Autonomous coding tiers currently run $200–500 per seat per month: Claude Code’s cloud tier at $200/mo, Devin Teams around $500/mo, and GitHub Copilot’s coding agent metering premium requests on top of a Copilot plan. Bundled into those prices are two very different things: orchestration (the product) and inference (the tokens). We think you should see them separately, because the token math is more interesting than the subscription prices suggest.

What a coding job actually costs in tokens

At July 2026 list prices, per million tokens:

ModelInputOutput
Claude Sonnet$3$15
Claude Opus$5$25
Claude Haiku$1$5

Now price a realistic job. A dependency bump or focused bug fix on Sonnet typically consumes something like 400K input tokens (the agent re-reads files, runs tools, accumulates context) and 30K output tokens:

  • Input: 0.4 MTok × $3 = $1.20
  • Output: 0.03 MTok × $15 = $0.45
  • Total: ≈ $1.65

A small lint-fix job on Haiku — say 100K in, 10K out — is $0.15. A gnarly refactor on Opus with a fat context — 1.5 MTok in, 80K out — is $9.50. Prompt caching (which agent harnesses lean on heavily, since they re-send the same repository context turn after turn) cuts the input side further; cached reads are billed at roughly a tenth of the base input rate.

So a team member running 30 typical Sonnet jobs a month spends about $50 in tokens. Run a hundred small Haiku cleanups instead and it’s $15. The tokens for a normal month of agent-assisted maintenance cost less than a nice lunch — yet the bundled seat for that same work is priced at $200–500.

Where the rest of the money goes

The gap isn’t fraud — it’s bundling. A managed vendor is covering sandbox VMs, idle capacity, egress, support, and margin, and smoothing heavy users against light ones. But bundling has two costs for you:

  1. You can’t see the markup. When inference is resold inside a flat seat, you can’t tell whether you’re subsidizing someone else’s usage or being throttled to protect the vendor’s margin. Rate limits become a product knob instead of a bill you control.
  2. You can’t arbitrage models. The single biggest cost lever in agent work is model choice per task — Opus where it matters, Haiku where it doesn’t, a local model where data can’t leave. A bundled seat locks that lever at the vendor’s setting.

The BYOK alternative

Nominos bills the two layers separately. The orchestration layer — dashboard, CRD controllers, merge queue, checkpoints, audit trail — is the subscription (Free to start, Pro at $39/user/mo on the Standard track). Inference is your own API key, billed by your provider at raw prices with zero markup, running on your own cluster.

Concretely, a five-person team doing serious agent work might look like:

  • Nominos Pro: 5 × $39 = $195/mo
  • Tokens at raw prices (150 mixed jobs): ≈ $75–150/mo
  • Total: ≈ $270–345/mo for the whole team

Versus $1,000–2,500/mo for the same five seats on bundled autonomous tiers — with less model flexibility and your code on someone else’s infrastructure.

BYOK isn’t automatically cheaper for everyone: if your usage is enormous and bursty, a vendor eating your inference bill can win, and you do take on cluster operations (though if you’re our customer, you already run Kubernetes). The point isn’t that bundles are always a bad deal — it’s that you should be able to check. When the token line item is your own bill, the markup conversation disappears, and cost optimization becomes an engineering decision you own: pick the model per job, cache aggressively, cap spend with maxCostDollars on the resource itself.

Transparency is a feature. Your cluster, your keys, your bill.

Nominos is in early access — see pricing or request access.