[elitecoder]

Moves the balanced model tier from Sonnet 4.6 to Sonnet 5.5 for all three providers (Bedrock, the Anthropic API, and Vertex).

These jobs use the balanced tier, and none of them sets a model override on this machine:

• The warm Slack session (bin/comms_session.py)
• The Observer (bin/pulse.py)
• The brief narrator (bin/narrate-brief.py)
• The strategist (bin/strategist.py)

Spawned workspaces don't change. They use the claude alias.

Rollout

comms-listen.py loads the tier table at startup, so restart com.assistant.assistant-comms after merging. The restarted daemon replaces a warm session on the old model before its next Slack message (warm_session_model_is_current). The Observer, narrator, and strategist pick up the change on their next run.

Also in this PR

• The Observer and fleet evals default to the balanced tier's model instead of a hardcoded Sonnet 4.6 id. Their qualification hashes include model_tiers.py.
• bin/metering.py prices Sonnet 5 models at $2 and $10 per million input and output tokens (Anthropic pricing). It used to price every Sonnet at $3 and $15.
• docs/assistant-comms-onboarding.md names the new default.

Validation

• The full suite passes on Python 3.13, and CI is green.
• claude -p --model 'claude-sonnet-5-5[1m]' answers on the Anthropic API backend this machine uses. A Bedrock call on us.anthropic.claude-sonnet-5-5[1m] answered on 2026-10-03; its message id starts with msg_bdrk_.
• Observer eval on claude-sonnet-5-5[1m]: 13 of 14 on the first run, then 14 of 14 three times.
  • The one miss was 03-refactor-pr-ci-green, where Sonnet 5.5 returned ready_for_merge; the fixture forbids that verdict. Run alone, that fixture returned it 1 of 17 times on 5.5 and 0 of 12 on 4.6.
  • merge-pr-dispatch.py squirrel_e2e_gate still refuses a Squirrel PR without an E2E attestation and asks the user instead.

🤖 Generated with Claude Code