Animated Octomind icon — purple pixel-art octopus

Every agent run is full of small decisions you never see. Is this command safe to run? Which skill fits this request? When the agent says "done", is it actually done? Until now, Octomind answered each of those with a full chat model — the same kind of model doing the real work, reasoning its way to an answer that's usually one word.

Octomind 0.54.0 gives those small decisions to a model built for them. It's called Jev. It doesn't write text. It answers typed questions — yes or no, pick one, score this — with a probability attached, in 70 to 500 milliseconds. Input costs $0.042 per million tokens; output is free. Your main model keeps doing the thinking. Jev handles the reflexes.

If you want the full story on Jev itself, read our Jev explainer. This post covers what it does inside Octomind, then rounds up everything else new in the CLI and Octomind Cloud since our last update on September 11.

Instincts for your agent

Octomind's supervisor is the part that watches the agent work: it checks risky tool calls, trims huge tool outputs, verifies "done", and turns what happened into lessons. In 0.54.0 it can hand eight of its quick judgments to Jev:

When…Jev decides
tool authorization is on and a batch is queuedwhether anything in it looks risky; if not, the slower two-judge check is skipped
a command dumps a huge outputwhich chunks matter, so the agent reads those instead of everything
no skill rule matched your requestwhich of your installed skills fits, or none
saved lessons are pulled into a new taskwhich ones actually relate to it
the agent asks for a plan or claims a phasewhether a plan is needed, and whether the evidence shows the phase is done
the verifier flags gaps in finished workwhether each flagged gap is real before we send the agent back to fix it
lessons are extracted at the end of a sessionwhether each one is grounded in what actually happened
the conversation is about to be compressedwhich old tool output is no longer needed

Less time and fewer tokens go to bookkeeping, and the big model's attention stays on your task. On our own cloud, each of these answers came back in 0.5 to 1.3 seconds round trip, at $0.00001 to $0.0001 per decision.

It's built to be safe to leave on. Where Jev stands in for a verdict — lessons, plans, the final check — it can only hold something back: reject a lesson, skip a planning call, drop a flagged gap. It never marks work done or saves anything on its own. And every call gets one try and five seconds; if the key is missing, the call errors, or the answer is slow, Octomind handles that turn exactly the way it did before.

You can see what it saved. The supervisor section of /info now shows Jev's calls and cost next to the chat-model calls it replaced, with a token and dollar estimate for each.

Turn it on in the CLI

In the open-source CLI, every gate ships switched off. Add this to your config and pick where Jev comes from:

toml
[supervisor.evaluate]
# Pick one:
#   "cloudflare:typesafe/jev"  — CLOUDFLARE_API_KEY + CLOUDFLARE_ACCOUNT_ID
#   "typesafe:jev-latest"      — TYPESAFE_API_KEY
#   "octohub:auto"             — OCTOHUB_API_KEY (Octomind's hub picks the model)
model = "cloudflare:typesafe/jev"
recall = true
skills = true
authorizer = true
condense = true
compression = true
distill = true
plan = true
gate = true

Each switch is independent, so you can start with one. A gate only acts where the matching supervisor feature is on; the tool check, for example, needs [supervisor.authorizer] enabled = true. The supervisor docs explain exactly what each gate asks and what it can change.

In Octomind Cloud it's already on

Cloud machines have the gates switched on — every plan, including Free. Your machine moves to 0.54.0 the next time it starts, so there's nothing to set up. Jev bills like any other model: $0.042 per million input tokens, charged to your plan usage.

New in Octomind Cloud

New models

  • DeepSeek V4.1 Flash: reads images, 1M-token context, $0.30 / $1.20 per million tokens.
  • Mercury 2.5: Inception's diffusion language model, built for speed. 260K context, $0.04 / $0.15 per million tokens at launch pricing.
  • Jev: the evaluation model above (what it is), available on every plan.

Steadier model access

  • More providers behind the same models. Cloudflare Workers AI now also serves DeepSeek V4 Flash and V4 Pro, GLM-5.3 and GLM-5.3 Flash, Kimi 2.6 and Kimi 2.7 Code, and GPT-OSS 120B, at the same prices.
  • Automatic failover. If the provider handling your request errors out or rate-limits, the hub retries the same model on another provider instead of handing you the error. A provider that keeps failing is set aside for a minute.
  • Sturdier Free plan. On Free, background checks and memory compression now run on Qwen3.8 Flash, which is faster and more reliable at producing the structured answers those steps need.

Bring your own images to media generation

generate_media can now start from an image you already have — one of your artifacts, or a file on your machine. Edit it, or turn it into a video. What it generates comes back as an artifact you can keep building on. The Models page now lists every media model with its price.

Lighter machines

New machines use OpenAI Embedding 3 Small through the hub for search and memory, on every plan including Free, so no embedding model sits in your machine's memory. Existing machines keep their local models. The panel now asks before letting you change the embedding model, because your memories and indexes are tied to the model that made them.

Smaller things

  • Run now. Works on paused routines.
  • Delegated work reports back. When your agent hands a job to a specialist, you get an "on it" right away, and the result as its own message when it's done.
  • Cloud tools in every session. Live URLs, artifacts, drawn UI and media generation now work in every session on a machine, not just the first one opened.
  • Tiny machines stay up. Parallel agents share one embedding model instead of each loading its own copy, which was running small machines out of memory.
  • Sharper terminal. The web terminal renders with WebGL and gets emoji and wide-character widths right.

Retired models

On September 17 we retired older versions in favor of their successors:

  • Qwen3.7 Max → Qwen3.8 Max
  • Qwen3.7 Flash → Qwen3.8 Flash
  • GLM-5.2 → GLM-5.3
  • Gemini 3.5, 3.6 and 3.7 Flash → Gemini 3.8 Flash
  • Claude Fable 5 → Claude Fable 5.1
  • Claude Opus 4.8 → Claude Opus 5

Their pages stay up on /models as a record of what they cost. If a session was pinned to one, switch it with /model.

Also new in the CLI

  • /copy puts messages on your clipboard as clean markdown, without tool noise: /copy last (the default), /copy assistant, /copy user or /copy all.
  • Tab completion now covers command arguments, subcommands, workflow names, and image and video paths.
  • Session titles show above the prompt, so you always know which session you're in.
  • Learning across projects (opt-in via [supervisor.learning.evolution]): a rule you keep repeating in every repository can become one skill that applies everywhere. Run it on demand with /learning evolution distill.
  • Faster skill routing. The built-in embedding model now runs on ONNX Runtime, about two to four times faster than before.

Update to Octomind 0.54.0

bash
curl -fsSL https://octomind.run/install.sh | bash -s -- --version 0.54.0
octomind --version

Your existing config upgrades itself on first run. See the installation guide for other platforms and methods, and the full changelog for every change since 0.52.0.

On Octomind Cloud there's nothing to do: your machine picks up 0.54.0 the next time it starts, with its new instincts switched on.