Changelog

Product updates, new endpoints, and behavior changes. Newest first.

Subscribe to the RSS feed or follow @realtellodb for announcements.

2026-09-30, ten new models

Added · models

• Hobby catalog: openai/gpt-6-luna, deepseek/deepseek-v4.1-flash, and google/gemini-3.8-flash.
• Paid catalog: openai/gpt-6.1-sol, anthropic/claude-sonnet-5.5, anthropic/claude-opus-5.5, x-ai/grok-4.7, xiaomi/mimo-v2.6-pro, meta/muse-spark-1.3, and cohere/command-a-plus.
• Provider context and completion limits, reasoning effort options, and mandatory-thinking settings for the new models.
• Updated pricing references and Xiaomi provider grouping.

All ten models are available through the shared catalog in debates, assemblies, Draft and Review, REST, and MCP. Pro and Pro Max retain the same paid model access. Gemini 3.8 Flash estimates use standard $1.50/M input and $7.50/M output rates, excluding the temporary 50% promotion. Existing saved panels and workspace defaults are unchanged.

2026-09-05, GPT-6 Astra

Added · models

• openai/gpt-6-astra with computer use and long-horizon reasoning, supporting a 1,050,000-token context window.
• Frontier model access for Pro and Pro Max subscribers, also accessible via prepaid compute credits.
• Standard OpenRouter metering at $5.00/M input tokens and $25.00/M output tokens.

You can use GPT-6 Astra in debates, assemblies, Draft and Review workspaces, and the paid model catalog.

2026-08-14, Gemini 3.7 Flash

Added · models

• google/gemini-3.7-flash with multimodal input, reasoning and tool support, and a 1,048,576-token context window.
• Hobby access matching Gemini 3.6 Flash.
• Standard metering at $0.75/M input tokens and $3.75/M output tokens, excluding OpenRouter's temporary 50% promotion.

You can use Gemini 3.7 Flash in debates, assemblies, Draft and Review workspaces, and the free model catalog. Gemini 3.6 Flash remains selectable because it is not a default.

2026-08-13, DeepSeek V4 Pro 0813 and Grok 4.6

Added · models

• deepseek/deepseek-v4-pro-0813 with a 1,048,576-token context window.
• x-ai/grok-4.6 with text, image, and file input and a 500,000-token context window.

Both models are available in debates, assemblies, Draft and Review workspaces, and the paid model catalog. Grok 4.6 provider pricing doubles when a prompt exceeds 200,000 tokens. Pre-run estimates use the standard rate and can understate those long prompts; final metering uses provider-reported usage.

2026-06-28, Assembly-first v1 API

Added · v1 API

The v1 API is now assembly-first. Two-step flow:

• POST /v1/assemblies, create a panel (models, format, threshold).
• POST /v1/debates, start a debate by passing prompt + assemblyId only.

Also new:

• GET /v1/assemblies, list your assemblies.
• GET /v1/assemblies/:id, fetch one assembly.
• GET /v1/debates, list your debates (paginated, filterable by status and assemblyId).
• GET /v1/debates/:id/status, lightweight status for polling.

Breaking · v1 API

POST /v1/debates no longer accepts inline models, maxRounds, agreementLimit, mode, format, or formatMeta. All panel config now comes from the assembly. The Idempotency-Key header is now required.

If you have existing integrations, migrate by:

1. Create an assembly once for each panel config you use.
2. Save the assemblyId.
3. Send { prompt, assemblyId } in subsequent POST /v1/debates calls.

2026-06-22, Production stability

Changed · orchestrator

• queue infrastructure retry budget increased from 1 to 3 (transient model failures now retry within the job before declaring FAILED).
• Stuck-session watchdog: any RUNNING session silent for more than 10 minutes is auto-failed.
• 5xx from model providers now triggers 3 model-call retries with exponential backoff (was 1).

2026-06-15, Documents and RAG

Added · web app

Upload PDF, DOCX, code, and text files to a debate. The system extracts text, chunks it, embeds it, and retrieves relevant chunks for each model's turn. The chairman's verdict cites pages and sections.

Limits: 50 MB per document, 1.5M tokens total context per debate.

2026-06-08, Webhooks for v1

Added · v1 API

Pass webhookUrl and a unique webhookSecret when starting a debate and the platform will POST the full session to that URL on completion (success or failure). Payload matches the GET /v1/debates/:id response. Deliveries are signed with HMAC-SHA256.

2026-06-01, Reasoning tokens

Added · models

Per-model enableThinking: true now supported across all major providers. When enabled, models emit reasoning tokens before their final response. Reasoning tokens are counted separately in the per-turn cost breakdown. Roughly 2-5x per-turn cost, substantially better answers on hard reasoning tasks.

2026-05-25, API key rotation

Added · web app

Generate multiple API keys per account. Useful for: separating production from staging, rotating without downtime, scoping access per application. Each key has independent usage tracking. Revocation is immediate.

2026-05-18, Prepaid credits

Added · billing

For usage beyond your monthly tier allowance. Credits don't expire. Use the Customer Portal to purchase.

2026-05-11, Oxford Union format

Added · formats

Formal Proposition vs Opposition debate with 4 rounds (opening, rebuttal, closing). Models split into sides via formatMeta.sides. The chairman declares a winner.

2026-05-04, Refinement and Devil's Advocate

Added · formats

The first two debate formats. Refinement is collaborative; Devil's Advocate is adversarial. Both are panel-based, with mode: "REFINEMENT" or mode: "DEVILS_ADVOCATE".

2026-04-27, v1 API launch

Added · v1 API

Initial public API. POST /v1/debates and GET /v1/debates/:id. Bearer-token auth. model providers model IDs accepted directly. Idempotency keys supported from day one.

Older entries are in the git history. For platform-wide announcements, follow @realtellodb.