Last updated: November 21, 2026
SID-2 is an agentic search model from SID AI, announced November 7, 2026. Instead of writing answers, it runs your keyword and vector search tools, reads results, adapts its queries and returns ranked document IDs. It costs $0.25 per million input tokens, has a 256K context, and works through the OpenAI Responses API.
Contents
- What is new in SID-2
- SID-2 vs. SID-2 Swarm vs. SID-1
- Pricing and rate limits
- How to make your first SID-2 request
- Running the agent loop
- Using SID-2 Swarm
- Workflows and prompt patterns
- Limitations
- Frequently asked questions
What is new in SID-2
SID AI builds models that do only one task: search. Its first model, SID-1, was trained with reinforcement learning to search iteratively and report a ranked list of relevant documents. The SID-2 technical report (November 7, 2026) introduces two second-generation models, SID-2 and SID-2 Swarm.
The main changes over SID-1, according to the report and the SID models page:
- Longer context: 256K tokens (the report says 262K), up from 32K on SID-1.
- Reasoning effort levels: low, medium and high, with an extremely low effort setting for latency-sensitive uses.
- Metric Steering: you choose whether the model optimizes for NDCG, F1 or recall@N, and it changes how broadly it searches and reports.
- Agent swarms: SID-2 Swarm runs a team of agents that fork, search in parallel and share findings.
SID reports that SID-2 at medium effort reaches 0.58 NDCG in 3.1 seconds of model inference on its benchmark of difficult search questions, matching Claude Opus 5.5 at medium effort in 25 seconds. With six agents, SID-2 Swarm reaches 0.70 NDCG in 18 seconds.
SID-2 vs. SID-2 Swarm vs. SID-1
Figures below are from the SID models page.
| Feature | SID-2 | SID-2 Swarm | SID-1 |
|---|---|---|---|
| Retrieval quality (SID rating) | Better | Best | Good |
| Easy question min. latency | 0.5 s | 9 s | 1.5 s |
| Hard question min. latency | 1.9 s | 13 s | 4 s |
| Reasoning levels | Low, medium, high | High (fixed) | Default |
| Context window | 256K | Effectively infinite | 32K |
| Objectives | NDCG, F1, recall@N | NDCG, F1, recall@N | NDCG |
| Input / cached / output per 1M tokens | $0.25 / $0.025 / $1.00 | $0.50 / $0.05 / $2.00 | $0.75 / $0.075 / $4.50 |
| Suggested uses | Knowledge bases, support, e-commerce | Legal, finance, healthcare, compliance | SID says use SID-2 |
If you currently use a reranker in a fixed pipeline, compare this approach with our Cohere Rerank 5 guide.
Pricing and rate limits
SID-2 is billed per token: $0.25 per million input tokens, $0.025 per million cached input tokens and $1.00 per million output tokens. Each turn resends the conversation, so the request documentation says to send a new session ID per episode in the x-session-affinity header to raise cache hits.
On SID’s API, new workspaces start at 100 requests per minute per model, with no token limit. Burst capacity is one sixtieth of the per-minute limit and at least 5. A 429 response includes retry-after. For SID-2 Swarm, the swarm usage page lists defaults of 12 concurrent agents, 10 swarms started per minute and 8 open connections per API key.
How to make your first SID-2 request
SID-2 is reached through the OpenAI Responses API on SID’s own API, on Baseten, or as an AWS Marketplace package on an Amazon SageMaker endpoint. The steps below use SID’s API, following the SID-2 documentation.
- Create an API key on the API keys page of the SID dashboard. Keys look like
sid-sk-.... - Install the SDK and OpenAI client:
pip install sid-sdk openai(Python 3.12+). TypeScript uses@sid-ai/sid-sdkon Node.js 22+, and Go usesgithub.com/sidhq/sid-go. - Point the client at SID: base URL
https://api.sid-1.com/v1, modelsid-2. On Baseten, usehttps://inference.baseten.co/v1and modelsid/sid-2-preview. - Define your tools: a keyword tool named
text_search, a vector tool namedsearch, a read tool, and the fixedreport_helpful_idstool. - Build one request object with
store=False,include=["reasoning.encrypted_content"],max_output_tokens=4096,reasoning={"effort": "medium"}andextra_body={"sid": {"tool_examples": ..., "reward": "ndcg"}}. - Send the question with
client.responses.create(input=history, **request), run the returned tool calls, and repeat until the model reports IDs.
Tool call arguments arrive as a JSON string, so parse them first.
Running the agent loop
You own the loop. After each turn, append the model’s output items unchanged (including encrypted reasoning), one function_call_output per tool call in the order the calls were made, and a user message such as “You have 7 turns remaining out of 10 total.” The agent loop page lists the reference defaults: MAX_TURNS = 10, 5 results per search, 50-word snippets and up to 8 parallel fetches.
The episode ends when the model calls report_helpful_ids, when the turn cap is reached, or when a response is incomplete. Then remove duplicates and translate the model-facing IDs back to your database IDs with the SDK’s DocumentCache.
Using SID-2 Swarm
SID-2 Swarm runs over the WebSocket form of the Responses API at wss://api.sid-1.com/v1/responses with the header OpenAI-Beta: responses_multi_agent=v1 and model SID-2-Swarm. It is only on SID’s first-party API. The swarm request page lists the key settings:
multi_agent.max_agents: 1 to 24, default 6, counting the root agent.agent_max_turns: per-agent turn budget, default 50. Reasoning is fixed at high.- Do not define tools named
fork,message_agents,waitorlist_agents; SID adds them.
Each agent reports its own ranking, and SID recommends fusing them with Reciprocal Rank Fusion. SID says swarms help most on multi-hop and comparison questions.
Workflows and prompt patterns
- Match the reward to the job: use
"f1"for precise results such as support answers or job matching,"ndcg"for broad research such as legal or compliance review, and"recall@10"when you need exactly N documents (N can be 1, 2, 5, 10, 20 or 50). - Give captioned tool examples: SID asks for at least one example per tool, such as a long descriptive query for vector search and short keywords for text search.
- Start at low or medium effort for knowledge-base lookups and raise it only for hard multi-part questions.
For other ways to add retrieval to agents, see our Amazon OpenSearch Memory guide, our Perplexity Context API guide for web retrieval, and our DeepSeek Search API guide.
Limitations
- It does not answer questions. SID-2 returns ranked IDs; you still need another model to write the answer.
- You build the tools and the loop. Quality depends on your search backend, snippet formatting and tool examples.
- Benchmarks are self-reported by SID, and the Baseten and SageMaker model ID is labeled preview.
- Swarm is first-party only and has a default quota of 12 concurrent agents.
Frequently asked questions
What is SID-2?
SID-2 is a search model from SID AI, announced on November 7, 2026. It does one job: it calls the search and read tools you provide, reads the results, searches again if needed, and returns a ranked list of document IDs. It does not write the final answer.
How much does SID-2 cost?
SID’s models page lists SID-2 at $0.25 per million input tokens, $0.025 per million cached input tokens and $1.00 per million output tokens. SID-2 Swarm costs $0.50, $0.05 and $2.00, and the older SID-1 costs $0.75, $0.075 and $4.50.
What is the SID-2 model ID?
On SID’s own API at https://api.sid-1.com/v1 the model ID is sid-2. On Baseten and Amazon SageMaker it is sid/sid-2-preview. The swarm version uses the model ID SID-2-Swarm over a WebSocket connection, and only on SID’s first-party API.
Does SID-2 work with the OpenAI SDK?
Yes. SID-2 uses the OpenAI Responses API, so you can point the OpenAI client at SID’s base URL. It does not support Chat Completions or stateful serving, so you must set store to false and resend the whole conversation on every turn.
What is the difference between SID-2 and SID-2 Swarm?
SID-2 is one agent with low, medium or high reasoning effort and a 256K context. SID-2 Swarm runs a team of up to 24 agents that search in parallel and message each other. SID rates its quality highest, but it is slower and costs twice as much.
Sources
- SID AI: Models, specifications and pricing
- SID AI: SID-2 Technical Report: Scaling Search Swarms (November 7, 2026)
- SID platform docs: SID-2 overview
- SID platform docs: Making a request
- SID platform docs: SID-2 Swarm, Making a request
Image: “A magnifying glass” by PLBechly, via Wikimedia Commons, licensed CC BY-SA 4.0; resized.