How much does T1 actually cost

"Is T1 cheaper than the alternatives" is one of the two questions we get every week. The honest answer is more interesting than a marketing yes.

T1 does not have a secret cheap-tokens moat. Per token we charge what the underlying providers charge, no markup. T1 Core is priced like Anthropic Haiku 4.5 because the tier IS Haiku 4.5. T1 Flow is priced like Anthropic Sonnet 5 because that is what is under the hood. T1 Max is priced like Opus 4-8 or Kimi K3, depending on what your prompt looks like.

So where does the cheaper-in-practice claim come from. Two places: the auto-router routes each prompt to the right tier instead of paying Max for a classify call, and running everything on one wallet instead of five means you never over-provision to avoid a rate-limit ceiling.

Below are three real workloads with the arithmetic spelled out. All rates are per one million tokens, current as of publish time.

Direct provider rates the T1 tiers map to:

Comparable models we will benchmark against:

Same app we walked through in the full-app blog post. Every ticket runs three calls: a classify, a draft response, an escalation check.

Typical shape per ticket:

• Classify: about 200 input tokens, 5 output tokens
• Draft: about 300 input tokens, 400 output tokens
• Escalation check: about 300 input tokens, 50 output tokens

The naive approach: everything on Opus 4.8. The reasoning is "use the best model so nothing gets fumbled."

Per ticket: (200 + 300 + 300) x $15/1M input + (5 + 400 + 50) x $75/1M output = $0.012 + $0.034 = $0.046 per ticket.

For 10,000 tickets: $460 a month.

The T1 Auto approach: router picks the tier. Classify goes to Core because it is a short, no-code prompt. Draft goes to Flow because Flow is the sensible default for prose that has to sound right. Escalate goes to Max because getting it wrong loses a customer.

Per ticket:

• Classify on Core: 200 x $1/1M + 5 x $5/1M = $0.000225
• Draft on Flow: 300 x $3/1M + 400 x $15/1M = $0.0069
• Escalate on Max: 300 x $5/1M + 50 x $25/1M = $0.00275

Total per ticket: $0.0099.

For 10,000 tickets: $99 a month.

Savings vs the naive approach: $361 a month, about 4.6 times cheaper. Same or better output. The escalate call is still on the deep-reasoning tier; only the cheap calls got demoted.

Internal dev tool. Mixed workload: about 60% short questions ("what does this error mean"), 30% coding prompts ("refactor this function"), 10% deep questions ("analyze the tradeoffs of using JWT here").

Assume output tokens run about half of input on average (models tend to reply shorter than the context).

Naive: everything on Sonnet 5.

Daily: 1M x $3/1M input + 500K x $15/1M output = $3 + $7.50 = $10.50 a day, $315 a month.

T1 Auto: router picks per prompt.

• 60% x 1M = 600K on Core: 600K x $1/1M + 300K x $5/1M = $0.60 + $1.50 = $2.10
• 30% x 1M = 300K on Flow: 300K x $3/1M + 150K x $15/1M = $0.90 + $2.25 = $3.15
• 10% x 1M = 100K on Max: 100K x $5/1M + 50K x $25/1M = $0.50 + $1.25 = $1.75

Daily total: $7 a day, $210 a month.

Savings: $105 a month, 33% cheaper. Not the crushing win of scenario 1 because most workloads sit in Flow territory anyway. But you get Max on the hard 10% for free (in the sense that the Core savings cover it), and you never accidentally use Max on the trivial 60%.

The bigger point: naive Sonnet is a real, defensible choice. T1 Auto is 33% cheaper AND catches the deep questions with the deep tier. If we ran the same numbers against naive Opus 4.8, T1 Auto would come in at about 6.5 times cheaper.

Analysis workload. Legal review, architecture design, research synthesis. Every prompt is genuinely hard; every answer needs to be right. Average shape: 8000 input tokens (the context you are pasting in) and 2000 output tokens (a substantial response).

Naive: everything on o1. The go-to when you want reasoning.

Per prompt: 8000 x $15/1M + 2000 x $60/1M = $0.12 + $0.12 = $0.24 per prompt.

Weekly: 100 x $0.24 = $24. Monthly: about $104.

T1 Max.

Per prompt: 8000 x $5/1M + 2000 x $25/1M = $0.04 + $0.05 = $0.09 per prompt.

Weekly: $9. Monthly: about $39.

Savings: $65 a month, about 2.7 times cheaper. And if you turn on T1 Max's extended thinking mode with a 4000-token budget, you pay for the reasoning tokens directly in output cost, still landing well under o1's per-prompt price.

Rough shape:

None of this comes from T1 charging less per token. It comes from routing. When you always use the biggest model, you pay biggest-model prices on every call, including the calls that would run cleanly on the cheapest tier. T1 Auto reads the prompt, picks the tier, and only spends the expensive tokens on the prompts that genuinely need them.

The number that matters for you is not the ratio in the table. It is your own workload. Plug the token shape of your top three calls into the arithmetic above; if the ratio looks like scenario 1 (a lot of short calls, a few deep ones), the savings are big. If it looks like scenario 2 (mostly one middle tier), the savings are real but smaller.

Being honest matters more than winning every comparison.

If your entire workload is short, boring classification and you would run it on GPT-4o-mini ($0.15 input, $0.60 output) or DeepSeek Chat ($0.27 input, $1.10 output), T1 Core at $1 / $5 is more expensive per token. If pure per-token cost on a Haiku-class workload is what you optimise for, direct routing to the cheapest provider still wins.

What T1 gives you at that price is: one wallet, one API key, one dashboard, one auto-router that catches the 5% of your traffic that actually needs Sonnet or Opus. If your model bill is $30 a month, the operational overhead of managing four provider accounts probably costs more than the price delta. Above that, the tradeoff shifts.

The other place T1 does not currently win: prompts that would benefit from Anthropic prompt caching. Prod today routes Anthropic-family traffic through OpenRouter, which strips the cache_control beta. Caching would cut cost on repeat context by another 90% for the read tokens. We are working on it; the SDK already exposes the accounting fields (see @taskclan/sdk 1.3.0).

T1 is priced at cost per token and cheaper in practice by 1.5 to 6 times on real mixed workloads because the router does not spend Max tokens on Core work. Plug your actual token shape into the arithmetic above and see where you land. If you want to try the numbers on your own traffic, grab a key at platform.taskclan.com and point the SDK at your workload for a week. The ledger tells you exactly what you spent per tier.