM-5.3-Flash, and OpenCode Zen no longer lists x-preview-f-free. The configuration below is preserved for editorial history and should not be treated as a live deal or current setup guide.

Ox Alpha appeared without a named developer, a model card, or a launch blog from the lab that built it. During its short preview it offered a million-token context window, tool calling, multimodal input, and a zero-dollar token price on two coding gateways.

The official name was Ox Alpha, with the letter O. You may also see it written as “0x Alpha” in older posts and searches. Its former model IDs were:

OpenRouter: stealth/ox-alpha
OpenCode Zen API: x-preview-f-free
OpenCode model picker: opencode/x-preview-f-free

OpenCode announced on August 20 that the model would be free for “the next week.” That pointed to roughly August 27, but neither gateway published an exact cutoff hour. Both catalogs still showed zero token pricing on August 26; by August 27, OpenRouter had changed to a retrospective identification page and Zen had removed the route. That is why the deals inventory records the preview as ended without inventing a calendar expiry.

What to do now: retire the preview configuration

Sources checked November 20, 2026. OpenRouter still identifies Ox Alpha as Z.ai GLM-5.3-Flash. This refresh checks documentation, not an authenticated model run. The August observations below remain historical; we have not rerun the preview or measured a replacement model’s performance.

There are two changes to understand before copying the archived setup:

  • OpenCode’s former Zen documentation now opens Console models. Its current rules restrict free-tier chat models to OpenCode. For another client, it directs users to paid models or Go. CC Switch’s protocol conversion does not remove that access restriction.
  • OpenRouter’s Claude Code guide now limits its compatibility guarantee to the Anthropic first-party provider. Treat a non-Anthropic model in Claude Code as an integration to evaluate, not a supported drop-in replacement.

1. Find where the expired route is still selected

Finish any active job before changing providers. Record your current provider card and make a private backup of any configuration you intend to edit. Do not share the backup: it may contain credentials.

Check the terminal that launches Claude Code, its shell startup file, ~/.claude/settings.json, project settings, and the active CC Switch provider card for stealth/ox-alpha or x-preview-f-free. Inspect locally rather than dumping environment variables or full settings into a support ticket. Note every role mapping that references the retired model, including background-task and subagent overrides.

A fresh terminal removes exports entered only in the old terminal. It does not undo settings saved in a shell profile, project file, or CC Switch. Remove or replace only the obsolete overrides you identified; preserve unrelated configuration.

2. Choose a currently documented route

If you want free-tier chat access through OpenCode, use its current /connect and /models flow and inspect the selected model’s terms. Do not search for the retired preview ID or assume a different model inherits its privacy policy.

If you want Claude Code through OpenCode Console, the current Claude Code guide documents an Anthropic Messages endpoint at https://opencode.ai/inference/anthropic for paid or Go models. Console handles translation for other model protocols. That is a different setup from the old Zen endpoint and local conversion recipe below. Follow the current guide for your chosen model and review account billing controls first.

If you want to restore your previous CC Switch setup, its switching manual distinguishes the active routing target from the provider used in Direct mode. Confirm the intended Direct provider before leaving Routing. Restart Claude Code if the session still shows the old route, then check /status and /model again.

3. Verify one request before permitting edits

Use a disposable folder containing a single fictional README and ask:

Read README.md and summarize its purpose in one sentence.
Do not edit files, run commands, install packages, or contact other services.

This instruction is not a security boundary; keep normal permission controls. Before proceeding, save a small private receipt with the date, harness version, selected model ID, gateway, observed charge, and whether the folder stayed unchanged. Check the provider’s usage record as well as the client display. A reply alone does not establish which upstream model handled it or what it cost. If accounting is delayed, wait for it before a longer run.

SymptomNext checkStop condition
Model not foundCompare the configured ID and endpoint with the current provider catalog.Do not keep retrying a retired preview ID.
Authentication errorCheck which provider issued the key and which authentication source the session uses.Never paste the key into a diagnostic report.
Unexpected charge or modelInspect role overrides, routing targets, and any fallback configuration.Stop the run until the usage record matches your intended route.
Text works but tools failCheck the provider’s harness support and protocol requirements.Do not infer tool compatibility from one text response.

Only after this check should you reuse the bounded coding comparison later in this article with an available candidate model.

What the preview established

ClaimEvidence statusPractical meaning
1,048,576-token context and up to 131,072 output tokensPublished by OpenRouter’s live model APIEnough capacity for large repository selections, logs, and tool history; not proof of perfect million-token recall.
Text, image, and video input; text outputPublished by OpenRouterUseful for code, screenshots, and visual debugging when the coding harness passes those inputs correctly.
Tool calling and reasoning controlsPublished in OpenRouter’s supported-parameter listThe model can participate in agent loops, but reliability still depends on the harness and gateway translation.
Free on OpenRouter and OpenCode ZenConfirmed on August 26; ended by August 27The zero-token price was temporary and is no longer a live offer.
Built by Z.ai or part of the GLM familyRevealed after the previewOpenRouter now identifies Ox Alpha as Z.ai GLM-5.3-Flash.
Better than a current frontier coding modelNot establishedOne successful demo or bug fix is a reason to test, not a general ranking.

Our Ox Alpha review and video companion documents the strongest observed signal so far: in Ron’s Helm’s Deep test, the model changed tactics when its first approach failed and finished with roughly 200,000 tokens. Comparable channel runs often used much more, but this was not a matched multi-run evaluation. Treat the recorded run as useful evidence, not an invented vendor scorecard.

Pick the route before you configure the app

The same model name does not mean the same service conditions.

Former routeBest use during the previewPublished free access at the timePublished data handling at the timeCatch
OpenCode + ZenFastest supported setup and heavier trial usex-preview-f-free was free for a limited timeZen said this route was zero retention and not used for model trainingZen required billing details; its current model table no longer lists this route.
Claude Code + OpenRouterSimplest direct Claude Code connectionstealth/ox-alpha showed $0 input/outputOpenRouter said the anonymous provider retained prompts and completions but did not train on themThe model page is now retrospective rather than a live zero-price offer.
Claude Code + CC Switch + ZenClaude Code interface with Zen’s preview routeZen kept the route free during the previewZen published zero retention for Ox AlphaRequired CC Switch’s local routing service to translate Anthropic Messages to OpenAI Chat Completions.

During the preview, OpenCode with Zen was the most direct route, OpenRouter was the shortest Claude Code connection, and CC Switch preserved Claude Code’s interface through a local protocol-conversion layer. Those routes are documented below as an archive, not as current instructions.

Archived route 1: Ox Alpha in OpenCode

August 2026 record; do not execute this retired setup. The names, billing example, and model ID below describe the preview period.

OpenCode Zen is the route the OpenCode team has tested for its own coding harness.

  1. Sign in to the official OpenCode Zen console. Zen’s documentation says account setup includes billing details even when the selected model is free.
  2. Before adding credits, inspect auto-reload and the workspace monthly limit. Zen says auto-reload can add $20 when a balance drops below $5 unless you change or disable it.
  3. Install or open OpenCode, then enter /connect in the TUI.
  4. Choose OpenCode Zen and paste the API key locally.
  5. Run /models and select Ox Alpha Free, whose configuration ID is opencode/x-preview-f-free.
  6. Start with a small read-only request: “Inspect this repository, identify the test command, and propose one bounded task. Do not edit files.”

Check the model shown in the session before granting edit or terminal permissions. “Free” should not turn into accidental use of a paid fallback model.

Archived route 2: Claude Code through OpenRouter

August 2026 record; the model ID below is retired. Use the current recovery checklist above for a new configuration.

OpenRouter exposes an Anthropic-compatible API surface specifically for Claude Code. This path needs an OpenRouter API key but no local proxy.

1. Create a scoped OpenRouter key

Create the key yourself in the official OpenRouter Keys dashboard. Give it a recognizable name and a small credit limit if the dashboard offers one. Never paste the real key into a chat, issue, screenshot, article, or committed file.

2. Test in one disposable terminal session

Use temporary shell variables first. Replace the placeholder locally:

export OPENROUTER_API_KEY="sk-or-your-key-here"
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="$OPENROUTER_API_KEY"
export ANTHROPIC_API_KEY=""

export ANTHROPIC_DEFAULT_FABLE_MODEL="~stealth/ox-alpha"
export ANTHROPIC_DEFAULT_OPUS_MODEL="~stealth/ox-alpha"
export ANTHROPIC_DEFAULT_SONNET_MODEL="~stealth/ox-alpha"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="~stealth/ox-alpha"
export CLAUDE_CODE_SUBAGENT_MODEL="~stealth/ox-alpha"

claude

The base URL is https://openrouter.ai/api without /v1. OpenRouter’s Claude Code guide says /v1 causes model-not-found errors. ANTHROPIC_API_KEY must be explicitly empty so Claude Code does not fall back to Anthropic authentication.

The leading ~ pins the exact OpenRouter model instead of allowing provider preference syntax to reinterpret the slug. Mapping every Claude Code role to the same preview model prevents a small background task or subagent from quietly switching to a paid model.

3. Clear an old Anthropic login if necessary

If Claude Code reports an authentication conflict or tries to find the model on Anthropic, run this inside Claude Code once:

/logout

Quit and relaunch claude. This clears a cached Claude login; it does not remove environment variables from your shell profile.

4. Verify the route before allowing edits

Inside Claude Code, run:

/status

You want to see ANTHROPIC_AUTH_TOKEN as the authentication source and https://openrouter.ai/api as the base URL. Then check /model, send one tiny read-only prompt, and confirm the OpenRouter Activity dashboard records stealth/ox-alpha at zero cost.

When the test is over, exit Claude Code and close that terminal. Because the variables were not added to .zshrc, .bashrc, or a project settings file, a fresh terminal returns to your normal provider.

Archived route 3: CC Switch and Zen

August 2026 record; not a current free-access method. Current Console free-tier chat restrictions apply regardless of local protocol conversion.

This route keeps Claude Code as the harness while CC Switch translates its Anthropic Messages requests into Zen’s OpenAI-compatible chat format.

1. Protect the configuration you already have

CC Switch writes the active Claude configuration to ~/.claude/settings.json. Before adding anything, export a CC Switch backup or duplicate your current provider card. If you keep hand-written settings in that file, copy those values somewhere safe and check the resulting diff after the switch.

Do not import provider deep links from strangers. A provider deep link can contain an endpoint and API key; review every field yourself.

2. Add the Zen provider

In CC Switch:

  1. Open the Claude app section—not Claude Desktop or Codex.
  2. Click Add Provider and choose Custom Configuration.
  3. Name it Ox Alpha via OpenCode Zen.
  4. Enter your Zen API key in the API key field.
  5. Set the endpoint root to https://opencode.ai/zen.
  6. Set the default model to x-preview-f-free.
  7. Expand Advanced Options and change the API format to OpenAI Chat Completions.
  8. If role mapping is available, map Fable, Opus, Sonnet, and Haiku to x-preview-f-free so background tasks cannot fall through to a paid model.
  9. Save the provider.

Do not paste the complete /v1/chat/completions URL into the normal endpoint field. CC Switch appends that path in its default prefix mode, so the root above resolves to Zen’s documented endpoint:

https://opencode.ai/zen/v1/chat/completions

3. Turn on the conversion route

Because Zen serves Ox Alpha through OpenAI Chat Completions rather than Anthropic Messages, the CC Switch proxy must stay running:

  1. Go to Settings → Routing → Local Routing.
  2. Turn on the Routing Master Switch.
  3. Enable routing for Claude.
  4. Return to the provider list and click Enable on Ox Alpha via OpenCode Zen.

CC Switch’s default local route is 127.0.0.1:15721. Keep CC Switch open for the whole session. Restart Claude Code if the current process does not pick up the switch.

4. Prove that the conversion works

Run /status in Claude Code, send the same small read-only prompt used in the direct test, and inspect CC Switch’s routing activity. Confirm three things before editing:

  • the request counter increases;
  • the upstream model is x-preview-f-free;
  • the Zen usage record shows zero cost.

If you receive 401 or 403, recheck the Zen key. If you receive 404, make sure the endpoint is the https://opencode.ai/zen root and that Full URL mode is off. If tool calls fail, update CC Switch and Claude Code before blaming the model; this route depends on protocol translation.

To restore your normal setup, enable your previous provider card or an official-login provider, turn off Claude routing if you do not use it elsewhere, and verify /status again.

A useful first test—not a token bonfire

A million-token window invites people to throw an entire private repository at the model. That is the wrong first move. Context capacity is not a quality score, and a free preview is not a security review.

Choose a public repository or disposable project and freeze the starting commit. Give your normal model and a currently available candidate the same task, permissions, and tests. During the preview, Ox Alpha was the candidate. A good task has a real failure mode and a finish line:

Read the repository instructions first.
Diagnose the failing test without changing unrelated files.
Explain the cause in three sentences.
Implement the smallest fix.
Run the relevant test and the nearest broader test suite.
Show the changed files, test results, and any unverified boundary.

Record:

  • accepted changed files;
  • tests passed and failed;
  • tool-call failures and retries;
  • input, output, and cached tokens;
  • wall time;
  • human review and correction time;
  • whether the model changed strategy after a failed approach.

The comparison is useful if this evidence tells you whether the candidate completes your kind of work more efficiently. “It used fewer tokens” is not a win if you spend an hour repairing the diff.

Privacy, limits, and other catches

The route changes the privacy promise

During the preview, OpenCode Zen described Ox Alpha as zero retention with no training use. OpenRouter’s retrospective page says the provider retained prompts and completions without using them for training. These were route-specific terms, not a policy for every later GLM endpoint.

The current OpenRouter data-collection policy, checked November 20, distinguishes opt-in logging from Safety Retention endpoints and CSAM review or legal-preservation exceptions. Review the selected endpoint and account controls before sending repository content; “not used for training” does not mean “never retained.”

Do not send either route secrets, wallet keys, customer data, private contracts, production .env files, unreleased strategy, or a repository you are not authorized to share. For OpenRouter, assume the provider will keep the content for an unspecified period.

Free OpenRouter requests are limited

Check the current OpenRouter limits documentation for the free model you actually select. A request allowance is not a promise of capacity, uninterrupted service, or access to a retired model. A coding agent can consume many requests during one task.

Do not buy credits to try to revive Ox Alpha. Review the key budget and fallback settings before testing a replacement, then inspect actual usage.

Zen had billing controls even while the model was free

Zen asked for billing details and documented an auto-reload feature. Although the Ox Alpha route was free during the preview, other models in the same account were not. That remains the useful lesson: tune auto-reload, add a monthly limit, and verify the selected model before a long session.

The model can change behind the same name

The operator was anonymous during the test window and did not publish a checkpoint history. OpenRouter’s later identification resolves the lab question, but not every routing or configuration detail from the preview. Improved or different output on another day could still come from a backend swap, system-prompt change, routing change, provider load, or sampling variance. Save the date, gateway, model ID, harness version, repository commit, and acceptance test with every result.

Verdict

Ox Alpha was exactly the sort of free preview worth testing: long context, tool support, and enough early evidence to justify one controlled coding task. Its abrupt end also shows why temporary pricing is not evidence of a permanent bargain or production-safe data handling.

Do not configure the former free routes from this archive. If you evaluate the revealed GLM-5.3-Flash model now, start from its current official provider listing and terms, then reuse the same bounded comparison workflow.

For any replacement route, keep the first job public, bounded, reproducible, and reviewed. The finish line is not “used a million tokens.” It is a tested diff you would accept.

Sources

Current migration and policy sources checked November 20, 2026. The archived route instructions and August preview observations are retained as history, not newly tested setup claims.

Put this to work

Separate the model, gateway, coding harness, price, limits, and data policy instead of treating 'free Ox Alpha' as one product.

Try

Retire the expired preview configuration, verify a currently supported route, and run one read-only task in a disposable repository.

Prove it worked

Compare accepted files, passing tests, retries, wall time, token use, and human correction time on the same commit.

Where it can pay

A repeatable model-routing and evaluation workflow helps freelancers protect margin when temporary free tiers disappear.

Keep in view

  • Ox Alpha was a short coding and reasoning preview that OpenRouter later identified as Z.ai GLM-5.3-Flash.
  • OpenRouter and OpenCode Zen offered different zero-cost routes, rate limits, billing controls, and retention terms during the preview.
  • The former model IDs are preserved below for editorial history, not as a current setup recommendation.
  • Current OpenCode free-tier chat models require OpenCode; a local proxy does not make them available to Claude Code.
Learn the workflow: the supervised Claude Code workflow