What we
ship.

• One click to copy a public link. On the Artifacts page, click the green published tag under any shared artifact (or the link icon on its tile) to copy its public link, without opening it first.

• You can leave a chat that is still answering. Clicking another chat used to keep you on the one that was writing until it finished. It now opens the chat you picked straight away; the first one keeps writing in the background, shows its purple dot in the sidebar, and picks up where it is when you go back.

• A published page is the whole page. Pages, drawings and decks you share now fill the browser under one slim bar, instead of sitting in a box under a heading. Documents keep their reading column and get the same bar.
• Tidier link previews. A document's preview text ends on a full sentence.

• Link previews show the artifact. Paste a published artifact's link into Slack, WhatsApp, X or LinkedIn and the preview now shows a picture of it with its title, instead of our logo. A page or a drawing appears as itself; a document or code shows its title and opening lines. Editing the artifact refreshes the preview.

• Share is a button now. It sits in the artifact's header instead of the ⋯ menu. Before the link exists you're told that anyone with it can view the artifact; once you confirm, the link is right there to copy. Open Share again to copy it or revoke it.
• Find what you've published. The Artifacts page has a Published filter, and shared artifacts are marked in the gallery.
• New Discord invite. The Discord link in the sidebar points to our new server.

• A full redesign. Every page is rebuilt on one calm design system: one typeface (Manrope), quieter colour, light and dark themes that match, and layouts that fit your phone, so nothing runs off the screen any more.
• Chat is here. Talk to any model on your plan right in the dashboard: attach files, edit or regenerate a reply and keep both versions, make artifacts, and share a chat with a link.
• Koda Remote has a new home. Your Koda sessions live under Koda in the dashboard. Open one from any device, see its history, and pick up where you left off; a continued conversation keeps the same session.
• A new model picker and sign-in page. The picker opens on eight recommended models, with everything else your plan can use one tap away under More models; and the sign-in page wears a different gradient each visit.
• Fixes: chat titles no longer start with a stray colon, and the chat header and composer stay put when a notice is showing or your keyboard opens.

Headings, numbers and the interface now all use Manrope, including diagrams in chat. Code, API keys and other fixed-width values stay in JetBrains Mono so they still line up.

• Mermaid diagrams are drawn. A diagram artifact used to show its source; it now renders as a picture, in the viewer and in the gallery. One that will not parse shows its source with the reason.
• Drawings no longer come up blank. An SVG that only set a viewBox rendered as an empty frame; it now fills it.
• A failed reply keeps saying why after you reload, instead of showing an empty message. And when a model simply is not available, you are told so straight away rather than asked to retry.
• Install commands keep your key out of the URL. The one-liners on the MCP Store, Koda and editor pages now send your key as a header. Commands you copied before still work.
• Settings works on a phone — every section used to be cut off on the right.
• Activity charts use your days, not UTC — an evening's work no longer lands on the wrong date.

Fixes & QOL: a deleted artifact's card in a conversation says so instead of opening an empty panel; the resources panel opens on narrow screens; What's New previews cut between lines; toasts no longer cover the Support button; the MCP Store is rebuilt; a missing conversation offers a way back; the Referrals monthly tile says what it counts.

The gateway now speaks the OpenAI Responses API at POST /v1/responses, next to Chat Completions and Anthropic Messages. Newer OpenAI SDKs, Codex CLI and the OpenAI Agents SDK use this format by default, so they now work with your OpenAdapter key out of the box.

What works: text and streaming (response.output_text.delta and friends), function tools with streamed arguments, reasoning summaries, image input on vision models, instructions, max_output_tokens, reasoning.effort and text.format for JSON output. Every model and every auto-* tier is available, with the same quota, failover and X-Auto-Model header as the other endpoints.

Not supported: previous_response_id (send the full conversation in input each call) and hosted tools such as web search or file search. Those return a clear error instead of doing something else.

Python:

from openai import OpenAI
client = OpenAI(api_key="sk-cv-...", base_url="https://api.openadapter.in/v1")
r = client.responses.create(model="GPT-6-Luna", input="hello")
print(r.output_text)

See the Responses section in the Guide for JavaScript and curl.

GPT-6-Luna is live and GPT-5.6-Luna is retired. Luna is the fast tier of OpenAI's GPT-6 line, the one OpenAI calls its most efficient model for focused, high-volume work. It carries a 1M-token context window with up to 128K output, reads text, images and files, and has a knowledge cutoff of May 2026 against February for its predecessor. Call it as GPT-6-Luna.

It is a pro-tier model at 1 request of quota per call, so it is available on every paid plan and the free tier. GPT-5.6 Luna was max-tier; this is a step down in price for the same job.

OpenAI has not published a standard benchmark table for Luna. What they do state:

We will add more figures as OpenAI publishes them rather than borrow numbers from elsewhere.

Calls to GPT-5.6-Luna now return a message pointing to GPT-6-Luna.

Sources: OpenAI announcement, model page.

Four new model names pick a model for you: auto-max, auto-pro, auto-lite and auto-free.

Each tier keeps a pool of its three to five fastest models, measured from the last 24 hours of real traffic. On the first turn of a conversation a small decisions model reads the request once and picks the model in that pool best suited to it: coding, agent loops with tools, long documents, images, or plain chat. The pick stays with the conversation, so an agent session does not hop models mid-task and prompt caching keeps working.

You always know what answered. The response's model field and an X-Auto-Model header name the model that served the request, and quota is charged at that model's rate. If the picker is unavailable, the fastest model in the pool answers. Plan rules are unchanged: auto-max needs a plan that reaches max-tier models, and auto-free works on every plan.

Koda uses auto-pro by default from its next release.

MiMo-V2.6-Pro is live. The flagship of the V2.6 series: 1.02T parameters with 42B active, a 1M-token context window with up to 128K output, and unlike V2.5 Pro it now reads text, image, video and audio. MIT licensed. Xiaomi positions it for complex projects, long-horizon agent work, security research and high-stakes tasks. Call it as MiMo-V2.6-Pro.

Xiaomi's published scores against the V2.5 Pro it replaces:

It is a reasoning model: it thinks before it answers, so give it an output budget of a few thousand tokens or short replies come back cut off. It costs 2.5 requests of quota per call and is available on every paid plan and the free tier. MiMo-V2.5-Pro stays available.

Source: MiMo-V2.6-Pro model card.

MiMo-V2.6-Flash is live. Xiaomi's efficiency checkpoint of the new V2.6 series: a 309B mixture-of-experts model with 15B active, a 1M-token context window, and text, image, video and audio input. MIT licensed. Call it as MiMo-V2.6-Flash.

Xiaomi's published scores, with the previous V2.5 Pro flagship for scale:

It costs 1.5 requests of quota per call and is available on every paid plan and the free tier, the same terms as MiMo-V2.5. It also joins the auto-pro pool as soon as it has enough traffic to be ranked.

Source: MiMo-V2.6-Flash model card.

MiMo-V2.5-Pro is live. It is Xiaomi's flagship: a 1.02T-parameter mixture-of-experts model with 42B active per token, a 1M-token context window, MIT licensed, and text only. Xiaomi built it for demanding software engineering and long-horizon agent tasks that span thousands of tool calls. Call it as MiMo-V2.5-Pro.

Xiaomi's published scores:

It costs 2.5 requests of quota per call and is available on every paid plan and the free tier. Use it for the hard, multi-step coding jobs, and MiMo-V2.5 when you need images, video or audio.

The old MiMo-V2-Flash name is retired. Requests for it now return a message pointing to MiMo-V2.5.

Source: MiMo-V2.5-Pro model card.

MiMo-V2.5 from Xiaomi is live. It is natively omnimodal: it takes text, images, video and audio in a single model, with a 1M-token context window. It is a 310B mixture-of-experts model with 15B active per token, released under the MIT license. Call it as MiMo-V2.5.

Xiaomi's published scores:

It costs 1.5 requests of quota per call and is available on every paid plan and the free tier, the same terms as the other flash vision models. Reach for it when the input is a document, chart, screen recording or audio clip rather than plain text.

Source: MiMo-V2.5 model card.

DeepSeek-V4.1-Flash is live. It reads text and images, carries a 1M-token context window, and is built for input-heavy agentic and coding work. Reasoning effort is adjustable, so it can answer quickly or think hard on request. Call it as DeepSeek-V4.1-Flash.

DeepSeek publishes these scores against the earlier V4 Flash, all at maximum reasoning effort:

It costs 2 requests of quota per call and is available on every paid plan and the free tier. DeepSeek-V4-Flash stays exactly as it is at 1 request per call, so nothing changes for existing setups. Pick V4.1 when the task is a long agent run or terminal work, and V4 when you want the cheapest fast model.

Source: DeepSeek-V4.1-Flash model card.

A batch of things that were visibly wrong.

• Inline timelines were unreadable — every event rendered as a narrow strip of text, one word per line. Diagrams also only had about half the colour palette wired up, so a chart asking for its third or fifth colour came back grey.
• Asking for a visual of a document produced a wall of CSS. The model was saving an HTML page as a new version of the existing document, which then rendered as source. Asking for the same subject in a different form now creates a new artifact instead.
• The composer settled down. No more pale box appearing around the text when the pointer crosses it, and no border lighting up the moment the page loads. The model picker sits next to the effort setting as one phrase rather than a name floating a hundred pixels from its own chevron.
• "New chat" from the chats list created a real, empty conversation every time it was clicked. It opens the compose screen now, and a conversation exists once something is said in it.
• Bring-your-own-keys is built from the same parts as the rest of settings, instead of section headings rendering as filled bands across the page.

The list gave you a title and a timestamp, so finding a conversation meant opening candidates until one of them was the right one.

Each row now shows where the conversation left off, and marks the ones that produced an artifact. Searching looks inside the messages too, not just at titles — so "the one where we worked out pricing" is a thing you can actually type. Sorting by name or by when a chat was started reorders every conversation you have, not just the ones on screen, and the list keeps loading as you scroll instead of stopping at the first two hundred. Day headings stay pinned while you scroll past them.

Deleting a conversation asks first, and tells you what goes with it.

Finding the timeline you made meant opening rows in a list of filenames until one of them was a timeline.

Every artifact is now a tile that renders the thing itself — the deck looks like the deck, a document shows its opening lines. Search by name or by the conversation it came from, sort by when it changed, when it was made, its name or its kind, and filter to just the pages or just the diagrams. Deleting one asks first, and says how many versions go with it.

Opening a tile uses the same split view a conversation does.

A page, deck or diagram used to open as a card inside the transcript, inset from every edge — something floating on the conversation rather than the other half of the screen.

It now opens as a proper split: pinned to the right edge, top to bottom, with its own bar carrying the title, the version picker, Source, Download and close. Drag the seam to make it as wide as you need — the width is remembered — or nudge it with the arrow keys. Escape closes it. The sidebar folds to icons while it is up and comes back when you close it.

Each kind of artifact also wears its own icon and colour now — a blue globe for a page, violet for a document, orange for code — so the rail, the transcript and the gallery all tell you what something is before you read its name.

The settings shell was rebuilt last week; the pages inside it had not caught up. All eight have now.

Billing leads with the plan you are on and its price, instead of burying it in a five-across grid where the plan you pay for was orphaned onto a second row. Usage replaces three full-width rulers with three readable meters and finally draws the last twenty-four hours of requests — the data was already there and nothing showed it. API keys, your own provider keys, teams and editor setup each got the same pass: content sits at a width it earns, a card's heading no longer competes with the page's, and a page is one page rather than two stacked on each other.

Removing a provider key or a team member now asks you to confirm properly, and the buttons that do those things no longer look exactly like the ones that don't.

The box you type into had a strip of buttons above it and a second row below, which made one control read as three. Everything now sits on a single line under the text: attach and mode on the left, the model and how hard it thinks on the right, and send.

The model and its effort are written as a phrase rather than boxed separately, because choosing one is choosing the other. Cited links in the right-hand panel show each site's icon, the way search results already did.

Seven dialogs had been built seven times over — four different backdrops, three corner radii, and three of them a fixed dark that ignored your theme. Fields were much the same: no two of them the same height, so a row of controls never lined up.

There is now one dialog and one set of fields behind all of them. Escape closes, the backdrop closes, focus lands in the first field and returns where it came from when you leave.

Along the way we found four toggles whose selected state had never rendered at all — a broken bit of markup meant the highlight was never applied. Those work now.

Settings had two page titles stacked on top of each other, a section list floating in a card that stopped a third of the way down, and the chat navigation still sitting beside it for a screen that has nothing to do with chat.

It is now a single full-height column of sections with the page beside it, the module navigation is out of the way, and signing out lives at the bottom of that column where the rest of your account already is.

A conversation used to be filed under the first sixty characters you typed, so a thread was called "Search the web for what a vector database is, then read o…".

The model now writes a real title — "Vector database basics" — while it is still answering, and the name changes in front of you. If it cannot think of one, the old truncation still stands.

A long answer would slide out of view as it was written: the view scrolled to where the bottom had been a moment earlier, fell a little further behind with every line, and eventually gave up following altogether.

It now keeps pace. Scroll up while an answer is being written and it stays where you put it, with a button to rejoin the bottom; scroll back down and it picks the answer up again.

Asking the assistant to search the web could come back with "Rate limit exceeded" — from us, to us. Search and page-reading were going out through the same door as model calls and queueing behind them, so an assistant that searched twice in one answer rate-limited itself.

They now run in place, with no round trip and no queue. Reading a page runs here too, which fixes the "invalid API key" error that had made it fail for everyone.

A page that hides its text behind JavaScript still cannot be read, and now says so plainly instead of returning a few words the model would have guessed from.

Occasionally a model breaks down and emits one character over and over — a wall of exclamation marks where an answer should be. It is a fault in the model's serving stack, not in anything you asked.

A run that starts repeating a single character is now stopped at four hundred of them and says what happened, instead of filling the thread with thousands and billing you for every one.

A turn keeps running when you close the tab, so that refreshing mid-answer rejoins it rather than throwing the work away. But Stop only knew how to reach a run this page had started itself — so after a refresh it quietly did nothing, and the turn carried on with no way to reach it.

Stopping no longer depends on knowing which run: it stops whatever is in flight for the conversation.

The sidebar is now two navigations doing two jobs. A narrow rail on the far left says which product you are in — Chat, API or Koda — and rests as icons until you pause on it, when it widens in place to show the names. Beside it sits that product's own navigation, which is always there.

So the pages you use daily are no longer competing with the pages you visit monthly. Skills, Knowledge and Tools belong to Chat. Keys, usage, billing and models belong to API. Sessions belong to Koda.

The second navigation collapses to icons with the [ key when you want the room, and it never disappears — that is the one you actually navigate with. The module you are in wears its own colour on the rail, the only place in the interface where a colour means something.

Every surface in the dashboard was re-cut against two apps we admire, by reading what they actually do rather than how they look — type sizes, ink values, borders, shadows, and the exact duration of every transition.

The change you will feel first is speed. One 220ms duration was doing the work of three: a colour change, a shadow and a panel opening all took the same time, and that is what made the interface feel heavy. Colour now settles in 90ms and only something that genuinely moves takes its time.

The rest is quieter. Cards sit on the page instead of dissolving into it. Secondary text is a single value that stays right on white, on a card and on a highlighted row, instead of three greys that drifted apart. Body text is a size larger and the black is a shade softer, because nothing we measured went as hard as we did.

When you create a key you now choose what it can do: Inference (chat, audio, embeddings, images) and/or Manage account (keys, billing, usage). New keys are inference-only by default — the right shape for a key you paste into an editor, because it can never rotate your keys or touch billing if it leaks.\n\nNothing changes for keys you already have: they keep exactly the access they had. Scopes show as small tags on the Keys page.

Diagrams drawn inside a chat reply were coming up a couple of pixels short and growing their own scrollbar, which turned an illustration into an attachment you had to scroll. They now measure themselves honestly and size to their content.

They also follow the theme switch straight away. Switching between light and dark used to leave any diagram already on screen in the theme it was drawn in.

A control sits next to the model picker in chat: no thinking, low, medium, or high. Medium is the default, and the setting belongs to the conversation — a thread you set to Low is still Low tomorrow.

It exists because thinking could eat the whole answer. A fast model handed a hard question would reason to the end of its output budget and return nothing at all: a "Thought" block, then silence, charged in full. The ceiling on a reply is now generous enough that there is always room left to answer, and when a turn still somehow spends everything on thinking, the reply says so instead of showing you an empty message.

Opening a live Koda session from the dashboard used to take two confirmations: matching a six-digit code in the browser, then running /remote approve in the terminal. Both are gone.

Reaching your session already required your OpenAdapter credentials — the same ones that authorize every model call Koda makes. A second and third confirmation guarded a door you had already unlocked, and the terminal step in particular defeated the point of remote: it asked you to walk back to the machine you had walked away from. Pick your session, and you are watching it.

Sessions stay end-to-end encrypted between your terminal and your browser; the relay still forwards frames it cannot read. Permission gates are untouched — a dangerous command still stops and waits for you to approve it.

Two fixes ship alongside. Refreshing the page no longer loses the conversation or the model and thinking controls: a long transcript exceeded the relay's frame limit, which quietly killed the connection every time you reattached. And a permission prompt raised while your browser was closed now waits for you instead of expiring unseen — the relay tells Koda when someone is actually watching.

Qwen3.8-Flash is live. It reads text, images and video, carries a 1M-token context window, and is the least expensive of the three vision models we added this week.

Alibaba positions it for coding assistance, agentic workflows, document and codebase analysis, chart reading and long-video analysis. Call it as Qwen3.8-Flash.

It costs 1.5 requests of quota per call and is available on every paid plan and the free tier — the same terms as GLM-5.3-Flash and DeepSeek-V4-Flash-Vision.

Alibaba has not published benchmark scores for this model yet. We will add them to the model page once they exist rather than borrowing figures from its siblings.

GLM-5.3-Flash and DeepSeek-V4-Flash-Vision are live. Both take images as well as text, and both carry a 1M-token context window.

GLM-5.3-Flash is built for coding and long-running agent work — it scores 84.3 on Terminal-Bench 2.1, within a point of the frontier models, and also accepts video input. DeepSeek-V4-Flash-Vision adds image understanding to the V4 Flash line, which posts 79% on SWE-bench Verified and 88.4% on LiveCodeBench, and is aimed at documents, charts and multimodal agent workflows.

Both are available on every paid plan and on the free tier, and cost 1.5 requests of quota per call. Call them as GLM-5.3-Flash and DeepSeek-V4-Flash-Vision.

Benchmark figures are as published by the model authors; DeepSeek's vision variant is marked experimental upstream.

Speech-to-text now works as audio arrives instead of only after a recording ends. Each sentence comes back the moment the speaker finishes it, so a voice agent can start responding before the person has stopped talking.

Open a WebSocket to wss://api.openadapter.in/v1/realtime/transcribe, send microphone audio, and read transcripts as they land. Your normal API key works — nothing to enable.

Available on parakeet, whisper-1, whisper-large-v3 and whisper-large-v3-turbo.

You're charged per transcribed sentence, not per connection or per second of audio — an open socket with nobody talking costs nothing.

Read the guide →

Qwen3.8-27B is live on Pro plans at 0.25× quota — four requests for the price of one.

It handles a 250K-token context, supports tool calling, and thinks step by step before answering (reasoning is returned separately, so you can show or hide it). Released by Alibaba on 14 August 2026.

Call it as Qwen3.8-27B from any client.

One note if you use it programmatically: thinking is on by default and consumes part of your max_tokens budget. Give it room — a few thousand tokens — or turn thinking off if you only want the answer.

Four models land on OpenAdapter today, and one is on its way out.

Hy3 and GPT-5.6 Luna — 1x quota

Both come in at 1x quota, the same cost as a standard model, which makes them the cheapest MAX-tier models on the platform. If you want frontier-class output for everyday agent loops without burning through your plan, start with these two.

Hy3 is Tencent's Hunyuan 3. Luna is the newest GPT-5.6 variant.

DeepSeek V4 Pro 0813 — 2x quota

DeepSeek re-post-trained V4 Pro. It serves under the same name you already use — DeepSeek-V4-Pro — so nothing in your config changes.

Qwen 3.8 Max — 4x quota

Alibaba's newest flagship and the largest Qwen we carry. It is the most expensive model on the platform at 4x quota, so reach for it when the task actually warrants it.

All four are MAX tier, so they need a paid plan.

GLM-4.7 is retiring

GLM-4.7 is being retired. You do not need to change anything — requests for GLM-4.7 now serve GLM-5 automatically, so existing configs, scripts and agents keep working. GLM-5 is the stronger model on every axis, and it costs the same 1x quota.

If you would rather be explicit, switch your config to GLM-5 at your convenience.

If web_search has been returning "No results found", this fixes it — but you need to re-run one command.

What was wrong

Two separate faults, which together meant search was broken for essentially everyone:

1. The search tool never reached our servers. The MCP we installed talked to a search instance on localhost:8888 — a local container almost nobody runs. So it had nothing to query.
2. Our own search fleet was being blocked. The engines behind it refuse traffic from datacenter addresses, so even a correctly-pointed request came back empty roughly eight times in ten.

What changed

Search now runs through the gateway. Your machine asks us, we handle the fleet and which engines to use. That means the next time an engine starts blocking, it is a fix on our side — not a reinstall on yours.

You need to re-run setup

The fix ships in the installer, so machines that already have the old search tool keep the old behaviour until you re-run it. From your terminal:

curl -sL "https://api.openadapter.in/api/setup/mcp-search?key=YOUR_API_KEY&tool=claude-code" | bash

Windows (PowerShell):

irm "https://api.openadapter.in/api/setup/mcp-search-ps?key=YOUR_API_KEY&tool=claude-code" | iex

Swap tool=claude-code for opencode if that is what you use. Your key is on the API Keys page. Restart your editor afterwards.

Windows users: the PowerShell installer itself was also broken and never completed. That is fixed in the same release.

Checking it worked

Ask your agent to search for something current. You should get real results rather than "No results found".

Koda is out of beta. Until today npm i -g @openadapter/koda gave you 1.0.0-beta.13 — latest had been frozen there for months while the real work happened on @beta. This release moves latest onto current code:

npm i -g @openadapter/koda

Debugging — new

Koda speaks the Debug Adapter Protocol. It can set a breakpoint, launch or attach, then read the real stack and actual variable values at the moment things broke — step through, and evaluate expressions in the paused frame.

It discovers 14 adapters on your PATH, with an xcrun fallback for the lldb-dap Xcode hides. It registers the tool only when an adapter genuinely exists on your machine, and never loads the engine until you actually debug.

This is the difference between reading a stack trace and watching the value change.

Delegation — new

A task tool with context-isolated subagents — general, explore, or your own in .koda/agents/*.md. They run in parallel and only their results come back, so the main context stays clean. A live line under the input shows how many are working and for how long.

Skills — new

koda skill search / add / remove / list / update, against the open skills.sh registry.

Editing — new

hashline_edit (vendored from oh-my-pi, MIT). Reads carry a content tag and line numbers; edits are line-anchored operations that never echo old text back. A stale tag fails closed with a re-read instruction rather than quietly corrupting the file.

Choose it in /settings → Edit tool, which enforces one or the other — running both just hands the model a coin to flip.

Type errors in the loop

After an edit, the errors your change introduced — including the call sites you broke somewhere else — come back in the tool result, so the model fixes them in the same turn. TypeScript needs no install; Python, Go, Rust, Java and C/C++ work through their standard language servers.

Koda finishes what it starts

The big one if you have ever watched an agent stop halfway.

There is now a completion gate on every model tier, not just the ones running the overseer. The old check only fired when the model had made a todo list — which happened in 13.6% of multi-step tasks, so on a cheap model nothing was checking at all. The new gate costs zero extra model calls.

Todos can also carry a verify command, and marking one done is only a claim until Koda runs that command itself and sees it exit 0.

Fullscreen mode

KODA_FULLSCREEN=1 gives a live sidebar: session, context, model, LSP status, and the todo plan marked [ ] / [x] / [✓] — that third mark meaning Koda verified it, not that the model said so.

Plus 15 themes on a four-role colour system. The old palette used accent 102 times against text's 16, so everything read as important and therefore nothing did.

Fixed

• Theme changes apply instantly instead of needing a restart
• The sidebar scrolls
• A language server's stderr can no longer shred the fullscreen frame
• A deleted working directory explains itself instead of throwing uv_cwd
• Stray ... on fullscreen rows
• Plan mode records a plan the gates can actually check
• The overseer toggle stops claiming to be on when it is suppressed
• Retried timeouts stay quiet
• A self-update rule that could have pulled beta users backwards onto latest

DeepSeek V4 Flash 0731 is a re-post-trained revision of V4 Flash, and it is now a primary backend behind DeepSeek-V4-Flash on OpenAdapter. Same model name, same quota cost. What changed is what it can do as an agent.

What it is

A sparse mixture-of-experts model: 284B total parameters, 13B active per token, tuned for coding, reasoning and agent workflows. Only a fraction of the network fires per token, so it answers at small-model speed while reasoning like a much larger one.

Independent evaluation

Measured by Artificial Analysis, who run models themselves rather than republishing vendor claims:

90.8% on GPQA Diamond is the number to look at — that is graduate-level science reasoning, from a model with 13B active parameters. The low CritPt and non-hallucination figures are equally worth knowing: it is strong at structured reasoning, weaker at research-frontier physics, and like most models it will still assert things confidently when it does not know. Do not use it as an oracle.

What the retrain changed

The weights are the same size. The post-training is not. DeepSeek's own agent-benchmark figures for this revision:

Note the shape: single-shot coding barely moved (SWE-bench flat at 79.0) while multi-step agent ability transformed. This model got better at operating, not at writing a function. Driven in a loop with tools you will feel it; on one-off questions you may not. DeepSeek claim 0731 now beats V4 Pro Preview on every agent benchmark they published — Terminal Bench 82.7 against Pro's 72.1.

Treat that second table with care: those are DeepSeek testing DeepSeek, on a harness that is not public yet, and two of the nine are internal sets the company built. The Artificial Analysis figures above and our own measurements below are the parts with independent backing.

How fast it is here

Measured through our gateway, median of three runs each:

Thirty-four thousand tokens of code answered in under five seconds is what makes it usable as an agent backend rather than just a chat model.

Capability check

We run every new model through the same suite before promoting it. This one passed all six:

Using it

Nothing to change. Point at DeepSeek-V4-Flash and you get it:

curl https://api.openadapter.in/v1/chat/completions \
  -H "Authorization: Bearer $OPENADAPTER_API_KEY" \
  -d '{"model":"DeepSeek-V4-Flash","messages":[{"role":"user","content":"Refactor this module"}]}'

The other backends stay in place as automatic failover, so if this one is ever busy your request still completes — you just will not notice.

OA-Robin-Mini-Preview is the first model we built and host ourselves, and it is live now.

Because it runs entirely on our own hardware, your prompts never leave our infrastructure and nothing is retained — the same privacy guarantee as our other Edge models.

• Private by default — great for confidential chat, brainstorming and creative work you would rather not send to a third party.
• 256K context with vision, tool calling and JSON mode, so it slots straight into agents and long-running workflows.
• 0.5x multiplier — every request costs half a normal one, so you get roughly double the usage from the same plan.

It is a preview, so expect it to keep improving. A much larger model in the same family is coming soon — we will announce it here when it lands.

Try it from the Models page, or just point your client at OA-Robin-Mini-Preview.

Share your referral link from the Referrals page. When a friend signs up and makes their first purchase, they get 20% off — applied automatically at checkout, no code to enter. You earn 20% of what they paid as credit, used automatically toward your own next plan. Track your link and referrals right from the dashboard.

Koda's first launch now connects to your account through the browser:

• Run koda — your browser opens the dashboard with a one-time code.
• Sign in (or create an account), check the code matches your terminal, and click Approve.
• Koda receives its own named API key (e.g. Koda CLI — MacBook-Pro) and finishes setup automatically — revoke any device from the Keys page.

Prefer the old way? koda setup sk-cv-… and the OPENADAPTER_API_KEY env var still work everywhere, including CI.

Browser-based, OpenAI-compatible apps — like Janitor AI, SillyTavern, and other web chat front-ends — can now call the API directly.

• Set the endpoint to https://api.openadapter.in/v1/chat/completions and use your sk-cv-… key.
• Previously these apps hit a browser "network error" because the API only accepted requests from our own domains. The /v1/* endpoints now accept any origin.

Note: only paste your key into apps you trust — a key used in the browser is visible to that app.

The limited-time free promo for MiniMax-M3 has ended. It's now a standard premium model:

• It counts toward your request quota like other top-tier models.
• It's rate-limited per your plan tier (no more special promo allowance).

Thanks to everyone who tried it during the promo. It remains available on your plan — just under normal quota now.

GLM-5.1 now advertises and serves a 128K context window — up from the previous 32K cap.

• Large codebases, long documents, and extended chats no longer hit the old limit.
• Requests are automatically routed to a backend that can handle the full window; very large prompts simply skip any route that can't.

No action needed — your existing calls just get more headroom.

The GO plan now includes the latest 0G Network models — DeepSeek-V4 (Pro & Flash), Qwen3.7-max, 0GM-1.0, and MiniMax-M3 — alongside the existing 0G lineup.

These models also group correctly under 0G Network on the Models page instead of slipping into "Other."