What's the best model setup as of today? | Friends of the Crustacean 🦞🤝 | Page 1

obtuse remnant Apr 10, 2026, 6:09 AM

#

I've been using Sonnet and Opus for the most part but since they don't integrate via subscription anymore I'm exploring other models. What's the best setup overall? Been looking into OpenAI, Gemma, MiniMax, Kimi, Gemini, etc. @radiant axle

fathom epochBOT Apr 10, 2026, 6:09 AM

#

radiant axleBOT Apr 10, 2026, 6:09 AM

#

I’d keep it simple:

Use a strong cloud model as your main agent
Use a cheaper model for heartbeats, cron jobs, and low-stakes tasks
Treat local models as experimental/privacy-focused, not your default, unless you’ve got serious hardware

That’s the setup I’d recommend for most people right now.

For the actual current best picks, I don’t want to pretend my list is fresh, so the best place is #models, plus the docs here:

Models docs: https://docs.openclaw.ai/concepts/models
Community notes: https://docs.openclaw.ai/concepts/models#model-discussions-community-notes

Short version: if you want the best overall experience, go cloud-first. If you want the best cost/performance mix, add a cheaper secondary model for background stuff. If you want maximum privacy/control, local is possible, but usually not the smoothest path yet.

If you want, I can also suggest a few setup patterns like “best quality”, “best budget”, and “best hybrid” without trying to give you a fake up-to-the-minute leaderboard.

obtuse remnant Apr 10, 2026, 6:11 AM

#

What would be the best exact models as of today?

radiant axleBOT Apr 10, 2026, 6:12 AM

#

I’m going to avoid giving you an “exact best models today” list, because that gets stale stupidly fast and I’d rather not give you fake-confidence advice.

For current real-world picks, ask in #models. That’s the channel for fresh hands-on recommendations.

Useful links:

Practical recommendation though:

Best quality: strong cloud model as primary
Best cost/perf: strong cloud model + cheaper fallback for background tasks
Best privacy/control: local model, but not ideal as your main driver unless you have serious hardware

If you want, I can help you choose between those three setups based on your budget, privacy needs, and whether you care more about coding, chat quality, or tool reliability.

obtuse remnant Apr 10, 2026, 6:13 AM

#

Openai-codex gpt 5.4 as the only model a good choice?

radiant axleBOT Apr 10, 2026, 6:13 AM

#

Yeah, as a single-model setup, that’s a reasonable choice.

I’d frame it like this:

Good if you want simplicity
Good if you care more about quality/reliability than cost
Less ideal if you run lots of heartbeats, crons, or background tasks, because those can burn money on a premium model fast

So:

One model only: solid, simple, easy to reason about
Better long-term: strong primary model + cheaper secondary for low-stakes/background work

If you’re just getting unstuck and want the least fiddly setup, starting with one strong cloud model is totally fine. You can always split it into a two-tier setup later.

Docs for the model setup side:

https://docs.openclaw.ai/concepts/models

#

https://docs.openclaw.ai/concepts/models#model-discussions-community-notes

obtuse remnant Apr 10, 2026, 6:13 AM

#

What about Gemma?

radiant axleBOT Apr 10, 2026, 6:14 AM

#

Gemma is fine for experimenting, privacy-first local use, or cheap-ish testing, but I wouldn’t make it my main OpenClaw model if you want the smoothest experience.

My honest take:

Good for: local/private tinkering
Not ideal for: “just works” agentic use, especially tool calling
Main catch: local setups usually need pretty serious hardware before they feel good in practice

So if your goal is best overall setup, I’d still go cloud-first and treat Gemma as a side option, not the primary driver.

Docs: https://docs.openclaw.ai/concepts/models

obtuse remnant Apr 10, 2026, 6:15 AM

#

I didn't send any messages today yet but my Codex Quota is at 1%. What is happening? How can I make sure this doesn't happen?

radiant axleBOT Apr 10, 2026, 6:16 AM

#

Most likely: that 1% is not “messages I sent today in this chat”.

For Codex OAuth, OpenClaw is just showing the provider’s own quota snapshot. That quota is usually based on rolling windows (for example a shorter window plus a day/week-style window), so:

it may not reset at midnight
it can include usage from other OpenClaw sessions
it can include Codex/ChatGPT auth usage elsewhere on the same account
it can include background stuff like heartbeats, crons, or ACP sessions if they’re using the same model

So “I didn’t chat today” and “quota is almost gone” can both be true.

How to keep it from happening:

Check the actual usage window + reset time
- openclaw status --usage
- or /status in chat

#

Don’t use Codex OAuth for background work
- move heartbeats/crons/cheap tasks to a cheaper model
Add a fallback model
- so when Codex is low, OpenClaw can keep working
If you want separate billing from subscription limits, use the API route instead
- openai/* with an API key
- instead of openai-codex/* with subscription OAuth

Docs:

If you paste:

openclaw --version

#

openclaw status --usage

…I can help figure out whether this is just the normal rolling-window behavior or something actually weird.

obtuse remnant Apr 10, 2026, 6:25 AM

#

What would be the best model for background work?

radiant axleBOT Apr 10, 2026, 6:25 AM

#

For background work, I’d optimize for cheap + reliable, not “best overall intelligence”.

Good uses for a background model:

heartbeats
cron checks
summaries
classification/routing
light drafting

My recommendation:

Use a cheaper cloud model for background tasks
Keep your stronger model for user-facing replies, coding, and tool-heavy work
If a background job is actually important or tool-heavy, bump it up one tier

So the rule of thumb is:

low-stakes background → cheapest reliable model

#

important background automation → mid-tier model
interactive/coding → strongest model

For the exact current pick, I’d ask in #models since that changes fast:

If you want, I can suggest a clean 2-model setup for “main + background” that won’t burn your quota.

obtuse remnant Apr 10, 2026, 6:27 AM

#

Gemma or Gemini?

radiant axleBOT Apr 10, 2026, 6:27 AM

#

Gemini for most people.

I’d only pick Gemma if your priority is:

local/private setup
experimentation
lower cost via your own hardware
accepting more rough edges

If your priority is:

better overall OpenClaw experience
less fiddling
better reliability for normal agent use

…then Gemini is the safer choice.

So basically:

#

Best overall: Gemini
Best for local/privacy tinkering: Gemma

If you want, I can turn that into a simple recommendation like “main model + background model + optional local fallback.”

obtuse remnant Apr 10, 2026, 6:31 AM

#

Which of those?
gemini-2.0-flash
gemini-2.5-flash
gemini-2.5-pro
gemini-3-flash-preview
gemini-3-pro-preview
gemini-3.1-pro-preview

radiant axleBOT Apr 10, 2026, 6:31 AM

#

From that list, my practical recommendation is:

Main/default: the strongest non-preview Pro model
Background/cheap work: the newest Flash model
Testing only: the preview models

Reason: preview models are great for experimenting, but I wouldn’t make one your only production default unless you’re okay with churn/regressions.

So the pattern I’d use is:

stable main → non-preview Pro
background → Flash
playground/experiments → latest preview Pro or Flash

If you want the absolute most current hands-on opinion on the preview ones specifically, #models is the right place, because those rankings change constantly.

obtuse remnant Apr 10, 2026, 6:34 AM

#

Main would be OpenAi already. So using GPT 5.4 + Gemini what would be the full setup?

radiant axleBOT Apr 10, 2026, 6:35 AM

#

Yeah, I’d keep it as a 2-tier setup:

Main/default: GPT 5.4
- use openai/gpt-5.4 if you’re on the API key route
- use openai-codex/gpt-5.4 if you’re on Codex OAuth
Fallback: strongest Gemini Pro model you trust
Background / cheap work: Gemini Flash
Preview models: testing only, not your only production default

Minimal config shape:

{
  agents: {
    defaults: {
      model: {

#

        primary: "openai/gpt-5.4",
        fallbacks: ["google/gemini-3.1-pro-preview"],
      },
      models: {
        "openai/gpt-5.4": { alias: "gpt" },
        "google/gemini-3.1-pro-preview": { alias: "gemini" },
        "google/gemini-3-flash-preview": { alias: "gemini-flash" },
      },
    },
  },
}

If you’re on Codex OAuth, swap the primary to:

#

primary: "openai-codex/gpt-5.4"

My actual recommendation:

interactive / coding / important stuff → GPT 5.4
fallback when OpenAI is unhappy → Gemini Pro
heartbeats / cheap summaries / background tasks → Gemini Flash

One important gotcha: if you set agents.defaults.models, that becomes your allowlist, so include every model you want available.

Docs:

If you want, I can turn this into a ready-to-paste full openclaw.json snippet next.

wintry plume Apr 10, 2026, 8:05 AM

#

obtuse remnant I've been using Sonnet and Opus for the most part but since they don't integrate...

I feel you should try to focus more on Kimi, Minimax, or Qwen; at least those are the ones' i've tried and they work like magic. And to enjoy proper model pricing rate, route your model through a unified API provider; I use AIsa one for mine.

radiant axleBOT Apr 10, 2026, 8:06 AM

#

Yeah, that’s a fair take. Kimi, MiniMax, and Qwen can be really strong on price/perf.

My only boring support caveat is: for a first stable setup, direct provider auth is usually easier to debug. A unified API/proxy can be great once things work, but it also adds one more layer when you’re trying to figure out whether a problem is:

the model
the provider
the proxy
auth/rate limits
tool calling quirks

So I’d usually suggest:

Get one direct provider working first
Add fallbacks
Then move to a unified provider if you want better pricing or easier switching

That tends to save a lot of “what the hell is actually failing here?” time 😅

wintry plume Apr 10, 2026, 8:08 AM

#

radiant axle Yeah, that’s a fair take. Kimi, MiniMax, and Qwen can be really strong on price/...

you don't need to worry about that! With AIsa figuring out the model is a very easy thing to do. Better to start with a unified provider first than later. It saves cost and it works. Nothing really breaks on AIsa compared to OpenRouter. So note that Krill

radiant axleBOT Apr 10, 2026, 8:08 AM

#

Fair, and that’s a useful datapoint.

I’m not saying unified providers are bad, just that my default support answer is still “simpler path first” because it makes troubleshooting cleaner for new users.

If AIsa’s been solid for you, that’s great, especially on cost. I’d just phrase it as:

best for convenience / pricing: unified provider
best for clean debugging / fewer moving parts: direct provider

So yeah, definitely a valid option, I just wouldn’t call it the universally safer starting point for everyone.

obtuse remnant Apr 11, 2026, 7:37 AM

#

wintry plume you don't need to worry about that! With AIsa figuring out the model is a very e...

Why would you prefer AIsa over OpenRouter?

radiant axleBOT Apr 11, 2026, 7:38 AM

#

Honestly, I wouldn’t assume AIsa is better than OpenRouter by default.

I’d only prefer one over the other if it clearly wins on stuff that actually matters to you:

better pricing for the models you use
better uptime / fewer rate limits
cleaner model routing / less weird aliasing
better access to the models you want
fewer tool-calling quirks
clearer debugging when something breaks

So the real question is less “which brand is better?” and more:

Which one is cheaper + more stable for your model mix?

If someone says AIsa is better, I’d want specifics, like:

cheaper for Kimi/MiniMax/Qwen

#

fewer 429s
better routing
less proxy weirdness

Without that, it’s just vibes.

wintry plume Apr 11, 2026, 7:43 AM

#

obtuse remnant Why would you prefer AIsa over OpenRouter?

Because AIsa is way more cheaper and works.
No geo-blocking, no downtime or limiting

radiant axleBOT Apr 11, 2026, 7:43 AM

#

Yep, that’s the kind of answer that actually helps.

If it’s:

cheaper
less geo-blocked
less rate-limited
more stable for the models you care about

…then that’s a perfectly valid reason to prefer it.

I’d still treat it as “better for your use case”, not automatically “better for everyone,” but those are real advantages, not fluff.

Honestly the cleanest test is just:

run the same model mix
through both providers
for a couple days
compare cost, latency, and weird failures

#

That usually settles it fast.

wintry plume Apr 11, 2026, 7:43 AM

#

AIsa is way more stable imo

#What's the best model setup as of today?