SID AI has launched SID-2, a second-generation agentic search model, alongside its Swarm mode. The system replaces rigid retrieval pipelines with a model that queries data, reads results, and decides the next search step. It returns ranked document IDs rather than generating text directly.
In this article
Technical specifications and performance
The model operates with a 262,000-token context window in research reports, though the live site lists 256,000 tokens. Developers can access it via a hosted API at platform.sid.ai, or through reference paths on Baseten and SageMaker. Latency figures assume a single B200 GPU handling eight concurrent queries.
Performance metrics show a single agent matching Claude Opus 5.5 on retrieval quality. The Swarm mode, which runs multiple agents, outperformed every model tested by SID. The best result came from SID-2 Swarm with 12 agents, scoring 0.720 NDCG. This beats GPT-6 Astra, which scored 0.704. The lowest tier, SID-2 Low, scored 0.556 NDCG, placing it 0.148 below GPT-6 Astra. Astra still holds the lead on California law, SEC filings, and FiQA tasks.
Operational mechanics
SID-2 is a search specialist, not a general chat model. Static pipelines previously ran rewrite, embed, search, and rerank steps in a fixed order. SID-2 searches, reads results, adapts, and searches again.
Developers connect search tools to their own backends. A text_search tool handles keyword lookups while a search tool covers vector queries. The model finishes each episode by calling report_helpful_ids with a ranked list. No reindexing or data movement is required because the model only sees what the tools return.
SID-2 Swarm runs several agents that message each other and fork their context into new agents. Each agent explores a separate slice of the corpus. The company argues search gains more from reading more documents than from reasoning harder about each one.
Reasoning effort and metric steering
The model offers low, medium, and high reasoning effort levels. Each level was trained as a separate expert, then merged using multi-teacher on-policy distillation. Metric Steering lets users target NDCG, F1, or Recall@k, with k ranging from 1 to 50. On SID’s 506-question validation set, SID-2 at high effort scored 0.14 to 0.22 higher per metric than the base model.
Training methodology
Training used a two-stage reinforcement learning process. Stage 1 optimises retrieval quality only. Stage 2 adds a multiplicative latency penalty that applies only to positive rewards. A proposer-solver self-play loop generated hard queries against a frozen SID-1 solver. SID reports a sixfold training-throughput gain over SLIME’s default setup.
Benchmark results
SID built a 506-question benchmark across 19 corpora. These include HotpotQA, Enron emails, SEC filings, PubMed, and BrowseComp-Plus. All numbers below are self-reported from Tables 2 to 4 of the report.
- SID-2 Medium achieved 0.585 NDCG in 3.11 seconds, versus Claude Opus 5.5 (medium) at 0.577 in 25.1 seconds.
- SID-2 Swarm with 6 agents reached 0.701 NDCG in 17.7 seconds, versus GPT-6 Astra (max) at 0.704 in 135 seconds.
- SID-2 Swarm with 12 agents scored 0.720 NDCG, the highest score in the table.
- Baselines include BM25 at 0.260 and vector search plus reranker at 0.287.
Swarm leads on 11 of 19 corpora. On WixQA, a single high-effort agent matches Astra at about 232 times lower cost. At its lowest effort, SID-2 beats four parallel SID-1 runs with 13 times fewer FLOPs.
Comparison with competitors
The following table compares SID-2 against GPT-6 Astra and Claude Opus 5.5.
| Feature | SID-2 (medium) | SID-2 Swarm (6 agents) | GPT-6 Astra (max) | Claude Opus 5.5 (medium) |
|---|---|---|---|---|
| Type | Agentic search model | Multi-agent search | General LLM | General LLM |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Context | 262k (report) / 256k (site) | Effectively infinite (per SID) | Not disclosed | 1M |
| NDCG, SID benchmark | 0.585 | 0.701 | 0.704 | 0.577 |
| Inference time per query | 3.11 s | 17.7 s | 135 s | 25.1 s |
| Cost per 1,000 queries | $6.54 | $263 | $2,141 | $131 |
| API price (input / output per 1M) | $0.25 / $1.00 | $0.50 / $2.00 | $10 / $50 | $4 / $20 |
NDCG, time and cost rows are measured by SID on its own benchmark. Retrieval backend time is excluded.
Deployment and pricing
SID-2 is available through SID’s API and on Baseten. SDKs exist for Python (sid-sdk), TypeScript (@sid-ai/sid-sdk) and Go (github.com/sidhq/sid-go). A SageMaker helper covers AWS deployments. The reference implementation uses the OpenAI client format.
SID-2 costs $0.25 per million input tokens and $1.00 per million output tokens. Swarm costs $0.50 and $2.00. Cached input is 10% of the input price.
What it means
Developers gain a tool that handles complex retrieval without manual pipeline configuration. The Swarm mode allows scaling search depth by adding agents rather than increasing model complexity. Pricing remains accessible compared to general-purpose models, though the Swarm mode costs more per query. All results are self-reported; size and weights are undisclosed.


