BroadRouter

Fact-check log

Updated

This log lists the facts we got wrong and changed after re-checking the source: 60 entries across 51 pages so far. 44 were corrected, 12 clarified, 3 unsupported claims were removed, and 1 cheaper competitor was added where we had left it out. Each entry gives the old claim, the new one and the source.

Why we publish this

BroadRouter is BroadNet's own site, so you should be able to see how its facts are checked. Prices for BroadNet come straight from its price file and cannot be typed by hand. Other providers' LLM API prices are read from their own pricing pages and stored with a source, date and quote per row (how). Everything else, such as competitor limits, tool settings and research claims, is checked against its live source, with AI assistance and automated build checks, and this page records what those checks changed. Some corrections went against BroadNet's interest, for example adding a cheaper speech-to-text provider. They are listed here too.

Reviews completed

  • , Pre-launch review, pass 1: Every article, guide, glossary term, alternatives page, tool and core page re-checked against live sources.
  • , Pre-launch review, pass 2: A second, independent check of the same pages, aimed at errors the first pass missed or introduced. Code samples were syntax-checked and worked examples recomputed.

Corrections

DatePageTypeIt saidIt now says
/tools/rag-chunking-calculator/CorrectedThe worked example said retrieving 50 chunks would need 26,860 tokens a query, "more than" BroadNet Gemma 4 VL 26B's 32,768-token context. 26,860 is less, so top-50 fits; the sentence was printed whatever the numbers.The example uses top-70 (37,100 tokens with the answer, more than the 32,768-token context; at most top-61 fits), and the sentence says "more than" or "within" from the computed result. Source
/blog/reduce-prompt-tokens/CorrectedFAQ: 40 rows of indented JSON took 34 to 38% more tokens than the same rows minified, and 57 to 61% more than CSV.Those figures are the savings in the post's own table: minifying saved 34 to 38% and CSV saved 57 to 61%. Put the other way, indented JSON took 52 to 61% more tokens than minified JSON and 132 to 159% more than CSV.
/data/tokenizer-efficiency/CorrectedThe dearest tokenizer needs 2.42× the tokens of the cheapest for the same Modern Standard Arabic (the ratio of two ratios to English), while the Hindi line used raw token counts.Every dearest-to-cheapest gap is now computed from raw token counts on the same text: 2.37× for Modern Standard Arabic (Phi-4 against Ministral 3); the dialect and Arabizi gaps use the same method. Source
/llm-api-pricing/providers/perplexity/CorrectedEvery one of the 22 models that Perplexity shares with other providers "is cheaper elsewhere" at the standard tier.Provider pages now count ties separately: Perplexity ties the lowest standard price for 18 of the 22 and is beaten on 4. Ranks give equal prices the same rank, and a provider that only ties is no longer called the dearest. The same template fix applies to every provider page.
/tools/gpt-oss-token-counter/CorrectedCheapest standard price: GPT OSS 20B on AkashML, a price AkashML labels "Limited Trial". The same row was the lowest Arabic price in the real cost by language dataset.Prices labelled "Limited Trial" are treated as promotional: the token counters and /data/real-cost-by-language/ do not rank them, and pricing pages tag them "promo". Source
/tools/llm-batch-job-planner/ClarifiedBroadNet documents only two endpoints, /v1/chat/completions and /v1/audio/transcriptions (also on the migration checker, the RAG chunking calculator, two blog posts and the embeddings glossary entry).BroadNet's API reference also documents GET /v1/models and account endpoints under /portal/. The pages now say it lists no batch, files or embeddings endpoint. Source
/tools/openai-to-broadnet-migration/Correctedverbose_json is not in BroadNet's docs (also in the timestamps glossary entry and the Arabic subtitles post), and the audio splitter said BroadNet states no billing increment.BroadNet's /v1/models metadata (checked 2026-11-15) lists json, text and verbose_json for both transcription models and prices them per second of audio. It does not say verbose_json carries timestamps or how part-seconds bill, and the pages say so. Source
/llm-api-pricing/llama-3.3-70b-instruct/Corrected"4 standard prices come with less [context], down to 12,288 tokens (Novita AI, Featherless and CoreWeave Forge)": three names for four prices.The template names every provider behind the count (here Novita AI, Featherless, CoreWeave Forge and Venice) and says when several prices come from one provider.
/blog/arabic-subtitles-srt-api/ClarifiedOpenAI lists whisper-1 as deprecated, shutting down on February 26, 2027.OpenAI's deprecations page gives the same February 26, 2027 shutdown for gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, and recommends gpt-transcribe or gpt-live-transcribe instead; the table now says so. Source
/alternatives/deepgram-alternative/CorrectedOpenAI's option was gpt-4o-mini-transcribe at 0.003 USD a minute, with gpt-4o-transcribe-diarize for speaker labels, and neither was marked deprecated.OpenAI's pricing page marks both deprecated, and its deprecations page says they shut down on February 26, 2027. The page now uses gpt-transcribe, OpenAI's recommended model, at 0.0045 USD a minute (0.27 USD an hour), and says the older models are deprecated. Source
/blog/speech-to-text-api-pricing-compared/CorrectedThe opening answer gave OpenAI Whisper's 0.006 USD a minute as OpenAI's price. The price table and the diarization row listed whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize without saying they are deprecated.The answer quotes OpenAI's current gpt-transcribe at 0.0045 USD a minute. The four older models are marked deprecated (shutdown February 26, 2027), and the voice-note example adds gpt-transcribe. Source
/blog/whisper-api-alternative/CorrectedStay with OpenAI's replacement models if you need streaming or timestamp and prompt options.OpenAI's guide documents word timestamps only for the deprecated whisper-1 and speaker labels only for the deprecated gpt-4o-transcribe-diarize. The page now says to stay for live transcription or for gpt-transcribe's prompts and keyword hints, and it cites the deprecations page (notice August 26, 2026, shutdown February 26, 2027). Source
/glossary/diarization/CorrectedOpenAI's diarization model was given as a current example of an API that matches speakers to reference clips.The page names the model, gpt-4o-transcribe-diarize, and says OpenAI lists it as deprecated, shutting down on February 26, 2027. Source
/models/CorrectedBroadNet Inference "serves six models on BroadNet-owned or leased hardware in the United States, with partner capacity as backup" (models page, seven alternatives pages and several articles).BroadNet's Terms of Service (§7.3) name DeepSeek V4.1 Flash as partner-hosted, and its model catalogue lists it among models BroadNet never ran on its own hardware. Pages now say the six models are served on owned or leased US hardware or, for partner-served models such as DeepSeek V4.1 Flash, with US inference partners, under zero data retention either way. Source
/blog/deepseek-api-alternatives/CorrectedBroadNet serves DeepSeek V4.1 Flash on BroadNet-owned or leased hardware in the United States, per its privacy statement.BroadNet sells DeepSeek V4.1 Flash through its API, and its Terms of Service say the model is partner-hosted; partner-served models run under zero-data-retention terms with US inference partners. Source
/models/cohere-transcribe-arabic/ClarifiedRanked #2 of 37 on the Open Universal Arabic ASR Leaderboard (reported by BroadNet), with no date on most pages.BroadNet reported #2 of 37 in September 2026. When we checked the live leaderboard on 15 November 2026 it listed Cohere Transcribe Arabic 07-2026 #4 of 39 (average WER 24.19), behind Jais-ASR, Sygma-ASR-Arabic and Audar-ASR-V1-Turbo. Every page that quotes the rank now gives both. Source
/compare/cohere-transcribe-vs-cohere-transcribe-arabic/CorrectedBroadNet does not state which Cohere release it serves, and does not publish its own English accuracy figure for BroadNet Transcribe.BroadNet's model pages name Cohere Transcribe Arabic 07-2026 and Cohere Transcribe 03-2026 (both 2B, Apache-2.0) and report FLEURS word error rates of 5.5 (English) and 3.2 (Spanish) for BroadNet Transcribe; the comparison now shows them. Source
/blog/multilingual-speech-to-text-api/CorrectedBroadNet Transcribe Arabic "is the one BroadNet publishes accuracy for".BroadNet also publishes FLEURS English and Spanish figures for BroadNet Transcribe; it publishes Arabic figures only for the Arabic model. The public Arabic leaderboard ranks both. Source
/zero-data-retention/ClarifiedAn automated retention audit runs every hour "to verify that no content is being stored".BroadNet's privacy statement says its gateway configuration is checked hourly for content-free operation. The page now says it is BroadNet's own configuration check and that no third-party audit report is published, and adds BroadNet's statement that inference data is not used to train, fine-tune or evaluate models. Source
/blog/free-llm-api-credit/ClarifiedNo time limit was given for the free starting credit.BroadNet's Terms of Service say the sign-up trial credit gives one person or organisation 30 days to try the service, and that added funds are prepaid credit valid for 12 months. Source
/llm-api-pricing/CorrectedAlibaba Cloud rows for 15 models (among them Qwen3.8 27B, Qwen-Plus, Qwen-Flash, GLM 5.3, GLM 5.2 and DeepSeek V4 Pro 0813) showed no cached-input price, so workload costs billed their cached tokens at the full input price.Alibaba's Context Cache page sets the cache-hit price for these models in the International (Singapore) scope: 20% of the input price (18.57% for GLM 5.1, 10% for DeepSeek V4 Pro 0813 and V4 Flash 0731). The rows now carry that cached price. Source
/llm-api-pricing/qwen3.6-plus/CorrectedRow notes for Alibaba Qwen3.6 Plus, Qwen3.6 Flash and the Qwen3.5 Plus and Flash snapshots said implicit-cache hits bill at 20% of the input price.Alibaba's Context Cache page does not list these models for implicit cache in the International (Singapore) scope used on these pages, so the notes now say there is no cached price. Source
/llm-api-pricing/command-a-plus/CorrectedCohere (and, on the GLM 4.7 Flash page, Z.ai) "publishes no price" for the model on its own API.The maker lists the model as free on its own API; the pages now say so, and the opening answer names the cheapest paid price. Source
/llm-api-pricing/cheapest/CorrectedPhi-4 and GLM 4.7 Flash were ranked as mid-size models, because their names carry no parameter count.They are ranked as small models: Microsoft's model card gives Phi-4 14B parameters and Z.ai's gives GLM-4.7-Flash 30B (A3B). Source
/llm-api-pricing/cheapest/ClarifiedCheapest frontier model: Gemini 2.5 Pro at DeepInfra, although Google charges the same price.When a host only ties the maker's own price, the maker's API is named first: Gemini 2.5 Pro on Google's Gemini API. Source
/llm-api-pricing/ClarifiedThird-party LLM API prices on the pricing pages, the cheapest ranking, the price index and the cost calculator came from an aggregator's price feed (what a router charges per provider), which can differ from a provider's own price.Every third-party price now comes from the provider's own pricing page, stored with its source URL, check date and a quote; the aggregator feed is no longer used. Rankings use standard-tier prices only, with batch, flex, priority and free rows shown but not ranked. Source
/tools/image-token-calculator/CorrectedThe Qwen rule rounded image sides that fall exactly halfway between multiples of 32 up (a 1040 × 1040 image: 1,091 tokens).Halves round to even, as in Alibaba's smart_resize reference code (Python's round): 1040 × 1040 is 1,026 tokens. Source
/tools/image-token-calculator/CorrectedThe width × height ÷ 750 rule of thumb gives slightly more tokens than the Claude patch count.It is only an approximation: it overestimates large images and underestimates small ones (200 × 200 is 64 tokens, not 53). Source
/tools/image-token-calculator/ClarifiedThe page "reproduces" OpenAI's worked example of 765 tokens on GPT-4o.OpenAI's vision guide gives worked examples only for GPT-6 Astra (1,229, 3,000 and 2,458 tokens), which the page reproduces; 765 for GPT-4o is our own result from OpenAI's tile steps. Source
/tools/embedding-cost-calculator/CorrectedMistral's page does not say its batch discount covers mistral-embed, so no batch price was applied.Mistral's model page lists batching for mistral-embed and its batch docs give 50% off, so mistral-embed now has a batch price of USD 0.05 per 1M tokens, like codestral-embed. Source
/tools/self-hosting-vs-api-calculator/CorrectedBreak-even utilisation ignored the monthly engineering cost, so it did not match the break-even volume shown next to it, and the self-hosted per-1M figure assumed a full server.Both break-even figures include engineering cost, and the per-1M figure is at your volume, idle capacity included (the full-server figure is shown as a note).
/data/llm-api-free-tiers-and-rate-limits/CorrectedGroq and Cerebras: training on free-tier data "not stated".No for both: Groq's Services Agreement says it is not permitted to train on inputs or outputs without permission, and Cerebras's terms give it no right to train on service content. Source
/data/llm-api-free-tiers-and-rate-limits/ClarifiedOpenRouter and Mistral: no statement on training found.OpenRouter has separate training settings for free and paid models and routes to providers that may train only if you allow it; Mistral (no free API) says paid API data may be used for training unless you opt out. Source
/data/speech-to-text-api-prices/ClarifiedSpeechmatics Standard 0.45 USD an hour, quoted as shown on its pricing page, with "no minimum spend".The price is the pay-as-you-go standard rate from the page's calculator; the page displays 0.23 USD after a 33% model-training discount and the top 25% volume discount, and the row says so. The unstated "no minimum" was removed. Source
/fact-check-log/CorrectedOne entry said BroadNet's docs do not state how API keys are stored; another cited a privacy URL that returns 404.BroadNet's docs say the portal stores only the key's hash (its privacy statement lists SHA-256 hashes of API keys), so that part of the entry was removed; the privacy source now points to the live statement. Source
/blog/flat-rate-vs-pay-per-token-llm-api/CorrectedFeatherless was listed as a flat-rate API option for agents.Featherless's flat-rate Chat plan is limited to human-driven use and rules out app or API traffic, background automation and coding agents; its Developer plan for agents is billed per token. The row and FAQ now say so. Source
/blog/arabic-speech-to-text-api-buyers-guide/DisclosedSoniox was not mentioned.Soniox is listed with its per-hour price, and the page says plainly that it is cheaper per hour than BroadNet Transcribe Arabic. Source
/blog/whisper-api-alternative/ClarifiedNo retirement date for whisper-1.OpenAI lists whisper-1 as shutting down on February 26, 2027; the page leads with it. Source
/blog/llm-api-spend-control/CorrectedAn exhausted pay-as-you-go balance returns HTTP 402.It returns 401; 402 applies only to purchases such as an hourly key. Source
/blog/llm-inference-provider-pricing-compared/CorrectedOpenRouter lists DeepSeek V4.1 Flash on 35 endpoints from 29 providers; one credit-fee figure.37 endpoints from 28 providers; credit fee 5.5% on Standard and 8% on Business. Source
/blog/choosing-an-llm-for-arabic/CorrectedFalcon-H1-Arabic was described as open weights.Only a chat demo is published; ALLaM is credited to SDAIA and its Hugging Face preview to HUMAIN. Source
/blog/arabic-llm-gulf-egyptian-dialects/CorrectedThe free starting credit runs Qwen, DeepSeek and GLM.Of the chat models, only Gemma 4 VL 26B runs on the free credit. Source
/blog/run-parallel-coding-agents-flat-rate/RemovedA second request on a plan session queues.BroadNet's documentation does not say whether it waits or is refused, so the claim was removed here and in three related posts.
/blog/best-open-model-for-coding-agents/CorrectedCline works only with native tool calls.Cline keeps a text-based fallback; same fix in the Cline vs Roo Code vs Aider comparison.
/blog/cheapest-vision-llm-api/CorrectedImage-token arithmetic for Gemma and Qwen was wrong.Recomputed; against a Full Moon plan, per-token DeepSeek V4.1 Flash input breaks even at about 340,000 images a month.
/blog/llm-api-key-security/CorrectedCited the wrong GitHub secret-scanning page.The correct GitHub page is cited. Source
/blog/handle-llm-api-429-rate-limits/CorrectedThe Python sample used await outside a function, and the Node sample used undefined variables.Both samples fixed and syntax-checked.
/blog/llm-api-sensitive-data/CorrectedThe redaction regex matched IBAN digits as phone numbers.IBAN matching runs first.
/guides/roo-code-openai-compatible/CorrectedInstructions assumed Roo Code is maintained.Roo Code was shut down on 15 May 2026 (repository archived); the guide points to Cline or the Zoo Code fork. Source
/guides/n8n-openai-compatible/ClarifiedNo mention of the Responses API toggle."Use Responses API" must be turned off for OpenAI-compatible providers; it is on by default in the node's v1.3.
/alternatives/fireworks-ai-alternative/CorrectedTogether AI H100 pricing "from 3.99 USD an hour".Together AI lists H100 GPU clusters at 3.99 USD an hour on demand and 1.99 USD preemptible; same fix on the Together AI page. Source
/alternatives/requesty-alternative/RemovedSaid no card is needed.Removed: not stated on its pricing page. Five alternatives titles also gave the wrong number of options.
/glossary/logprobs/CorrectedPerplexity was defined loosely.Perplexity is the exponential of the negative mean log-probability.
/glossary/open-weight-model/CorrectedThe OSI definition requires the training data.OSI requires detailed information about the training data, not the data itself. Source
/glossary/vector-database/RemovedEmbeddings cannot be inverted back to text.Removed: published inversion attacks recover text from embeddings. Source
/data/llm-api-price-index/CorrectedWidest price gap for one model: 59×; median spread 2.0×.Promo-discounted and flex endpoints are excluded from spreads: widest gap 29×, median 1.8×. The unfiltered figure is kept as a caveat.
/data/tokenizer-efficiency/Corrected1,000 English words ≈ 1,142 tokens, measured on 12 short texts; Arabic 1.17–1.67× English.Version 2 measures 57 openly licensed texts; every page now quotes the dataset figures, which are computed at build time so they cannot drift.
/llm-api-pricing/cheapest/Corrected"Flex" endpoints were ranked as cheapest.Flex is a best-effort tier that can be slow or refused; it is shown and labelled but not ranked.
/models/CorrectedModels run on "GPUs BroadNet operates itself".Aligned with BroadNet's privacy statement: owned or leased hardware in the United States, with partner capacity as backup. Source
/tools/api-request-builder/CorrectedA shared link could insert shell commands into the generated curl command through the base URL.The URL is shell-quoted, and a warning shows when a link points your key at a non-default server.

Found an error?

Email info@broadnet.ai with the page address and the source. If we confirm it, we fix the page, update its date and add it here. See the editorial policy and about BroadRouter.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity