Alfrada Docs

Appearance

Release Notes ​

User-facing product updates, new docs coverage, and notable workflow improvements.

How To Read These Notes ​

Each note answers four questions — in this order:

  • What you can do — concrete capabilities with specifics. Names, numbers, limits.
  • Where this shows up — real moments you'd actually hit this, in your voice.
  • Try it — copy-ready prompts you can paste today.
  • Heads up — only when there's something you genuinely need to know: pricing, limits, breaking behaviour, migration.

Current Releases ​

July 14, 2026 ​

  • Value Modelling — See The Expert Hours & Dollars Alfrada Saves You — Settings → Account → Hours Saved shows expert hours saved, dollar value at per-profession rates you can edit, compression multiples, and per-conversation receipts with satisfaction labels. Analysis runs on local Gemma (private, tool I/O stripped), follows Anthropic's Economic Index methodology (task-level value at expert wage rates, conservatively flexed 50%), and backfills up to 90 days on demand.

July 13, 2026 ​

July 10, 2026 ​

  • GPT-5.6 Is Live — Luna, Terra and Sol Scale to the Job — three GPT-5.6 models join the picker with 1.05M context, 128K output, tools, Swarm, and complexity-aware reasoning: Luna for fast high-volume work (Budget 1×), Terra for everyday coding and agents (Premium 6×), and Sol for the hardest reasoning and multi-step workflows (SOTA 12×). Auto can route across all three; Luna can also work inside multi-agent runs.

July 6, 2026 ​

  • GLM 5.2 Now Routes Direct To Z.ai — Free This Weekend — GLM 5.2 now uses a direct Z.ai API route (OpenRouter vendor retired); same 1M context, reasoning, tools, and swarm eligibility with lower latency. Free Fri–Sun 10–12 July 2026 (UTC) — zero token-budget and usage cost — then automatic return to 6× premium billing from 13 July (UTC). Pick GLM 5.2 in the selector; the old GLM 5.2 (via OpenRouter) entry is gone.

July 5, 2026 ​

  • Memory Now Remembers How Things Changed — long-term memory becomes a version-controlled timeline: every change is a dated commit with a before/after diff (+ created / ~ superseded / − invalidated / ± edited) attributed to you, the agent, a migration, or a re-mine. Facts are superseded, not overwritten — the old one is kept and stamped "true until…" while the new one takes over. New Timeline and Entities views (a lane per person/org/project/place) live in Digital Twin → Memories, with a per-fact history drawer. A background, billing-exempt, resumable re-mine rebuilds your timeline from your whole chat history, dating each fact to the conversation it came from. Existing memories import automatically as dated genesis commits; superseded ≠ deleted.

July 4, 2026 ​

  • Wake Me — Alfrada Checks Back In The Same Conversation — wake_me schedules a wake-up that resumes this same chat after 5 minutes to 24 hours so Alfrada can follow through on time-dependent work (awaited email, long render, intraday check) without polling or opening a new session. One pending wake per conversation; 10 consecutive self-wake turns max; five pending wakes per user. Always on and free. Not WhatsApp/Argus — use Task Scheduler for recurring or clock-anchored jobs.

July 1, 2026 ​

  • Claude Fable 5 (Mythos) Is Back — Anthropic's Strongest Model, Selectable Again — after the June 13 pullback, Claude Fable 5 (Mythos) returns as a deliberate manual pick [Not Private]: highest quality rating of any Alfrada model (above Opus 4.8), adaptive thinking, 1M context, 128K output, and the Overkill band (24× tokens). Still non-ZDR (provider may retain inputs/outputs) so it's kept out of Auto routing; routed through Bedrock's Mantle endpoint (Global by default), with EU-only geofencing supported. Supersedes the June 10 and June 13 notes.
  • Claude Sonnet 5 Is Live — Anthropic's New Workhorse for Coding and Agents — Claude Sonnet 5 joins the picker with 1M context, adaptive thinking, 128K output, Premium-band pricing (same rates as Sonnet 4.6), Workhorse Auto routing, and swarm eligibility. Global Bedrock only — hidden when EU Data Residency is enabled. Cloud release only; the desktop Inference settings tab does not ship to the web app.

June 28, 2026 ​

  • Memory & Playbooks, Now Focused Tools — long-term memory's single multi-purpose tool is now five focused tools: memory_search (search or list), memory_create, memory_update, memory_delete, and a dedicated playbook tool. Saves now require the namespace and text together, so a memory either lands completely or says exactly what's missing — no more half-saved entries or retry loops, on any model. Same "Agent Intuition" feature; existing memories and playbooks carry over and default playbooks auto-update to the new names.
  • Self-Checking Answers + Pick How Hard Alfrada Works — on bigger tasks, Alfrada now writes a task-specific rubric, grades its own draft against it, and revises before you see it. Medium does one self-review pass; Beast keeps revising until the rubric clears and also runs a vision pass over generated charts/images (axes, legends, cut-off text). A new performance slider lets you pin Fast / Medium / Beast or leave it on Auto, with a badge showing the mode that actually ran and sticky persistence across the session. Trivial turns are never graded. Grader usage is billed.
  • Model Picks That Fit Your Machine — the picker now reads your device's AI memory, sorts you into a RAM band (<16 / 16–32 / 32–64 / 64–128 / 128GB+), and recommends one strong local primary per tier — never a model that can't fit. Cloud-only servers skip the probe and show cloud picks. GPT-OSS 120B context grew 32K → 131K tokens.

June 17, 2026 ​

  • GLM 5.2 Is Live — Z.ai's Long-Horizon Agent Flagship — GLM 5.2 joins the picker as Z.ai's upgraded long-horizon agent flagship: 1M-token context (up from GLM 5.1's ~203K), reasoning enabled, tool support, swarm eligibility, and SOTA-band placement at the same premium pricing as GLM 5.1. Suited for project-level software engineering, long-running agent workflows, and complex multi-step automation. Auto-routable in the SOTA tier on every plan (scored on quality and speed), and directly selectable too.

June 13, 2026 ​

  • Kimi K2.7 Code Is Live, Fable 5 Pulled Back — Kimi K2.7 Code replaces Kimi K2.6 in the public specs sheet as Moonshot's coding-focused SOTA model: 262K context, reasoning enabled, tool support, and swarm eligibility. Anthropic Claude Fable 5 (Mythos) has been pulled back after a DoW notification and is no longer selectable while the notice and related obligations are reviewed.

June 10, 2026 ​

June 9, 2026 ​

  • Edit Uploaded Word Documents In Place, Formatting Intact — the new Doc Patch tool surgically edits a .docx you uploaded instead of rebuilding it: inspect every paragraph, table, image, and comment with stable references, then replace exact text while keeping run formatting, patch table cells and rows, swap an image, or insert/delete blocks — preserving surrounding text and table styles. It can also author a brand-new .docx from structured sections. Every edit runs against a required preservation contract and returns an honest change report. Use Report Generator for net-new reports from markdown. Free on every plan.
  • Edit Uploaded PowerPoint Decks In Place, Design Intact — the new PPTX Patch tool surgically edits a .pptx you uploaded instead of rebuilding it: inspect every slide's shapes, text, image positions, notes, and animation XML, then replace text, swap an image while keeping its exact size/crop, move/resize/hide/delete a shape, duplicate a slide, or update notes — preserving the original design, masters, and existing animations. Every patch runs against a required preservation contract and returns an honest report of what changed and what couldn't be guaranteed. Use Presentation for brand-new decks; new-animation authoring is not supported. Free on every plan.

June 6, 2026 ​

  • Nemotron 3 Ultra Is Live — NVIDIA's Frontier Orchestration Reasoner — Nemotron 3 Ultra [Global] lands in the picker: NVIDIA's 550B / 55B active hybrid Transformer-Mamba MoE for long-running agentic planning, 256K context, tool support, swarm eligibility, and Workhorse-band pricing. In the Auto routing pool on every plan; global via OpenRouter — use Nemotron 3 Super 120B [EU] when you need EU residency.

June 3, 2026 ​

  • Live Video Studio — Edit Chat-Built Videos Without Starting Over — the Live tab now opens Video Studio while Alfred runs video_editor or video_lab: one panel per session, a step list during the build, then preview + horizontal timeline (drag clips, trim, transitions, colour grade, audio, subs) after finalize. Update preview rebuilds your edits on the CUDA sidecar.

June 02, 2026 ​

  • MiniMax M3 Is Live — Long-Context Workhorse For Agentic Tasks — MiniMax M3 is now available as a Workhorse-tier model: 524K context, tool support, swarm eligibility, budget-band pricing, and a conservative quality score based on the first judged leaderboard run. Strong first signal on baseline tool orchestration; not promoted to SOTA yet because deep prediction and swarm timed out under the benchmark harness.

May 30, 2026 ​

  • HTML Dashboards — Beautiful, Interactive Reports That Render Live And Export To PDF — the new HTML Dashboard tool builds live, interactive dashboards (KPI walls, research readouts, comparisons, roadmaps) from six prebaked templates, renders them in a new Dashboards panel, refreshes the same dashboard in place when data changes, and exports to PDF / PNG / HTML. Self-contained (Chart.js, inline data), on-brand with auto dark mode, free on every plan.

May 29, 2026 ​

  • Claude Opus 4.8 Is Live — Anthropic's New Frontier Flagship — Claude Opus 4.8 is in the picker: Anthropic's most capable frontier model with adaptive thinking, a 1M-token context window, and the highest quality rating of any model Alfrada offers. Built for strategic planning, financial analysis, and deep research; swarm-capable; EU-only geofencing supported for in-region data residency. Replaces Opus 4.7 and 4.6 at the same price. Available on every package.

May 25, 2026 ​

May 23, 2026 ​

  • Qwen 3.7 Max Is Live — Alibaba's New Frontier Flagship — Qwen 3.7 Max replaces Qwen 3.6 Max Preview in the picker: frontier MoE for agentic tool use and long-horizon reasoning, 256K context, swarm-eligible, same pricing as the preview. Pinned 3.6 Max Preview sessions autoupgrade on the next message; not in Auto routing yet.

May 20, 2026 ​

  • Auto Just Got Cheaper — Gemini 3.1 Flash-Lite Is The New Default, Gemini 3.5 Flash Joins The Picker — Auto's default model swaps from gemini-3-flash-preview to gemini-3.1-flash-lite at half the per-token cost, with the same 1M context and tool support. Background intent/Q&A/summarisation moved off Gemini 2.5 Flash-Lite onto 3.1 Flash-Lite in the same step. Separately, Gemini 3.5 Flash lands as a Workhorse-tier pick — 1M context, reasoning on at medium effort, swarm- and worker-eligible — for harder agentic loops where you want frontier Flash quality at a Workhorse band cost.
  • xAI Grok Imagine Joins Image Lab Through OpenRouter — pin provider: "xai" in Image Lab to render 1K/2K posters, ads, packaging, menus, social graphics, and reference-image edits through x-ai/grok-imagine-image-quality on OpenRouter. Gemini remains the default path, Higgsfield Soul remains the moderation fallback, and xAI's named-entity rendering is documented as capability rather than legal clearance.

May 19, 2026 ​

  • xAI Imagine Video Joins Video Lab — Native-Audio Clips With A Safety-Net Fallback Chain — Video Lab now talks to three providers (Runway, Higgsfield, xAI). Pin provider: "xai" for short native-audio clips on x-ai/grok-imagine-video via OpenRouter (1–15s, 480p/720p, eight aspect ratios, up to seven reference images). provider: "auto" now runs a three-stage fallback chain — xAI → Runway → Higgsfield — with an Image Lab bridge so a text-to-video request never silently falls back to a Google image search. Every auto response carries a fallback_chain block showing which provider succeeded.

May 18, 2026 ​

  • Image Lab Now Auto-Falls-Back To Higgsfield Soul When Gemini Blocks — when Gemini's safety filter refuses an image (named franchises, IP overlap, silent refusals), Alfrada now quietly re-runs the same prompt on Higgsfield Soul and tells you which model produced the file. Pre-emptive prompt rewriting catches known-blocking combos (real person + named IP + photoreal) before submission, and you can chain refinements ("darker lighting", "more rain") through Soul without ever going back to Gemini.
  • Higgsfield Joins Video Lab — Cinematic Camera Moves Runway Can't Quite Do — pin provider: "higgsfield" on any video_lab call for native cinematic camera arcs and landscape pans. Four Higgsfield models ship today: DOP Lite for previews, DOP Turbo for working drafts, DOP Standard for keepers, and Kling 2.1 Pro for landscape pans (5 or 10 seconds). The new video_provider_routing playbook explains when to pick which provider; failures surface with retry suggestions and never silently fall back to Runway.
  • Google News — Localised News Search That Cites Itself — google_news runs Google News via SearchAPI with country/language auto-localised to your detected location (Berlin → German edition, London → UK edition, Karachi → Pakistan edition). Top 5 articles auto-save as .news.json citations in the session out/ folder, indexed for report_generator and presentation. Thumbnails decode from SearchAPI's base64 data once on the backend so they render in the chat card without inflating LLM token cost. Filters: recency (last_hour through last_year), sort order (most_recent or relevance), pagination. Per-row inline citation surfaces the publisher hostname + saved filename next to each article.

May 16, 2026 ​

  • Google Images — Localised Image Search That Saves Straight Into Your Session — google_images runs Google Images via SearchAPI with country/language auto-localised to your detected location; top 5 results download into out/ by default and are indexed for presentation, video editor, and report generator. Filters cover image size (16 options including megapixel thresholds), color (15), image type (photo, clipart, line_drawing, gif, face), aspect ratio, recency, usage rights (Creative Commons / commercial), SafeSearch, and pagination. Each saved file ships with a credit line for inline rendering.
  • Amazon Search & Amazon Product — Localised Shopping On Your Storefront — amazon_search and amazon_product ship together; both auto-route to the user's local Amazon (amazon.de from Berlin, amazon.co.uk from London, amazon.in from Bangalore) across 21 storefronts. Results are filtered to items deliverable to the user's country. Search supports sort orders (featured / price / review / newest / bestsellers), price min/max, and pagination. Product takes an ASIN or full Amazon URL and returns title, brand, rating, buybox (price + availability), feature bullets, attributes, and variants. URL form auto-detects the storefront.

May 12, 2026 ​

  • Argus Owner Gets Ludicrous Mode, Group Members Get The Builder Bundle — owner Argus auto-merges every active tool in the catalog, no tool pre-selection needed; group/external senders now get a curated 33-tool allowlist covering research, analysis, and artifact generation (PDF/DOCX/XLSX/PPTX, code execution, image/music/video, finance and social search, medical_vision) while remaining locked out of owner data, owner state, and every Composio MCP.
  • Discord — Identity, Servers & Invites — connect your Discord account for a 10-tool read-only account surface: profile, server (guild) listing with member counts, invite resolution (codes, vanity codes, discord.gg URLs), widget/template inspection, third-party connection inventory, username/avatar management, and leave-guild. No channel messaging (Composio toolkit limitation); routes you to Slack/Teams/Gmail for that.
  • Argus Is A Full WhatsApp-Powered Agent — Argus now has a full operating guide, clearer setup path, autonomous heartbeat check-ins, model selection, expanded tool access, and use cases for running a capable agent from WhatsApp.

May 10, 2026 ​

  • Autonomous Workflows Are More Dependable — scheduled tasks run with fewer approval dead-ends, long-form briefs get expanded when too short, presentation creation retries drop, mobile chat controls stay in frame, and artifact outputs now end with concrete next-step offers

May 05, 2026 ​

  • Grok 4.3 Lands In The Picker — xAI's newest reasoning model is selectable now: 1M context, always-on reasoning, image input, tool calling, workhorse pricing; manual-pick only for now (not yet in Auto or Swarm)

May 04, 2026 ​

May 03, 2026 ​

April 30, 2026 ​

  • Qwen 3.6 Family Goes Live — Flash, 27B, Max Preview — three new Qwen 3.6 models in the picker today; Flash with 1M-token context (new tier), 27B dense reasoner, and Max Preview ~1T frontier MoE; three Qwen 3.5 entries retired with automatic autoupgrade for pinned sessions; every credit band stayed the same

April 29, 2026 ​

  • Argus On WhatsApp — dedicated Argus operating guide now live (activation flow, ACL/group mode rollout, sender trust boundaries, and whitelist/blacklist warnings), Tool Library now links directly to the Argus page, and the default server model for Argus sessions moves to Kimi K2.6

April 27, 2026 ​

April 25, 2026 ​

April 24, 2026 ​

April 23, 2026 ​

  • Medical Vision — AI Second Opinion On Your Scans (Local, Experimental) — new medical_vision tool runs MedGemma 4B locally via Ollama for X-rays, CT/MRI, ultrasound, dermatology, lab reports; inspect mode re-reads any image with forensic detail; 50k context window, pinned-warm residency, UI tagged Experimental, always disclaimed
  • Runway Gen-4.5 Takes Over Video Generation — video_lab now calls Runway Gen-4.5 for every clip (2-10s, preset aspect ratios, first+last keyframe, optional watermark); video_editor is fully NVENC-accelerated; retake/extend/upscale/ic_lora/id_lora/create_from_audio return temporarily_unavailable with migration guidance

April 21, 2026 ​

April 20, 2026 ​

April 19, 2026 ​

April 18, 2026 ​

April 17, 2026 ​

April 16, 2026 ​

  • Claude Opus 4.7 — Anthropic's frontier model with adaptive thinking, 1M context, swarm support

April 15, 2026 ​

Writing Standard ​

Each release note follows a human-centred template:

  1. Title and date at the top — title is the change, not the jargon for the change.
  2. Lede — one or two sentences in the user's voice, answering: what is this and why would you use it today?
  3. What you can do — concrete capabilities with specifics (names, numbers, limits). List every meaningful thing the feature lets the user do — don't abbreviate the capability list. Name sub-tools or sub-tasks where relevant.
  4. Where this shows up — two or three real moments where someone would actually hit this, in their voice. "You want X. You used to have to Y. Now you just Z."
  5. Try it — copy-ready prompts the user can paste.
  6. Heads up — only when there's something the user genuinely needs to know. Skip the section if there isn't.

The template lives at _template.md.

Built for the Alfrada platform.

Copyright 2026 Alfrada