What "verified" means here

• 14 pages read in full on 5 Oct 2026: intro, async, batch, models, prompt caching, reasoning, Anthropic compatibility, skill, autobatcher, one model page, llms.txt, the GitHub org page, the skill repo, and the marketing homepage.
• About 12 more seen as search snippets (Pi, Rig, Daytona, Agno, Mastra, CrewAI, Google ADK, Control Layer pages, skill-page variants).
• ~140 pages listed in the site's own llms.txt (my count). Most are indexed here by title and description only.
• The sidebar of the live docs was used to check every docs link in this deck.

What changed from v5

• Full price table rebuilt from the page; v5 had ~half of it from memory.
• "Async up to 50% off / batch up to 75% off" removed. Computed from the table instead: ≈25% / ≈50% below realtime.
• "Control Layer removed from docs" was wrong; see the ledger in Section 09.
• Skill repo name, reasoning levels, Anthropic setup and Autobatcher details corrected.
• Restored from v4: four diagrams and the docs-review panels.

Tier discounts (list price, 31 models with realtime)

• Async ≈ 25% below realtime, tightly clustered (24–25% for the middle half).
• Batch ≈ 50% below realtime, and ≈33% below async.
• So the large savings in the marketing ("up to 90%") come from model choice (open-weight vs closed), not from tier choice alone. Output price spans 141× across the catalog ($0.10 to $11.25 per 1M async output).

Cache read is not always a discount

• 29 models have a cache-read discount: median 80% below input (range 27–97%).
• 7 models price cache read at the input price: Hy3, Qwen3.6 35B A3B, Gemma 4 31B IT, the three OCR models and Qwen3 Embedding 8B. The docs don't say whether caching is unsupported there or simply undiscounted.

Anomalies worth checking before you budget

• Qwen3.5 9B dottxt: async output $0.70 and batch $0.58 are higher than realtime output $0.30. Could be a typo or real; confirm in the Console.
• 5 models have no realtime price: Qwen3.5 4B, Qwen3.5 397B dottxt, DeepSeek OCR 2, LightOnOCR 2, olmOCR 2.
• Embedding output is $0.00; you pay input only.
• Reasoning table covers 19 of 36. Not listed: Kimi K3, Inkling, Hy3, MiMo V2.5 Pro, DeepSeek V4 Pro 0813 and V4.1 Flash, Qwen3.8 27B, Muse Glimmer, Gemma 4 26B and others. The DeepSeek V4 Flash page lists efforts, so the table may simply lag the catalog.

• Implicit: on by default for cache-supported models, no code change. Best-effort; reused tokens bill at the cache-read rate; no write charges.
• Explicit: add cache_control: {"type":"ephemeral","ttl":"5m"|"1h"} to a content block or tool. Guaranteed reads, sliding TTL, up to 4 breakpoints, shared across an org's keys. Minimum prefix is model-specific (1,024 tokens on most).
• Write prices: 1.25× input (5m), 2× input (1h). Cache read = r × input, model-specific.
• Models without caching ignore markers and bill normally, so markers are safe to send.

"messages": [{"role":"system","content":[{
  "type":"text","text":"<~2,000 tokens>",
  "cache_control":{"type":"ephemeral","ttl":"1h"}}]},
 {"role":"user","content":"How do I reset my password?"}]
# check usage.cache_read_input_tokens and
# usage.prompt_tokens_details.cached_tokens

• Seven levels: none, minimal, low, medium, high, xhigh, max (OpenAI-compatible). v5 showed three.
• Chat Completions: reasoning_effort. Responses: reasoning.effort.
• Unsupported effort is rejected, not silently changed. Omit the field for the model default.
• Some models map effort to a reasoning-token budget; your output limit must exceed it or the request is rejected.
• Send only these OpenAI-style fields. Do not send thinking, thinking_token_budget or chat_template_kwargs.
• Qwen3 non-thinking: append /no_think to the latest user prompt; it is Qwen-specific, not an effort value.
• Anthropic thinking config is mapped to reasoning effort (Anthropic page).

curl https://api.doubleword.ai/v1/chat/completions \
  -H "Authorization: Bearer $DOUBLEWORD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V4-Flash",
    "messages": [{"role":"user",
      "content":"Solve this carefully."}],
    "reasoning_effort": "medium",
    "max_completion_tokens": 16384
  }'

from anthropic import Anthropic
client = Anthropic(
    base_url="https://api.doubleword.ai",  # no /v1
    auth_token=KEY)       # NOT api_key=
m = client.messages.create(
    model="deepseek-ai/DeepSeek-V4-Flash",
    max_tokens=1024,
    messages=[{"role":"user","content":"Hi"}],
    extra_body={"service_tier":"flex"})  # optional

api_key= makes the SDK send x-api-key; Doubleword authenticates on Authorization: Bearer.

• Works: messages, streaming, GET /v1/models in Anthropic shape (with anthropic-version), flex tier via service_tier: "flex" in extra_body, caching on /v1/messages.
• Batch: native /v1/messages/batches is not available. Use the OpenAI-format Batch API with "url": "/v1/messages" on every JSONL line and endpoint="/v1/messages"; results come back as Anthropic messages.
• Anthropic's auto / standard_only service tiers are accepted but do nothing.
• The page does not mention tool use or vision. v5 claimed tool use; that was not supported.

# pip install autobatcher   |   npm install autobatcher
from autobatcher import AsyncOpenAI      # or BatchOpenAI
client = AsyncOpenAI(
    api_key=KEY,
    base_url="https://api.doubleword.ai/v1")
r = await client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4-Flash",
    messages=[{"role":"user","content":"Hi"}])

• Supports chat.completions.create() and embeddings.create() only. No Responses API (v5 said otherwise); everything else passes through to the OpenAI client.
• No completion_window option in the current table (v3/v4 had one).

f = client.files.create(file=open("in.jsonl","rb"), purpose="batch")
b = client.batches.create(
    input_file_id=f.id,
    endpoint="/v1/chat/completions",
    completion_window="24h",
    metadata={"description": "daily eval job"})
# output_file_id and error_file_id exist immediately

r = requests.get(f".../v1/files/{b.output_file_id}/content",
                 headers=H)
if r.headers.get("X-Incomplete") == "true":
    resume = r.headers["X-Last-Line"]   # ?offset={resume}

• Partial results while running: download the output file early; X-Incomplete and X-Last-Line headers tell you where to resume with ?offset=. The page notes this differs from OpenAI.
• Cancel any time: unstarted requests are cancelled and not charged; in-flight cancellation is best-effort; finished results stay downloadable.
• The Console can create a batch by uploading a file; no separate trigger.
• url is typically /v1/chat/completions; /v1/messages is also accepted (Anthropic page).

• The site index lists 16 CLI pages under /dw-cli: introduction, installation, authentication, quickstart, batches, streaming, real-time, JSONL format, local file tools, projects, examples, command reference, accounts, API keys & webhooks, usage & analytics, global flags.
• The CLI docs also appear at /inference-api/dw-cli and on GitHub Pages: three places.
• Gotcha from the skill README: with API-key-only login, dw whoami may fail because API keys lack admin API access. The skill probes with a tiny realtime call instead.

dw whoami                       # browser login only
dw models list                  # needs full platform login
dw realtime "$MODEL" "Reply with OK." \
   --temperature 0 --max-tokens 2 --no-stream

Control Layer (open-source gateway)

• Docs are live at /control-layer. The site's llms.txt lists 11 pages: overview, getting started, 7 how-tos (API, endpoints, users & groups, health monitoring, production, model pricing, payments), configuration reference, "how billing works". It is simply absent from the top nav (Inference API and CLI only).
• Repo doublewordai/control-layer: Rust, Apache-2.0, 92★, updated 13 Aug 2026. Vendor claim: "450× less overhead than LiteLLM" (benchmark not re-run here).
• Quick start (search snippets of the docs): wget the repo's docker-compose.yml, docker compose up -d, open localhost:3001; needs PostgreSQL; image ghcr.io/doublewordai/control-layer:latest; OpenAI-compatible at /ai/v1. A Helm chart repo exists (control-layer-chart).
• Set a real SECRET_KEY (openssl rand -base64 32) and change default credentials before exposing it.

Inference Stack (self-hosted GPU deployment)

• Still marketed: a product link in the homepage footer, and a public repo doublewordai/inference-stack ("easiest & most performant way to run genAI infrastructure in your private environment"; Go templates, 10★).
• No Inference Stack pages appear in llms.txt, and a usage page returned 404 in my earlier session. So its docs are gone or moved; I could not confirm which.
• Related research repos: zerodp (zero-copy data parallel for MoE serving), inference-lab (an LLM inference simulator, not a request daemon as v3/v4 said).

Customers named on the homepage

Detail · nPlan · OpenMed · Ken AI · Element Materials · UnaGo AI · Dataiku · NXL · University of Colorado Boulder. Featured in Sky News, Sifted, Forbes, Bloomberg, TechCrunch.

All case studies (Kelp, NXL, Ken AI, UnaGo AI, Dataiku) · Seen in the Wild

Inference engineering (technical blog)

Multipliers are the vendor's headline claims. The older conceptual pages (why batch matters, batched endpoints, parallel primitives, ZeroDP) now sit in the docs Archive section: why batch · batched endpoints · parallel primitives · ZeroDP. The "Behind the Stack" episodes 4, 8, 9 cited in v4 are not in llms.txt.

Read these next (highest value)

• Dedicated Realtime, Data Sources / S3 (Beta), Get Support, Tool calling
• CLI pages, especially install and "Usage & Analytics" (could settle the analytics-endpoint question)
• Webhooks and Organizations (March data only)
• Control Layer how-tos (the nav omits them; re-score them)
• Whether batch-skill still redirects to doubleword-skill
• The community skills (Nnamdi Odozi, OpenClaw, Termo) which were not re-checked

Cheapest reliable refresh path

• Fetch docs.doubleword.ai/llms.txt for the page list; most pages have a .md twin.
• Re-copy the Models table into the data block of this file; the table, calculator and every discount figure recompute from it.
• Static doc pages fetch well; the interactive API-reference viewer returns little useful text (from earlier crawl notes).
• Docs deploys were out of sync on 5 Oct (D3): compare a page's sidebar before trusting one fetch.