# Label thousands of items cheaply

> Label thousands of items cheaply: 10 companies fully do it, 12 with a limit, 30 carry the tag. Every listing in Decision models with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/decision-models.classify-at-volume/
- Markdown: https://www.anchorterminal.com/niches/decision-models.classify-at-volume/index.md (~2,200 tokens)
- Slim: https://www.anchorterminal.com/niches/decision-models.classify-at-volume/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/decision-models.classify-at-volume/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 10 | 12 | 30 | 15 | 7 |

## What counts as done

Every item labelled with a probability per option at under $0.10 per 1,000 items.

Measured, for our own tests: median latency under 300 ms per call, with no parsing failures.

How agents ask for it: categorise these 10,000 products; tag every email in this inbox; flag which of these charts shows an anomaly; score sentiment and spam on every comment

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Celeris-1 Decision (Marqo Inc)**: $0.004 per million input tokens, output free, 512 questions per request, published default of 1,500 requests and 10 million input tokens a minute. Images count as input. No batch endpoint, so the caller runs its own pool.
- **Decider (Mark Marosi (Mapika))**: Free local weights from 0.8B to 35B; the 1.1.0 server batches across requests at a 50 ms median, MPS 133 ms, GGUF on 8 CPU threads 0.12 to 0.7 s. Up to 1,024 scoring rows a request.
- **Decision 3.0 (vLLM Semantic Router project)**: Free Apache-2.0 weights in six sizes; d3-mini answers 16 text questions in 109 ms and d3 in 548 ms, each question its own forward pass over the shared input. The authors' free API logs requests and may train on them.
- **GLiClass (Knowledgator Engineering Ltd.)**: Free local pipeline with batched inference, a Ray Serve endpoint with dynamic batching, ONNX edge models and a CPU flag; the v3 card gives examples a second on an A6000 by label count. The HTTP route takes one text a request.
- **mammoth-1 (Baseten)**: Baseten serves it at $0.10 per million input tokens, about $0.03 per 1,000 short items; a 1,342-question community benchmark measured 240 ms median and 84.9 percent, four answers behind GPT-6 Sol.
- **Drex 1.5 (Nace AI, Inc.)**: $0.03 per million input tokens on drex-max-v1.1, up to 512 questions a request with the state counted once in Max's whole-request budget, and published free and paid requests-per-minute tiers.
- **OpenAI Decisions API (OpenAI OpCo, LLC)**: $0.10 per million uncached input tokens, about $0.03 per 1,000 short items, 1 to 400 questions a request, prompt caching on shared input, 0.16 s p50 at Arize. Not in the Batch API.
- **pplx-decider (Perplexity AI, Inc.)**: $0.02 per million input tokens with the state billed once for grouped questions on the 27B; 10 requests a second default, raisable on request. The card measured 96 ms median a request on the fast path.
- **kettei (Sqwish Ltd, trading as Sqwish Labs)**: kettei-core at $0.02 per million input tokens with the context paid once for up to 128 decisions; the comment-scoring example runs three questions a comment. Rate limits are per tier and read from the account.
- **Jev (TypeSafe AI, Inc.)**: $0.042 per million input tokens with the state billed once for any number of questions within 64k; a user categorised over 10,000 items in minutes for under a cent, a 1,000-email test ran four questions at about 200 ms a message, Amplitude paid about $0.06 per 1,000 charts.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Clef (Cloudflare, Inc.)**: Clef-flash at $0.038 per million input tokens and 64 questions a call meets the price. Throughput is bounded by the Workers AI default of 300 requests a minute, and Clef-flash's hosted context fell to 24,576 tokens on 9 October.
- **Laya (Convai Innovations Pvt. Ltd.)**: Free and local at 33 ms a decision on a T4 or 193 to 464 ms on CPU, all questions in one pass. Arize measured 59.8 percent on RAGTruth with calibration error 0.330, and the context is 512 or 1,024 tokens, so accuracy and not cost is the limit.
- **Drex Max v1 (Nace AI, Inc.)**: Free to download, but the 27B needs about 49 GB of GPU memory in BF16, 26 GB in 8-bit or a 24 GB Mac for the affine Q4 text build; the licence cap applies. No throughput figures on the card.
- **elephant-1 (Baseten Labs, Inc.)**: Baseten serves it at $0.035 per million input tokens with zero data retention, well under the price bar, but no latency figure is published for the 9B and local serving needs the patched runtime.
- **Granite Switch 4.1 (International Business Machines Corporation)**: Free to run on your own vLLM with prefix caching across decisions, but every question is its own chat request with hand-built prompts, and no latency figures are published for the adapter.
- **Kev (Jared Palmer)**: Free local weights at 18 to 41 ms for six questions and about 101 requests a second on an H100; a user reported Kev-4B outperforming other models on entity resolution. Issue 281, open, reports VRAM growing with each invocation.
- **JEV-27B-VL (AutoTrust AI Lab)**: One question per request, so several labels per item multiply calls; about 67 GB of weights with no quantised build. Free to run.
- **joey-1 (Baseten Labs, Inc.)**: Baseten's $0.01 per million input tokens is the cheapest hosted price in the chunk, but the index's Full score is 35.91 with 10.2 on new domains, so labels outside its training mix need checking first.
- **Liquid d1 (Liquid AI, Inc.)**: Hosted billing charges every question for its text and all images again, so three questions an item cost three states; still about $0.03 per 1,000 short items at 0.20 s p50. d1-2B is free locally.
- **Microsoft-Decision-1 (Microsoft)**: $0.042 per million input tokens meets the price bar, but OpenRouter's measured latency is 0.42 s a request, over the 300 ms median, and Foundry needs a deployment set up by a person.
- **pplx-decider-flash (Perplexity AI, Inc.)**: Flash bills the state again for every question in a request, so from three questions an item the 27B costs less; one question an item at $0.01 per million is the cheapest hosted rate in the chunk.
- **ScaleDown (Carbon ScaleDown, Inc.)**: Every question in a request is a separate model call with usage summed, so the state is billed per question (about $0.045 per 1,000 items with three short questions at $0.05 per million). The chunks batch on /classify bills the shared profile once.
- **SPEV (Sprag AI, Inc.)**: Price is fine, but the default limit of 60 requests a minute makes 100,000 items a 28-hour job unless raised; questions about one piece of evidence go in one request.
- **Strands Decider 2B (Amazon Web Services, Inc.)**: Free and 115 ms a question on a GPU, but on a 2 vCPU microVM a decision took 11 to 13 s and cost more than a Nova Pro answer; batch requests are an open feature request.
- **Vela 2.0 (vLLM Semantic Router project (with KR Labs))**: Free local weights with multiple named questions a request and a 0.3B that serves on CPU, but general decision accuracy is weak (index 31.63) and no per-decision latency for the choice path is on the card.

- not this job: RightNow System Zero (RightNow AI, Inc.): USD 0.00013 per answered question plus USD 0.000055 per answered call on top of USD 0.19 per million text tokens: one question on 1,000 items costs about USD 0.19 before tokens, over the USD 0.10 per 1,000 bar.

1 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- per-question billing of the state (pplx-decider-flash) multiplies cost from three questions up
- no batch endpoint, so the caller runs its own worker pool
- rate limits unpublished and set per organisation

## Around it

Before: web-scrapers.page-to-markdown, document-extraction.table-from-text, email.list-messages. After: databases.write-rows, spreadsheets.append, inference.llm.deep-dive-on-flagged.

## Where this comes from

Weight 15 of 100 in Decision models: 7 of 41 accounts; two HN commenters, the Amplitude chart benchmark, the email test and the Kev entity-resolution issue

As data: https://www.anchorterminal.com/api/v1/use-cases/decision-models.classify-at-volume.json, or the MCP tool get_use_case with id decision-models.classify-at-volume. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
