AI Models & Cost Guide

Understanding how SPECTRA spends AI usage since the Claude Code migration (v2), and how to control it.

The cost model changed completely in v2

Every pre-v2 version of this guide described a GitHub Copilot premium-request (PR) multiplier table: SPECTRA ran its own AI SDK session, chose a provider/model from ai.providers, and billed against your Copilot plan’s monthly PR allowance.

That entire billing surface is gone. v2.4.0 deleted the GitHub Copilot SDK, ai.providers, ai.critic (as a config block), and spectra auth outright. SPECTRA no longer opens its own AI session and makes no model calls of its own for generation, analysis, criteria extraction, or updates. There is nothing left in spectra.config.json to route requests to a provider, because legacy configs that still carry ai.providers/ai.critic load cleanly, but the keys do nothing.

All inference is your interactive Claude Code session. SPECTRA’s CLI only does the deterministic bookkeeping either side of a model call:

spectra ai compile-prompt ...   →   you generate the answer, in-session   →   spectra ai ingest-tests ...

The middle step is an ordinary Claude Code turn (or, for grounding verification, a subagent call). There is no separate “SPECTRA AI cost” anymore, since it is exactly your normal Claude Code usage, and it is billed, throttled, and reported the same way every other piece of work in your session is: against your Claude subscription’s usage limits (Pro / Max / Team / Enterprise), or against your API/Console token balance if you run Claude Code with API billing. See Claude’s pricing page for current plan and per-token rates, since SPECTRA has no visibility into or influence over those numbers.

What actually makes a model call

Seam	Command sequence	What runs the model
Behavior analysis	`compile-analysis-prompt` → turn → `ingest-analysis`	Your session’s active model
Test generation	`compile-prompt` → turn → `ingest-tests`	Your session’s active model
From-description test	`compile-prompt --from-description` → turn → `ingest-tests`	Your session’s active model
Criteria extraction	`docs changed` → `compile-extraction-prompt` → turn → `ingest-criteria`	Your session’s active model
Targeted update	`compile-update-prompt` → turn → `ingest-update`	Your session’s active model
Grounding verification	`compile-critic-prompt` → `spectra-critic` subagent call → `ingest-verdict`	The subagent’s pinned model

The first five rows all run as ordinary turns in whatever model your Claude Code session is currently using, so pick it the same way you’d pick a model for any other piece of work (/model, or your plan’s default). The batch size for generation is ai.generation_batch_size (default 30 tests per turn); fewer, larger turns cost less overhead than many small ones, at the cost of a longer single response.

The critic is different: it always runs as a context: fork subagent (.claude/agents/spectra-critic.md in your project, seeded from SPECTRA’s bundled copy), invoked once per generated test. context: fork means each critic call starts from a clean, isolated context (just the test artifact + its source docs), so it doesn’t grow your main session’s context window, but each call still counts as normal Claude Code usage. The critic’s model is a static model: field in that agent file (shipped as claude-sonnet-4-6); there is no spectra.config.json knob for it anymore, so edit the agent file directly (and re-apply your edit after spectra update-skills if you’ve customized it).

Choosing a session model

Because generation and analysis literally run as your session’s own turns, “choosing a generation model” now means choosing what model your Claude Code session is running when you drive spectra-generate, spectra-criteria, or spectra-update.

Tier	Best for	Trade-off
Haiku	Straightforward CRUD docs, light re-verification, fast iteration	Shallower behavior analysis, meaning fewer implicit edge cases surfaced
Sonnet	Default for most suites, the balance of depth and speed	Good default; matches the critic’s own family
Opus	Complex business logic, deep implicit rules, large multi-doc suites	Slower and more expensive per turn; reserve for suites where Haiku/Sonnet analysis feels shallow

There is no dual-provider requirement anymore. Prior guidance to pick a different model family for the critic than the generator (for independent hallucination detection) no longer applies, because ARCHITECTURE-v2 §32 found a same-family critic (Sonnet verifying Sonnet-authored tests) reads as more useful in practice, since it shares the same standard of evidence. The critic stays pinned to the Sonnet family by default for that reason, independent of whatever model your own session uses.

Tuning levers that still exist

Everything provider/model-routing related is gone from spectra.config.json. What’s left under ai is purely about shape and pacing, not cost selection:

{
  "ai": {
    "generation_batch_size": 30,
    "generation_timeout_minutes": 5,
    "analysis_timeout_minutes": 2
  }
}

generation_batch_size (default 30) sets how many tests are requested per generation turn. Larger batches mean fewer round-trips through compile-prompt → turn → ingest-tests, at the cost of a bigger single response for the model to produce in one turn.
generation_timeout_minutes and analysis_timeout_minutes bound the deterministic CLI side of the seam (parsing, validation, index writes), not the model turn itself, since the turn now runs in your interactive session rather than an SDK call SPECTRA is waiting on.

Tracking what you’ve spent

SPECTRA can no longer report per-call token usage the way it did through v1, since it isn’t the process making the model calls anymore, so it has nothing to observe. .spectra-debug.log’s old RUN TOTAL … tokens_in=… tokens_out=… line and the Copilot-SDK-era AssistantUsageEvent capture were both retired along with the SDK.

To see what a generation/analysis/criteria/update run actually cost, use Claude Code’s own usage surfaces, whether your plan’s usage indicator or the API/Console token dashboard if you’re on pay-per-token billing, the same way you’d track the cost of any other work in the session.

Migrating from a pre-v2 config

If you’re upgrading from a release before the Claude Code migration:

Before (v1, GitHub Copilot SDK):

{
  "ai": {
    "providers": [
      { "name": "github-models", "model": "claude-sonnet-4.5", "enabled": true }
    ],
    "critic": {
      "enabled": true,
      "model": "claude-sonnet-4-6"
    }
  }
}

After (v2, Claude Code session):

{
  "ai": {
    "generation_batch_size": 30,
    "generation_timeout_minutes": 5,
    "analysis_timeout_minutes": 2
  }
}

You don’t need to edit anything by hand: a pre-v2 spectra.config.json loads unchanged (the providers/critic keys are simply ignored) and spectra init/spectra validate no longer ask about or warn on AI providers. spectra auth and gh auth login are gone; there is nothing to authenticate, since you’re already authenticated as whatever account is running your Claude Code session. See Claude Code v2 vs. the GitHub Copilot SDK v1 for the full picture of what changed beyond cost.