Jordan Urbs did not cancel his $200 Claude subscription because Claude Code stopped working. He cancelled because it had become the default for his projects, and he wanted to know whether he could leave without rebuilding his entire workflow. In his video, he describes a move toward OpenCode, multiple open-weight models, task-based routing and Venice Max. It is an ongoing experiment, not a measured proof that the replacement is cheaper, more private or equally capable for everyone.
Keep the workflow you can move: project instructions, skills, tests, data and a record of completed tasks. Then test another harness and provider on the same work. Jordan's reported API-equivalent bill is a warning to examine dependence, but public token prices do not reveal Anthropic's internal costs or predict a plan change. Today, the former $200 Max 20x plan is called Claude Ultra, while Venice Max is a separate $200 plan with a $225 monthly metered API allowance. Those are different products with different limits.
Watch Jordan's Account
Disclosure: Jordan says he works for Venice. He calls this video unsponsored, but he is an advocate for the product and provides a creator-specific Venice link. Read his product recommendation with that interest in mind. This article is not sponsored by Venice and did not reproduce his migration test.
What His Token Math Does and Does Not Show
Jordan reports two days above 700 million tokens each and estimates those two days would cost more than $50,000 at a simple public API rate, or roughly $1,800-$2,000 after caching. The video does not include a reproducible split of input, output, cache reads, cache writes, models and tool calls. The amounts are his estimates, not invoices. More importantly, API list prices are what a customer is charged, not the provider's cost of serving a subscription user. The calculation cannot establish that Anthropic lost money on his account or will change the $200 plan.
For a concrete reference, Anthropic lists these Opus 5.5 API rates as of 2 November 2026. They are for API usage, not a conversion table for Ultra's included use.
| Token category | Opus 5.5 API rate per million | Why it changes the estimate |
|---|---|---|
| Uncached input | $4 | New prompt and context tokens use the full input rate. |
| Output | $20 | Generated tokens, including billed reasoning, cost much more than input. |
| 5-minute cache write | $5 | Storing reusable context costs more than ordinary input. |
| Cache read | $0.20 | Repeated context can be far cheaper, if the cache actually hits. |
The correct comparison uses the usage fields for each category, multiplied by its own rate, plus any other billed features. A large total token count alone cannot produce one credible dollar figure. Also compare completed work per month, not just theoretical token spend: a fixed-fee plan with usage limits and a metered API allowance expose different risks.
The Real Lock-In Is in the Workflow
Jordan's strongest point is practical. After a year of projects, instructions, skills and habits, he opened Claude Code before considering another tool. A substitute model does not automatically move that context, reproduce a tool call or pass the same tests. The portable part is the material you control: repository files, documented decisions, prompts, tests and examples. The harder part is vendor-specific behavior: authentication, memory, tool permissions, hosted features, model judgment and the way an agent interprets instructions.
OpenCode documents AGENTS.md as persistent project guidance and supports named agents with model preferences and permissions. That makes an alternative setup possible, not a guaranteed drop-in replacement. In OpenCode V2, CLAUDE.md is not a fallback for AGENTS.md; move or reference important project rules deliberately. Check the installed version and test each skill, tool and permission rule. Do not move secrets into a shared instruction file merely to make migration convenient.
Route Work by Job, Then Measure It
Jordan proposes a stronger model for planning, a cheaper one for implementation and a very inexpensive model for repetitive work. That is a hypothesis worth testing against your own jobs. A cheap run that needs several retries, more supervision or an expensive rescue pass may cost more in total.
| Work | Candidate route | Acceptance check |
|---|---|---|
| Architecture and risky decisions | High-capability planner plus human review | Trade-offs, edge cases and rollback plan are explicit. |
| Routine implementation | Lower-cost coding model | Tests pass; the diff stays within scope. |
| Mechanical updates | Fast, cheap model or deterministic script | Output matches an exact schema or golden example. |
| Final review | Independent reviewer, human for high-risk work | Defects and unsupported claims are caught before merge. |
OpenCode's provider selector and agent profiles support trying this division. Keep a spending cap and measure accepted results, elapsed time, retries and human corrections alongside the invoice.
What the Venice Alternative Actually Buys
Venice currently lists Max at $200/month with $225 of monthly API spend, with three-month banking up to $675. API requests are metered at the chosen model's rates; the allowance is shared across API keys and metered Venice features. Its public catalog lists hundreds of text and media models, but the count is not a promise that every model is suitable for coding or is offered under the same privacy terms. Venice's OpenCode integration guide shows a custom provider connection. Test a small request and inspect the balance before routing a full agent session.
A local harness is not local inference. OpenCode may run on your computer while prompts and code excerpts go to a remote model API. Venice also documents four distinct privacy modes. In its anonymous mode, an upstream model provider can see and may retain prompt content; its private mode relies on Venice or zero-retention partner commitments; TEE and end-to-end encrypted modes have different technical protections and feature limits. Check the mode for each model you select and separately audit web search, MCP servers, logs, plugins and shell tools. Jordan's concern about profiling is his view, not proof that a named competitor builds a psychological profile of him.
A One-Week Portability Test
Before cancelling, pick three real tasks: one planning job, one ordinary code change and one repetitive edit. Save the expected result and an existing good example. In a disposable project, document the current workflow, connect one alternative provider, and run the same tasks with identical inputs where practical. OpenCode's /connect and /models flow is a starting point; for Venice, follow its provider-specific guide and store the API key outside the repository.
Record actual billed usage, elapsed time, retries, corrections, tests and whether the result was accepted. Run read-only first; approve edits only after confirming tool permissions and a rollback point. If one route wins on repetitive work but loses on architecture, use a hybrid setup. If the pilot exceeds its cost cap or mishandles private data, stop. Owning your workspace is valuable even if you keep the Claude subscription.
Copy a Portability Audit Brief
This original brief turns Jordan's dependency question into a bounded comparison. It asks for evidence before recommending a switch.
Audit your agent stack
Three tasks, actual costs and a clear stop rule.
Video Chapters
The reason for leaving
| Time | Topic |
|---|---|
| 00:00 | The $200 power-user confession |
| 01:56 | Token math and the estimated API bill |
| 03:03 | Subsidy and lock-in argument |
| 04:36 | Audit how much work depends on Claude |
| 05:36 | Why replacing Claude is the wrong question |
The proposed stack
| Time | Topic |
|---|---|
| 07:01 | Open-weight models and OpenCode |
| 08:09 | Privacy and the home-network example |
| 09:38 | Routing models by task |
| 11:53 | A concern about the next generation |
| 12:17 | Venice Max experiment |
The open question
Source Note and Limits
Checked 2 November 2026. Jordan's video and supplied transcript are the source for his usage, cancellation, motivations and ongoing Venice experiment. Find his own updates on X and Substack. His API-equivalent estimates and privacy concerns were not independently audited. Current plan names, limits and prices were checked against Claude Ultra, Anthropic API pricing, Venice API plans and Venice privacy documentation. Workflow capabilities were checked against OpenCode's V2 documentation and Venice's integration guide. The cost worksheet, routing checks and audit prompt are our analysis, not claims that Jordan completed those tests. Prices, plan terms, model catalogs and privacy modes can change.