Skip to main content
by artokun · July 20, 2026 · Codex · GPT-5.6 · provider tracking Back in June, we wrote that the Agent Panel lets you pick a provider, not a port: the same canvas-driving agent runs on either Claude or ChatGPT, on your own subscription, no API keys. The promise implicit in that design is that when a provider ships a new frontier family, you shouldn’t have to do anything — the same panel just offers the new models. This week that promise got tested. OpenAI shipped the GPT-5.6 family — Sol (the flagship), Terra (the balanced middle), and Luna (fast, and the cheapest of the three) — and as of comfyui-mcp v0.41.0, all three are in the panel’s model picker via the Codex integration. Same chip, same Connect button, same subscription billing. Pick GPT-5.6 Sol and ask it to build you a WAN video graph. Getting there was not just “add three strings to a list.” The interesting parts are a new kind of reasoning effort, a catalog that has to be honest about what your account can run, and a dependency pin that only exists because a user filed a great bug report. That’s what this post is about.

​
Ultra is not a bigger dial

The Codex effort scale used to top out at xhigh. GPT-5.6 extends it with two new levels: max, available to all 5.6 users in Codex, and ultra — and ultra is worth pausing on, because it is not just a longer thinking budget. Per OpenAI’s GA post, an ultra-effort turn coordinates four agents in parallel by default: the model fans the problem out, works branches concurrently, and merges. For the kind of request the panel gets — “audit this 60-node graph and fix whatever’s mis-wired” — that’s a materially different execution mode, not a notch on the same dial. The panel’s effort dropdown honors per-model ceilings, and those ceilings come from the provider’s live model/list, not from our assumptions: Sol and Terra accept effort through ultra; Luna tops out at max. The dropdown you see is the intersection of the panel’s scale and what the selected model actually supports — pick Luna and ultra simply isn’t offered, rather than being offered and rejected at turn time. One pleasant simplification fell out of this: the panel’s cross-provider effort mapping used to downmap Claude’s max to Codex’s xhigh when you switched providers, because Codex had no max. Now it does — max maps to max, natively.

​
The catalog is live, and it’s about your account

The bigger change in 0.41.0 is philosophical: the model picker stopped being a hardcoded list and became a live, account-aware catalog. The Codex backend asks the codex app-server for model/list on connect, and that RPC is account-aware — a given ChatGPT plan only sees the models that plan can actually run. We learned to want this the hard way: the old static list once advertised a model a ChatGPT-plan account couldn’t use, and picking it produced a 400 on every turn. A picker that offers you something unrunnable is worse than a shorter picker. With GPT-5.6 out, the live catalog gained a policy layer: pre-5.6 models are deprecated, and the picker hides them whenever your account has the 5.6 family. If your account hasn’t been migrated yet and has no 5.6 models, you keep your full existing catalog — the panel never bricks an account that lags behind the provider’s rollout.

​
Hidden is not unrunnable

There’s a design decision buried in that filter worth making explicit. Hiding a deprecated model from the picker and forcing you off it are two different things, and 0.41.0 deliberately does only the first. Internally the backend keeps two views of the catalog. The full, unfiltered account catalog answers the question “can this account legally run model X?” — and that’s what every turn’s model resolution clamps against. The picker gets the deprecation-filtered view. So if you have a tab pinned to a pre-5.6 model that your account can still run, it keeps running on that model. You won’t find it in the dropdown for new picks, but your existing choice is respected until the provider itself pulls the model. Hiding a model from new selections must never force-switch a still-runnable session mid-conversation — that’s the invariant. The one place we do switch you automatically is when there’s no legal alternative: the ChatGPT-direct backend’s default model moved from gpt-5.4-mini — now retired by OpenAI, not merely deprecated — to gpt-5.6-luna, its natural successor on the fast-and-cheap end. Saved tabs that had the old default stored migrate automatically; without that, a tab would have kept its retired model selected and 400’d every turn while looking perfectly configured.

​
The pin: a compatibility war story

Now the part where a fast-moving provider bites back. comfyui-mcp bundles the Codex SDK (@openai/codex) as an optional dependency so the ChatGPT backend works with zero extra installs. For 0.41.0 that bundle is pinned to 0.145.0-alpha.24 — an alpha, pinned exactly, on purpose. The stable release at the time had a crash with a very specific shape: if Codex Desktop 0.145 had already written its model cache to disk, the stable SDK would crash while renewing that cache, because the newer cache entries carry a field — supports_reasoning_summaries — that the stable release didn’t know and refused to tolerate. So the failure mode wasn’t “new feature missing”; it was “having OpenAI’s own desktop app installed breaks the panel’s backend.” Two official components of the same provider, one version apart, sharing state on disk and disagreeing about its schema. We would likely have chased this for days if not for the field report in GitHub issue #241, which was the kind of bug report maintainers dream about: the exact crash, the exact Codex Desktop version, the observation that it only happened on machines where Desktop had touched the cache, and the verification that the alpha pairing worked. The pin that shipped is the field-verified pairing, not a guess. Thank you — reports like that are how a small project keeps up with a provider that ships weekly. The general lesson is the real content here. Tracking a frontier provider means living with three moving parts — the models, the SDK, and the provider’s other software on the same machine — that don’t version in lockstep. The answers that held up: pin exactly what you verified, derive capabilities from live endpoints instead of hardcoding them, and always leave a path for accounts and binaries that are one step behind (the static fallback catalog still exists for older CLIs that can’t serve model/list at all — and since such a binary can’t run any 5.6 model, that fallback keeps a runnable pre-5.6 escape hatch).

​
What you do with it

Nothing new — that’s the point. Sign in with codex login, open the Agent tab, pick the ChatGPT chip, and the picker now shows Sol, Terra, and Luna with the effort ceilings each one really has. The same panel_* live-canvas tools, the same model-family knowledge, the same cost guardrail, the same Ctrl+Z — now with a flagship that can fan out four agents on your gnarliest graph, and a fast default that’s the cheapest way yet to have an agent wire your nodes. You picked a provider; the port kept up.
Drive ComfyUI from an autonomous agent on your own Claude or ChatGPT subscription: install comfyui-mcp and add the Agent Panel — see Backends / providers for the full capability matrix. Star the repo or file an idea at artokun/comfyui-mcp.