Kilo Pass vs Anthropic API direct: the cost math

On January 9, 2026 Anthropic blocked third-party tools, including Kilo Code, OpenCode, and OpenClaw, from authenticating with Claude Pro and Max subscription OAuth tokens. The policy was formally documented on February 19, 2026. Pro and Max subscriptions are now valid only for the official Claude Code CLI and claude.ai (The Register). Tokens that worked with Kilo on January 8 stopped working on January 9.

Kilo Code launched a marketing campaign called “Claudepocalypse” (kilo.ai/claude-pocalypse) to capture the displaced users. The pitch: keep using Claude through Kilo Pass, a credit-based platform that resells Sonnet, Opus, GPT, and a few hundred other models at claimed discounted rates. “Claudepocalypse” is Kilo's framing, not Anthropic's.

We ran the math. For a steady moderate Claude user, Kilo Pass Starter comes out a few dollars a month cheaper than going direct on the Anthropic API once its bonus credits ramp, heavy users save a little more, and light or uneven users pay more. If you can run a local model, a free desktop IDE takes the inference bill to $0, which neither option matches. Here is the worked example.

What did Anthropic actually change?

The change was about credential portability, not access. Anthropic's API is open to anyone who pays. Sonnet 4.5 and 4.6 list at $3 per million input tokens and $15 per million output tokens (platform.claude.com pricing). That has not changed. The newer Claude Sonnet 5.5 lists at $2/$10, which makes going direct cheaper still.

What changed is who can use a Claude Pro or Max subscription. Before January 9, third-party tools could route requests through a user's Pro subscription using OAuth. After January 9, those tokens only work with Anthropic's own clients. Third-party harnesses had to either switch to direct API billing (paying per token) or move users onto a different middleman that handles its own billing.

Kilo Code chose the second path. Kilo Pass is the result.

What does Kilo Pass actually offer?

Kilo Pass is a credit-based platform with three tiers (kilo.ai/pricing):

Credits are spent at provider list rates. Kilo's pricing page says there is no markup, so the bonus credits are the whole discount. Kilo Pass is billed monthly, cancel anytime. On monthly plans the bonus starts at 10%, pays a one-off 50% welcome bonus in month 2, and reaches a 40% cap from month 8 of continuous subscription; the optional annual plan pays 50% every month. Bonus credits expire at the end of each month. Base credits roll over but expire after a year. Every Kilo Pass purchase, monthly or annual, and every credit top-up carries a 5% processing fee.

Separate from Kilo Pass, Kilo used to sell KiloClaw, a $55-a-month hosted agent. It was never part of this coding comparison, and Kilo is now retiring it: the September 2026 Kilo Code release (v7.7.3) removed it from the extension and CLI, and it no longer appears on Kilo's pricing page.

One important clarification: on March 23, 2026 Kilo Code ended its free hosted-inference tier (blog.kilo.ai). The core VS Code, JetBrains, and CLI extension is still free and open-source under Apache 2.0 for users who bring their own API keys. The change closed off Kilo's subsidized cloud inference, not the extension itself.

The math: 5 million tokens a month

A moderate Sonnet user running an agentic coding workflow burns roughly 5 million tokens a month. The split skews toward input because agents read a lot of files and emit short edits: assume 4 million input tokens and 1 million output tokens, the typical 80/20 ratio for tool-heavy workflows.

At Anthropic's list price:

On Kilo Pass Starter, you pay $19.95 a month ($19 plus the 5% fee) for $20.90 in credits in month 1, rising to $26.60 by month 8 of a continuous subscription. Even fully ramped, that runs just short of $27 of usage, so you top up the gap, and the top-up carries the same 5% fee. Month 1 lands at about $26.36. A fully ramped month lands at about $20.37.

At the $49 Pro tier ($51.45 with the fee), you get $53.90 to $68.60 in credits. That is headroom you do not need at 5 million tokens: you pay $51.45 a month for $27 of usage, and the unused bonus expires at month end. Unused base credits roll over, but expire after 12 months.

Where does Kilo Pass break even versus the direct API?

We can frame the comparison as “monthly cost to deliver $X in equivalent Sonnet inference”:

Ranges run from a fully ramped month (40% bonus, low end) to month 1 (10% bonus, high end). The pattern is consistent. Light users pay more on Kilo Pass than direct, because the $19 floor dominates a small bill and the unused bonus expires; skipping Kilo Pass for pay-as-you-go credits still costs direct price plus 5%. From moderate usage up, Kilo Pass comes out cheaper than direct, by about 2% in month 1 and up to about 25% once the bonus is fully ramped. That saving is entirely bonus credits minus the 5% fee: Kilo says it charges list rates with no markup, so there is no per-token discount underneath.

Two caveats. First, the math takes Kilo's no-markup statement at face value and prices credits at Anthropic list rates; we cannot audit Kilo's per-model billing from outside. Second, the bonus is fragile. It expires at the end of each month, and the ramp depends on an unbroken subscription. If your usage is uneven, paying for a tier you do not burn through is dead money.

The three-year math: Bodega One Code enters the picture

Stretch the comparison to 36 months and the picture changes again, because Bodega One Code is free to use, commercial work included, with no recurring cost. BYOLLM means you bring your own Anthropic API key and pay Anthropic directly.

On pure dollars, a steady moderate user does best on Kilo Pass Starter with small top-ups: up to about $239 less than going direct over three years, if the subscription never lapses and the bonus never goes unspent. Overbuy to Pro for headroom and you pay about $880 more than direct over the same stretch. Bodega One Code adds nothing on top of direct API, for personal or commercial work, so its row matches the direct row exactly. Where Bodega One Code changes the math is local models: run an open-weight model on your own GPU and the inference line drops to $0. That is not Sonnet on every task, but no Kilo tier gets you to $0.

The deeper comparison is who you are paying. With Kilo Pass, you pay Kilo, plus 5%, and Kilo pays Anthropic. With BYOK, you pay Anthropic. With Bodega One Code plus BYOK, you pay Anthropic and run a free desktop IDE that does not depend on a third-party subscription chain.

When does Kilo Pass actually win?

We are not arguing Kilo Pass is a bad product. It is a routing layer with real value for specific use cases. Naming them honestly:

• Heavy multi-provider routing. If you genuinely route across Sonnet, Opus, GPT, Gemini, and a handful of open-source models in the same week, the routing layer saves you from managing five API key relationships and five separate bills.
• One bill for accounting. Solo contractors and small studios that expense AI usage to clients benefit from a single line item. A consolidated invoice is sometimes worth a 5% processing fee.
• Steady usage you never skip. If you reliably spend your tier's credits every month, the bonus structure comes out ahead of direct API pricing at every tier from moderate up, and the annual plan's 50% bonus widens the gap.
• Already on the Kilo extension and happy with it. The friction of switching workflows is real. If Kilo Code in VS Code or JetBrains fits your workflow, staying inside the same tool while paying Kilo Pass instead of Anthropic Pro is a small change.

These are real wins. The size matters, though: for a moderate solo Claude user, the steady-usage win is a few dollars a month, and it disappears the first month you skip.

The bigger pattern: subscription roulette

The January 9 Claude Code restriction is one event in a longer pattern. Every AI coding tool that depends on a third-party subscription chain is one policy change away from forced migration. A short selection from the last 12 months:

• January 9, 2026. Anthropic blocks third-party tools from authenticating with Claude Pro and Max OAuth tokens. Documented February 19 (The Register).
• March 23, 2026. Kilo Code ends its free hosted-inference tier (Kilo blog).
• April 24, 2026. GitHub starts using Copilot Free, Pro, and Pro+ interaction data, including prompts, suggestions, code snippets, and navigation, to train its AI by default (GitHub blog). Opt-out lives in Privacy settings.
• March 2026. Windsurf moved from monthly credits to daily-and-weekly quotas, mid-contract, after the Cognition acquisition.
• September 2026. Kilo starts retiring KiloClaw, its hosted agent, removing it from the Kilo Code extension and CLI in v7.7.3 (GitHub).

We have walked through the broader subscription cost pattern in our 3-year subscription analysis. The honest read: tool-by-tool cost math matters, but the structural risk is harder to price. The user did nothing wrong on January 8. On January 9, their workflow stopped working. That kind of breakage cannot be modeled in a per-token rate sheet.

What we'd actually recommend

Three named recommendations, depending on where you are.

1. Stay on Kilo Pass if you genuinely route across many providers, your usage is steady enough to spend the bonus every month, and the single-bill, single-extension workflow is worth the 5% fee to you. Kilo's extension is mature, the routing layer is real, and every tier is competitive when you actually use it.

2. Go BYOK direct if your usage is light or uneven. Use the free Kilo Code core extension in BYOK mode, or Cline. Pay Anthropic directly at $3 per million input and $15 per million output. No platform fee, no processing fee, no expiring credits, no monthly minimum. For a light user, or anyone who skips weeks at a time, this is the cheapest cloud path month over month. Our broader Kilo Code alternatives post covers the BYOK options in more detail.

3. Move off the extension model entirely if you want to stop renting the surface your AI runs on. Bodega One Code is a standalone Electron desktop IDE: free to use, commercial work included. Monaco editor, AI chat, autonomous coding agent, 10+ provider presets including Anthropic direct. Air-gap mode with nine enforcement layers if you need to verify nothing leaves your machine. No subscription, no platform fee, no January 9 risk on the IDE itself. You bring the model, and the tool runs on your machine. Free to use, and the beta is open to everyone. Full launch coming later this year.

For more context on Kilo specifically: our Kilo Code comparison page, the full Kilo Code pricing breakdown, and the Roo Code shutdown alternatives post if you came to Kilo via the Cline → Roo → Kilo lineage.

Sources

• Anthropic third-party access restriction (January 9, 2026 OAuth block; documented February 19), The Register, February 20, 2026: theregister.com
• Anthropic Claude API pricing (Sonnet $3/$15 per million input/output; Sonnet 5 $2/$10 introductory through Aug 31, 2026): platform.claude.com pricing
• Kilo Pass tiers, credits, and bonus structure (Starter $19, Pro $49, Expert $199, billed monthly; 5% processing fee on Kilo Pass and top-ups; provider rates with no markup): kilo.ai/pricing and kilo.ai/pricing/kilo-pass
• Kilo Code free hosted-inference tier ended March 23, 2026 (Apache 2.0 BYOK extension still free), Kilo blog: blog.kilo.ai
• KiloClaw removal from the Kilo Code extension and CLI (v7.7.3, September 17, 2026; the PR describes the feature as being retired): github.com/Kilo-Org/kilocode
• GitHub Copilot interaction-data training policy (effective April 24, 2026), GitHub blog: github.blog
• Bodega One Code pricing (free to use, commercial work included): bodegaone.ai/pricing

Anthropic API pricing verified April 2026 from platform.claude.com/docs/en/about-claude/pricing. Kilo Pass tiers, bonus structure, and the 5% processing fee re-verified October 13, 2026 from kilo.ai/pricing. Anthropic third-party restriction date and scope verified from The Register's February 20, 2026 reporting. The math prices Kilo credits at Anthropic list rates, per Kilo's no-markup statement, plus the stated bonus structure and fee; top-ups are assumed to be bought in the exact amount needed. Run your own numbers against your own usage before switching either way.