Changelog

New features, improvements, and fixes to the DeepRead platform.

November 10, 2026

Performance

Extract and Deep Extract return results faster, at the same price. Field extraction runs on a quicker model that matched or beat the previous one on most of our benchmark sets. Requests with speed=fast are unchanged

November 9, 2026

Feature

Fast mode. Add speed=fast to POST /v1/process and Deep Extract runs about twice as fast at the same accuracy, billed as Deep Extract Fast at $90 per 1,000 pages. Extract also accepts speed=fast at the Extract price. Available on Standard and Enterprise plans; on other plans the request is refused with 402. Requests without speed run, and are billed, exactly as before

Feature

mode=bundle on POST /v1/process. Send one file that holds many documents, such as a summary bill with one section per account, and get one record back per document

Fix

Long tables full of numbers keep their values. When a column held the same number on most rows (a column of 100.0, say), Extract could blank it or fill it with stray digits; those columns now come back as read

Fix

A PDF saved from an email client is accepted. Such a file can carry a few lines of email wrapping around the PDF, and the upload was refused as Invalid file type; the wrapping is now removed and the PDF is processed

Fix

The /signup link opens the signup form instead of a page-not-found

Fix

The Compare us to other providers button on the homepage shows its full outline

October 23, 2026

Improvement

n8n node 0.5.0. The DeepRead node for n8n gains Reviewer Email, Review Due (Hours) and Review Expiry (Days), a Webhook URL so a workflow can continue from an n8n Webhook node when the job or the review finishes, Bounding Boxes (include_markers) and Per Page. Polling waits for Retry-After on a 429 or a transient 5xx instead of failing the item. Update n8n-nodes-deepread from Settings → Community Nodes

October 22, 2026

Fix

A database that does not answer is a 503, not a 401. When our database did not answer while an API key was being checked, the API replied 401 Authentication failed although the key was fine, which tells a client to stop retrying or to rotate the key. It now retries once on a fresh connection and, if that fails too, answers 503 with Retry-After: 5 and "Your API key could not be checked. Retry in a few seconds." A bad key is still 401

October 21, 2026

Feature

Have a person review a document in DeepRead. Send reviewer_email with a job on POST /v1/process, one address or a team of up to 10 separated by commas, and that person is emailed a link when the document is ready. Only they can open it as a reviewer: they sign in with a code sent to their address, see the document and every extracted value, correct what is wrong and complete the review; in a team the first to complete it completes it, and their /reviews page lists everything waiting for them with how many values were flagged. The job completes as usual with review.status pending; your webhook_url then receives review.completed with the corrected data, who completed it and changes[] (each value’s previous and new reading, who changed it and when), and review.updated for every later fix. review_due_hours sets a due time (review.overdue when it passes; the review stays open), review_expiry_days closes an unfinished review (default 30 days; review.expired carries the unreviewed data), and PATCH /v1/jobs/{id}/review reassigns or cancels. On the preview page the sender sees who the review is waiting on and every correction with who made it. reviewer_email turns the preview on and cannot be combined with incognito

Feature

Spend per API key, and a monthly limit per key. The key list on the dashboard shows what each key has spent this month, and the last six months behind a click, so one key per project is one budget per project. Set a monthly limit when you create a key or from its row; a job that would take the key past its limit is refused with 402 (key_limit_reached) until the month turns or the limit is raised. The account balance still has the last word

Feature

Auto-recharge for Enterprise accounts. The billing page of an invoice-billed account has the same Auto-recharge and Payment method panels as a Standard account: set a threshold and an amount, and when the balance falls below the threshold the saved card is charged that amount and the credits are added. A card is saved by buying a credit pack with it

Fix

Auto-recharge rules. The monthly cap counts the calendar month (UTC), not a rolling 30 days. After a card is declined on every attempt, no new recharge starts for a day, and the email says so and links the billing page. When the cap holds a recharge back, you get one email that month with the amount recharged so far, the cap and when charges resume. A charge that Stripe left unconfirmed is checked again on the next attempt instead of blocking later recharges, and a failed charge leaves no open invoice behind

Improvement

Every webhook says which event it carries: event in the body and the X-DeepRead-Event header. job.completed and job.failed for every job, and review.completed, review.updated, review.overdue, review.cancelled and review.expired for jobs sent with reviewer_email. Accounts pinned to the legacy response shape are unchanged

Improvement

The billing activity list reads in plain words: a job whose cost is held while it runs is "In progress" and a settled one is "Done", instead of hold and settle

October 20, 2026

Fix

Emails quote your account’s own rates. The email after a credit-pack purchase and the onboarding email take their prices from your account, the negotiated rate where one is set, else the list price, instead of a price written into the template

October 18, 2026

Improvement

The favicon is the brand logo on a transparent background, and the landing pages carry the same icons as the app

October 17, 2026

Improvement

Extract runs on the engine the benchmarks page measures, for every account: Free accounts from October 16, every other plan from October 17. Your request, the pipeline value in the response, the product and the rate are unchanged, and page images are stored only when you ask for a preview, as before. Handwritten and low-confidence lines are read a second time by a second OCR model, and tables in long documents are copied row by row and checked against the main read

Fix

A JPEG with a valid header and no image data is reported as a corrupt upload. The job fails with that reason, the way a truncated or password-protected file does, instead of a server error

October 16, 2026

Feature

Enterprise accounts buy credit packs by card. The billing page of an invoice-billed account offers Buy credits and a card top-up panel, the same packs as Standard; a top-up is spent at the account’s own rate per product and the account stays on its contract. When a purchase order is spent, the 402 (credit_line_exhausted) says to top up at /dashboard/billing or to ask for a term

Improvement

Deep Extract reads money, labelled fields and tables more carefully. Money figures that pages disagree on are settled by the statement page, an exact figure beats its own rounding, column units such as £’000 are multiplied out, a computed total yields to a printed one, and a minus sign or credit-note notation the document prints is kept. Identifiers, dates, totals and terms the model left null or contradicted are settled from the document’s own labels. Long-document table copies keep a legitimately constant column and rows with blank optional cells. On the public benchmarks Deep Extract went from 87.7 to 92.4 on Kleister-Charity and from 94.5 to 97.8 on LongExtractBench, the highest average of the systems compared

Fix

The benchmarks page carries the 15–16 October runs for every system, scored the same way, and the Deep Extract row is back. Price claims compare DeepRead’s cheapest row with the best competitor and accuracy claims use our best row; the pricing rule is one sentence saying what a self-serve customer pays

October 15, 2026

Improvement

The phone menu lists Solutions as one row that opens into the industry pages, the way the desktop dropdown does, instead of eleven "Solutions · …" lines. On a solutions page the row starts open with the current page highlighted. The docs menu on phones and tablets does the same

October 14, 2026

Improvement

The benchmarks page compares DeepRead with Reducto, Extend and LandingAI on three public benchmarks: DocILE, Kleister-Charity and LongExtractBench, 2,614 pages per system, the same documents and schemas for everyone. It shows each system’s accuracy per benchmark, its price per 1,000 pages and what the whole run cost

Fix

Benchmarks page corrections. The headline price multiple is measured against the next most accurate API, not the cheapest one. The Kleister-Charity scores of the other providers come from the completed 40-document run; one document per provider had still been pending when the first table was built, and Charity is now a tie for first. The table shows Extract against the five competitor modes

Improvement

The brand green is the same across the site and every landing page, the site scrollbar is visible as a thin rounded thumb, and the benchmarks page follows the dark and light toggle like every other page

October 12, 2026

Feature

Low-balance alerts. Ask us to put an account on watch with a threshold and up to five extra recipients, and it is emailed once when its available credit falls below the line and once more when it reaches zero; a top-up above the line arms it again. Setting or moving the line never sends an alert by itself

October 9, 2026

Docs

The API reference describes layout blocks. Besides headings, paragraphs and tables, a page’s layout can carry a list (bulleted, numbered or a checklist of marked and empty boxes), a chart with the labels and values read from it, and an image with a role: logo, stamp, signature, barcode, photo or other. The reference says what each engine returns for checkboxes, charts, stamps, signatures and barcodes

October 8, 2026

Improvement

Two products, lower prices. Extract is $15 per 1,000 pages and Deep Extract is $35 (they were $20 and $40). Parse is retired as a separate product: pipeline=extract without a schema is now Extract, and returns the same Markdown, layout blocks and bounding boxes as before. Nothing changes in your requests. Jobs run before today keep the product and rate they were charged at

Feature

Your rates on the billing page. The billing page lists what your account pays per 1,000 pages for each product, and marks a rate as negotiated when it differs from the list price. Enterprise accounts see the rate of their current term the same way

Improvement

Negotiated rates can be set per product and for any plan, with an end date. When a negotiated rate ends, the list price applies again on the next job; a job is always charged at the rate in force when it was submitted

Fix

webhook_url must be an address our worker can call: http:// or https://, a public host, no username or password, up to 2,048 characters. localhost, private and link-local addresses and internal names are refused with 400 before the job is created

October 6, 2026

Improvement

A one-time API key link stays valid for 30 days. The link we send a newly onboarded account to collect its API key used to expire after 7 days unopened; new links now last 30 days. Links already sent keep the expiry they were created with

October 2, 2026

Fix

The reserved confidence field is gone from extracted fields. extraction.fields[].confidence was always null and was often read as a score; it is no longer returned. Doubt about a value is reported by needs_review and review_reason, and there is no numeric confidence. If your code reads confidence, drop that read

Improvement

Preview on every plan. Send preview=true with any Extract or Deep Extract job, Free plan included, and the job carries a link to the visual review page. Without preview=true no page images are stored, as before

Improvement

Request limits are checked up front: a schema may be up to 256 KB and a field description up to 2,000 characters. A larger request is refused with 400 before any page is processed

Fix

A page whose layout carries a malformed block order no longer fails the document. One bad value from the OCR model used to fail the whole job; the block now keeps its place and the job completes

October 1, 2026

Improvement

Values inside tables are located on the page. Each cell of a table row (line_items[6].amount, for example) now has its own bounding box with include_markers=true, and in the review page clicking a row value jumps to it on the document. Only top-level fields were located before

September 30, 2026

Improvement

OCR fills in what one reader missed without discarding the rest. When the second OCR reader finds text the first left out and the page image confirms it, only the missing blocks are added; the page used to be replaced wholesale. On our 53-document benchmark Deep Extract accuracy went from 92.3% to 94.7% at a slightly lower cost per page

Improvement

Extract and Deep Extract are more careful with long documents. A malformed reading window no longer triggers costly field-by-field searches across the document, an incomplete model reply is read again instead of turning into silent nulls, and a value taken from a repaired window is flagged for review

September 29, 2026

Fix

A file sent to a JSON endpoint gets a clear 422 that names the field, not a 500

Fix

Searchable PDF links no longer go dead. The link in artifacts.searchable_pdf_url used to stop working 7 days after the job finished. It is now signed each time you read the job, so reading the job again always gives a link that works for the next 7 days. This covers every job already processed; nothing needs to be run again

Fix

Per-page text is markdown. With per_page=true, pages[].content.text carried the page layout as a JSON string although its format said markdown. It is now the page as markdown: headings, key and value pairs, tables. The structured layout has its own new field, pages[].layout, already parsed. If your code reads JSON out of pages[].content.text, read pages[].layout instead

Fix

A blank page no longer leaves raw data in the document text. A blank page inside a document used to add a block of JSON to document.content.text; it now adds nothing

Improvement

Pages per document follow your plan: 50 on Free, 1,000 on Standard, 2,000 on Enterprise. A longer document is refused with 400 page_count_exceeded, which states the page count and the limit. Deep Extract accepts the full 2,000 pages; a cost ceiling per job used to refuse documents above 1,250

September 28, 2026

Improvement

New model generation under every engine. Google retires the Gemini 2.5 models on 20 October, so Parse, Extract, Deep Extract, form fill and PII redaction now run on Gemini 3.5 Flash-Lite, Gemini 3.1 Flash-Lite and GPT-6 Luna. Measured on 52 real documents before the switch: accuracy unchanged, Extract about a third faster, Deep Extract the same speed. Prices, plans and the API are unchanged

Improvement

Deep Extract jobs report engine version 1.2.0. The previous models can still be requested with version=1.0.0 until 20 October

September 25, 2026

Feature

Accounts we set up for you start with a temporary password that you replace the first time you sign in, and your API key reaches you through a one-time link that shows it once and then dies, so the key never sits in an email or a chat

Fix

Review flags now name the value they are about. One unsettled disagreement between the two OCR engines used to flag every field on that page with the same reason; now only the field whose value the engines disagreed on is flagged, with what each engine read, and the document-level flags list only what is still unsettled after resolution

Improvement

Preview links are opt-in. POST /v1/process no longer creates a public preview link unless you send preview=true; anyone with the link can open the document, so it now exists only when you ask for it. The older include_images=true still works. New links use 12-character tokens

Docs

The process response carries id (the docs said job_id), and the API description now promises a review flag and reason per field, not a per-field confidence score that was never returned

September 22, 2026

Improvement

Enterprise billing page now reads like the Standard one — available balance and what running jobs hold in the hero, the Purchased / Promotional / Held / Last charge tiles, and the activity table with export. Both Downgrade buttons work: an Enterprise account can move itself to Free or Standard and keeps its balance

Fix

X-RateLimit-* headers are no longer sent to Standard and Enterprise accounts. They describe the Free plan’s monthly page counter, and on an account that pays from its balance they read "limit 0, remaining 0", which SDKs took for a throttle. Free accounts keep them

September 21, 2026

Feature

Enterprise, prepaid — Enterprise is now sold as a prepaid term of pages: you receive a Stripe invoice, and the pages land on your account the moment it is paid. Warnings at 50 / 20 / 5% left, a hard stop at zero, no rollover

Improvement

Enterprise billing page shows the term — pages remaining, pages held by running jobs and the expiry date. While an invoice is unpaid it shows the amount, the pages it buys and a button to pay it, since nothing can be processed until then

September 20, 2026

Fix

Pricing page — the comparison table’s plan headers now stay pinned while you scroll, the phone menu scrolls so "Sign Up Free" is always reachable, and there is no sideways scroll at iPad widths

September 19, 2026

Improvement

Accessibility — every icon-only control (copy API key, download result, delete blueprint, the tick cells of the pricing table and more) now has a name for screen readers, and no page scrolls sideways at 150% text size

Fix

Pricing and billing plan cards line up at every width — the "Most popular" badge no longer overlaps the plan name, prices stay inside their column on tablets, and the site header no longer overflows between 768 and 1100 px

September 18, 2026

Performance

Billing page loads faster — it used to exchange the session for a fresh token once per request on the page (five in parallel); concurrent requests now share one

Fix

A crash in the page-load progress bar on the landing page is fixed

Fix

A job is no longer marked failed when another worker had legitimately taken it over after a heartbeat hiccup — the first worker now concedes it

September 17, 2026

Feature

Two engines by name — pipeline=extract (one OCR pass) and pipeline=deep-extract (two passes, an LLM judge, a verification read). The old names fast, standard and searchable keep working and keep answering with the name you sent; searchable is now deep-extract with searchable_pdf=true. GET /v1/pipelines lists both

Feature

Three products, priced per page — Parse (extract, no schema) $10, Extract (extract with a schema) $20 and Deep Extract $40 per 1,000 pages. Every job carries its product and rate, and the Jobs page names what a job was billed as

Feature

Standard is pay as you go from prepaid credits — buy a $50, $100, $500 or custom pack on the billing page; each job holds its cost at submission and settles on completion. Optional auto-recharge from the saved card with a threshold, an amount and a monthly cap. Credits never expire and there is no monthly fee. The first pack moves a Free account onto Standard; Downgrade takes you back to Free with the balance kept

Feature

New pricing page and plans — Free (2,000 pages a month, API only), Standard (pay per page) and Enterprise (custom). One comparison table, with the plan cards generated from it so the two cannot disagree

Feature

idempotency_key on POST /v1/process — send the same key on a retry and you get the job the first request created, never a second one

Feature

Signed webhooks — every delivery carries X-DeepRead-Signature: t=<unix>,v1=<hex>, an HMAC-SHA256 over the raw body. Your secret is at GET /dashboard/v1/webhooks/secret and can be rotated

Feature

More file types — Standard adds TIFF (multi-page), WebP, BMP, GIF, DOCX and TXT; Enterprise adds DOC, ODT, RTF, PPT, PPTX, XLSX, XLS, CSV, HTML and more, plus PSD, APNG and other image formats. Everything is converted to PDF in a sandboxed process before the page count. A type off your plan answers 415 naming the plan that has it

Improvement

Plan limits, stated and enforced — submissions per minute 10 / 100 / 500, pages in flight 16 / 200 / 500 and polls per minute 20 / 60 / 150 (429 with Retry-After); file size 15 / 50 / 500 MB; and a hard maximum of 2,000 pages or 500 MB per document on every plan (413)

Improvement

Features by plan — webhooks, blueprints, the optimizer and the review UI from Standard; PII redaction, form fill, searchable PDF, incognito and retention_days on Enterprise. A feature off your plan answers 402 naming the plan that has it

Improvement

Field locations — every located schema field now carries location.bounding_box in the response, not just its page. Dense pages are located in chunks, so documents with many fields no longer come back with no boxes (deep-extract coverage on the benchmark set went from 38% to 99%)

Docs

Every docs page rewritten for the release — engines, products, file types, limits, idempotency, signed webhooks, plan gates, credits and billing

September 16, 2026

Fix

Free plan pages now reset monthly on your signup anniversary, as every page has always said — nothing was resetting the counter before. The 429 body, X-RateLimit-Reset and the dashboard show your account’s own next reset date

September 15, 2026

Fix

PNG and JPEG uploads now count against the Free allowance up front, the same as PDFs, and are refused with 429 when none is left instead of being counted after the run

September 14, 2026

Performance

A page whose OCR read loops to the model’s output ceiling is handed to the fallback engine at once instead of being retried, so a one-page document can no longer take ten minutes

September 11, 2026

Feature

Retention add-on — pass retention_days=N (1–365) to POST /v1/process and everything about the job is deleted N days after submission: the document, preview page images, searchable PDF and the extracted results. Until then the job works like any other; from the deadline on GET /v1/jobs/{id} returns no content, the preview link answers 410, and data_deleted_at confirms once every stored object is gone. Combine with incognito=true for a document that is never stored in the clear and results that expire

Improvement

Deleting a job from the dashboard now also deletes its stored document, preview page images and searchable PDF, not just the record

Fix

Running jobs are no longer interrupted by scale-down — worker capacity is only reduced once the queue is fully idle, so a long document never restarts from page one part-way through

Improvement

Processing time limits no longer fail large documents — the per-page budget is doubled to 30 s per page (4 h cap), and a job that still runs out of time is retried with a doubled ceiling, up to three attempts, instead of failing

September 10, 2026

Performance

Batch throughput — worker capacity now scales to up to 50 replicas within seconds of a backlog (previously capped at 10) and drops straight back down once the queue is empty, so large batches finish sooner

September 9, 2026

Feature

Result-first playgrounds — the OCR, PII and Form Fill playgrounds now open with a finished sample result (three documents, zoomable viewer), and you can upload your own document without signing in or creating an API key: 3 runs a day, up to 5 pages and 10 MB. Sign in to save a result, and every run from that browser is attached to your account

Performance

Very large documents (1,000+ pages) now process reliably — worker memory no longer grows with page count. A page whose OCR call times out is retried in a narrower pass and then falls back to the direct OCR engine; a page that still has no text is flagged for review and listed in errors, never returned blank silently

Performance

Faster job status lookups and dashboard job lists — the database indexes the queue and per-user queries needed are now in place (a 30-day job list went from 1.9 s to under 10 ms)

September 8, 2026

Fix

New accounts receive the welcome email again — a race between sign-up and the first dashboard load had been skipping it

September 4, 2026

Improvement

failed is now final — transient problems on our side (provider outages, capacity) never fail a job. It is retried with backoff for up to 48 hours while reporting queued or processing, and exactly one failure webhook fires, only on a final verdict. Applies to OCR, PII and Form Fill alike

Fix

Extra worker capacity now actually comes online during a backlog — the replica change used to be recorded without a rollout, so bursts drained on a single worker

Docs

Jobs & Webhooks and Error Handling guides updated for the retry contract — put retry logic on the POST /v1/process submission (429/5xx), not on failed jobs

September 1, 2026

Fix

PII redaction fail-safe — if PII detection cannot run on a page, the job now fails with a clear error instead of ever returning an unredacted document marked complete. Transient provider errors are retried automatically first

Improvement

All account emails — welcome, password reset, optimizer results, shared results and playground results — now share one branded design; three of them previously went out unstyled

Fix

Browser tab title now follows the page you are on — after a visit to the blog, every later page kept the blog post title

August 31, 2026

Feature

Incognito processing — pass incognito=true to POST /v1/process and we crypto-shred your document the moment the job finishes: the stored file is deleted and its single-use encryption key destroyed, so nothing readable remains on our side. Available on fast and standard tiers

Improvement

Optimizer small-dataset guidance — POST /v1/optimize now returns a warning when you supply fewer than 8 training documents, and result emails report the accuracy of the schema version that is actually live, measured on held-out documents

Improvement

PII redaction on scanned pages is more reliable — page OCR now runs in an isolated, resized path so large scans no longer slow down other jobs

August 26, 2026

Feature

PII and Form Fill playgrounds — try PII redaction and PDF form filling directly in the docs playground from your browser, no code required

August 25, 2026

Fix

Searchable PDF add-on now works on PNG and JPEG uploads, not just PDFs

Docs

Docs cleanup — corrected every API example, fixed a code-block rendering crash, and refreshed pricing and optimizer content

August 24, 2026

Performance

Blog loads 98% lighter — cover images are now served resized and in WebP, and article bodies are fetched per post instead of shipping every article to every visitor. The listing also fits 12 posts per page so rows always fill

August 6, 2026

Feature

Blog at /blog — search-indexed articles rendered as static HTML with full meta tags, JSON-LD, sitemap, and RSS feed

July 29, 2026

Feature

Optimization results by email — we email the requester when a run finishes, with the accuracy it reached and whether the tuned schema was activated

July 27, 2026

Feature

Self-serve Optimizer — optimization requests now run automatically instead of being scheduled for manual processing, and the optimizer page shows live progress (round, accuracy, spend) while a job runs

Improvement

Optimization progress is now readable while a job runs — poll GET /v1/blueprints/jobs/{job_id} for status, iteration, accuracy and spend, or watch the same numbers on the optimizer page

July 24, 2026

Feature

API playground in the docs — upload a document, pick a tier, and run a real extraction from the browser without writing any code

July 17, 2026

Improvement

Plan-aware pipeline default — requests that omit `pipeline` now default to `fast` on the free plan and `standard` on paid plans. Pass an explicit tier to override

June 27, 2026

Feature

Logistics landing page — a vertical landing at /solutions/logistics (in the Solutions menu) with a "click one" demo that turns three real freight documents (proof of delivery, freight invoice, bill of lading) into clean structured JSON

May 20, 2026

Feature

BYOK redesign — path-picker UX with primary + secondary slots. Three real setups: OpenRouter (1 key, identical accuracy), cross-family pair (OpenAI + Google for genuine cross-family second opinion), or single direct provider with same-family tier swapping

Feature

BYOK Jobs tab — a dedicated job list on the BYOK page showing per-job provider(s) and your BYOK cost. The main Jobs page now only lists DeepRead-billed jobs

Improvement

BYOK reliability fixes — Gemini direct multipart-response parsing, max_output_tokens enforcement, defensive guards across the pipeline, and surfaced failing-step diagnostics on crashes

April 16, 2026

Feature

Jobs page — unified job history across OCR, Forms, and PII with multi-select product and status filters, date range presets, file name search, and inline result previews/downloads

April 1, 2026

Feature

BYOK (Bring Your Own Key) — use your own AI provider API keys for document processing with zero DeepRead cost

Docs

BYOK documentation with setup guides for OpenRouter, OpenAI, and Google providers

March 17, 2026

Feature

PII Removal — automatically redact sensitive data from documents

Improvement

Schema View UI improvements for easier field configuration

March 10, 2026

Feature

Preview page now supports per-page viewing and inline editing

February 21, 2026

Feature

Form Fill API — programmatically fill PDF forms with extracted data

February 10, 2026

Feature

Benchmark diff viewer to compare extraction results side-by-side

Improvement

API key management — regenerate and revoke keys from the dashboard

Improvement

Password reset flow for account recovery

January 20, 2026

Feature

Human-in-the-loop (HIL) review interface for low-confidence fields

Improvement

Benchmark page with competitive scoring against other OCR providers

Improvement

Dark mode enabled by default

January 13, 2026

Feature

Blueprint Optimizer — create and manage extraction schemas visually

Feature

Webhook support with signature verification for secure callbacks

Fix

API response format now consistent across all endpoints