Skip to content

[assemblyai] Verify regional pricing, cache rates, and decision models - #1020

Open
mayankjain-panw wants to merge 8 commits into
mainfrom
feat/assemblyai-provider
Open

mayankjain-panw wants to merge 8 commits into
mainfrom
feat/assemblyai-provider

Conversation

@mayankjain-panw

@mayankjain-panw mayankjain-panw commented Nov 5, 2026 •

Copy link
Copy Markdown
Collaborator

Final qualification - 2026-11-17

Merged current main without rewriting history. Both CI validation checks pass. Both review threads are resolved with product-specific pricing sources and end-to-end evidence.

  • Rechecked the official Sync ($0.30/hour) and Dictation ($0.62/hour) rate cards. Their nonzero additional-unit rates are intentional. The older zero-rate scope no longer applies to these synchronous products.
  • The real local stack recorded 0.5 cents for a mocked Sync minute and 62/60 cents for a mocked Dictation minute in both Rust and TypeScript paths.
  • Full-stack qualification passed 20/20 scenario groups per runtime, including actual persisted analytics/raw logs. Haiku's formerly missing rate, global pricing and default-model TTS also passed.
  • Luna displayed the complete 73-entry integration catalog. Generated Gateway and Albus bundles remain tied to this Models source.
  • No price or model-ID change was needed for the review comments. Async/WebSocket accounting and the approved inherited Jev assumption remain explicit limitations.

Qualification report. No real AssemblyAI credential or invoice was used.

Summary

Add the AssemblyAI catalog and pricing in the repository's existing cents per billable unit format. Current head: 29fb9ba062412726adb2f6ad1d6919f28b2874fc. Includes main at c12277ef98ffc7145ff557840f47597b6e6e7e98.

Models is the source of truth. Gateway and Albus bundle generated copies through their existing configuration mechanisms; they do not maintain independent rates.

Coverage and rates

  • All 60/60 models in the live LLM Gateway catalog now have general metadata and pricing. Previously only four were included.
  • All published global, us, eu and regional_increase_percent data is represented. Forty models have regional uplifts. Three Nemotron models have explicit US/EU rates. No unpublished cache prices, long-context tiers or other dimensions were invented.
  • Published cache reads, five-minute writes and one-hour writes are represented with existing schema fields.
  • Sync 3.5 and 3.6 use internal pricing keys universal-3-5-pro-sync and universal-3-6-pro-sync: 0.5 cents/minute = $0.30/hour. This replaces the earlier, now-stale assumption that Sync pricing was unpublished.
  • Dictation uses the internal key assemblyai-dictation: 62/60 cents/minute = $0.62/hour, including LLM cleanup. Duration is charged through additional_units.request_audio_token; token prices are zero to avoid double charging.
  • Native/default TTS and the six language IDs retain 0.0015 cents/character = $15/million. Provider-reported billed characters take precedence in Gateway when available.
  • Jev retains the previously approved TypeSafe/OpenRouter rate: $0.042/million input tokens, output free. AssemblyAI-specific Jev billing remains unconfirmed. This is an inherited assumption, not a newly verified vendor rate.
  • Audio and Decisions remain disabled in Playground. The 60 chat entries are enabled; unsupported unified streaming is not advertised.
  • Async transcription and WebSocket-session accounting remain explicitly unmetered. This PR does not invent a session/job accounting implementation or apply an async rate to synchronous traffic.

Sources checked 2026-11-16

Validation

  • All 60 live IDs covered; 375 exact-decimal rate assertions passed against the captured official catalog.
  • Existing pricing-schema validation, duplicate-key checks, focused ESLint/Prettier and whitespace checks passed. Normal commit hooks passed; CI JSON validation is green.
  • Gateway's bundled pricing and Albus's bundled general/pricing files are semantically identical to these source files.
  • Production TS/Rust calculator regressions verify actual bundled rates, including Haiku at $6.60 for 1M input + 1M output in-region, Sync at 0.5 cents/minute, Dictation at 62/60 cents/minute, and the three explicit Nemotron EU overrides.
  • Luna's actual model feature helper recognizes all 60 as chat models and does not advertise unsupported unified streaming.
  • Full-repository Models lint has 2,653 pre-existing formatting errors in general/scenario-ai.json, reproduced on exact main. That unrelated file is unchanged.

No authenticated AssemblyAI calls or invoice reconciliation were performed: no approved test credential was supplied.

Rollout and companions

Publish this data to the normal remote catalog/pricing store before enabling priced Gateway traffic. Existing negative pricing-cache entries can survive for an hour. The normal Gateway pricing-sync workflow is manually triggered and opens a PR/MR; it does not deploy automatically.

The companion Gateway MR includes the local rate-card bundle, TS index, Rust loader and registry coverage. Albus includes the offline model-config bundles and provider-row SQL migration. Existing deployments must apply that SQL through their normal deployment process.

mayankjain-panw and others added 3 commits November 5, 2026 16:39
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Use live AssemblyAI model rates. Apply the missing Gemini regional uplift and expose global rates through custom pricing regions. Include the documented Decisions models; leave the unavailable Jev price explicitly null.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@mayankjain-panw mayankjain-panw changed the title feat: add assemblyai model config and pricing [assemblyai] Verify regional pricing, cache rates, and decision models Nov 5, 2026
mayankjain-panw and others added 3 commits November 5, 2026 21:26
Use the same Jev 1.13 input rate as TypeSafe and OpenRouter: 0.0000042 cents per token, with zero output-token cost. This follows the requested cross-provider model-rate convention; AssemblyAI-specific Jev pricing remains unconfirmed. No cache discount or regional surcharge is inferred.

Validated actual model selection, usage extraction and cost calculation against the TypeSafe rate for 0, 1000 and 1000000 input tokens, including nonzero output usage. Pricing JSON formatting, lint, syntax and duplicate-key checks pass.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Update the AssemblyAI catalog and pricing PR with current main while preserving its provider entries.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Remove the DashScope-only formatting change introduced by the merge commit hook. AssemblyAI catalog and pricing entries are unchanged.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@mayankjain-panw
mayankjain-panw marked this pull request as ready for review November 16, 2026 06:52
Copilot AI balanced review requested due to automatic review settings November 16, 2026 06:52

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔵 Needs a closer look

Billing correctness depends on unavailable gateway behavior and an unverified inherited Jev rate for AssemblyAI, requiring human confirmation before rollout.

No open issues 🎉

What changed in this PR

Adds AssemblyAI model metadata and pricing to the repository’s catalog, which supplies rates through the existing pricing sync.

Changes:

  • Registers speech, text-to-speech, chat, and Decisions models with capabilities and output limits.
  • Adds per-unit pricing, regional/global overrides, and supported cache rates.
  • Includes the inherited Jev pricing assumption and zero placeholders for unpublished Sync pricing.
File Description
pricing/​assemblyai.json Adds AssemblyAI rates, regional overrides, cache pricing, and unpriced defaults.
general/​assemblyai.json Adds model IDs, capabilities, output limits, and playground restrictions.

Review effort: Balanced

Cover all 60 public LLM Gateway models with regional and cache prices. Set Sync and Dictation duration billing without token double charging.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot AI balanced review requested due to automatic review settings November 16, 2026 10:54

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Changes recommended

Additional-unit rates remain nonzero for Direct Sync and Dictation despite the stated zero-pricing policy, without verified product-specific billing sources.

2 open issues

Review effort: Balanced

Comment thread pricing/assemblyai.json Resolved
Comment thread pricing/assemblyai.json Resolved
Keep the reviewed AssemblyAI model catalog and published synchronous audio rates unchanged while incorporating current main.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot AI balanced review requested due to automatic review settings November 17, 2026 07:12

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔵 Needs a closer look

Billing correctness requires human cross-repository validation of regional pricing and metering units against companion Gateway/Albus consumers unavailable in this checkout before rollout.

2 open issues

Review effort: Balanced

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants