Critique Auto Gets a New Model Roster

Model catalogs age in a surprisingly specific way. The router still works, but the names around it start telling the wrong story: a “3.7” label that should be “3.8,” a former “Flash” route that has become a dated alias, or a transport suffix that leaks into a product selector where a person only needs a clear name. That drift is costly because model choice is part of the product contract. It affects latency expectations, the amount of context a turn can accept, token pricing, benchmark interpretation, and the fallback a coding agent reaches when its first plan does not survive contact with a repository.

This update brings the current model names into Critique Auto, the direct Inference API offer list, the hosted CritiqueCode selector, Remedy, the model catalog, and Pricing. Internal route identifiers remain implementation details. The UI shows the model name, provider mark, access level, and the relevant credit or token-rate context. Existing stored aliases still resolve forward, so a policy that references a previous release does not become a dead configuration overnight.

What changed in Critique Auto

Auto is a session policy, not a random model carousel. On the first turn, Critique reads the request shape: tool use, editing, completion, long context, vision, video, reasoning depth, and the expected agent horizon. It ranks models that can actually accept that shape. Once a model is selected, the session keeps it sticky while the loop is healthy. Stickiness gives the agent a stable behavioral surface and preserves the practical value of context and cached prefixes.

The refresh moves the Auto portfolio forward without changing that policy. Efficient now includes Kimi K2.7 Code, MiniMax M3, KAT-Coder Pro, GLM-5.3 Flash, Mercury 2.5, Hy4, and GPT-5.6 Luna, plus the existing Flash and specialist set. Frontier is GPT-5.6 Sol, Claude Opus 5, Kimi K3, Muse Spark 1.3, GLM-5.3, Qwen3.8 2.4T, Grok 4.6, and Fugu Ultra. Trinity and Kimi K2.6 are out of Auto. Muse Spark stays on 1.3.

The important behavior is the boundary around those names. Auto chooses from a fixed coding portfolio: listed Inference specialists plus a few catalog ids scored for SWE work. It does not silently pull an arbitrary model from the managed review catalog. That is intentional: an Auto request should have a known route, a known billing surface, and an explainable fallback list. If you want a specific managed model outside the rungs, pin it from the Models shelf. If you want Auto, let the session policy do the selection and inspect the routed model in the response metadata.

The refreshed model shelf, in plain English

There is a difference between an Auto candidate and a catalog listing. A candidate is scored against a task vector and is eligible for a specific rung. A catalog model is a product offer with a credit floor, role support, and a published display identity. The two sets overlap — Sol, Opus 5, and Luna are on Auto because they scored for those rungs — but Auto is still not the whole shelf. If you want a model outside the rungs, pin it.

Current token pricing for the refreshed routes

The table below is a directory snapshot of provider-reported input and output rates, normalized to dollars per one million tokens. It is here to make model selection legible, not to imply that a token rate predicts successful repository work. Output is often the expensive side of an agent loop; context length is not the same thing as useful context; and a low per-token price can still lose if the model needs repeated repairs. The managed credit floor shown elsewhere in Critique is the product rate for a review run, while these numbers are the underlying token-rate reference for direct model calls.

One common benchmark is useful — if we keep it honest

A model catalog can contain dozens of benchmark claims and still have no clean comparison. The answer is not to mash every score into one leaderboard. For this roster, the cleanest shared coding benchmark is SWE-bench Verified: real GitHub issues, a recognizable unit, and enough overlap across the catalog to show a useful shape. The chart includes only models for which the catalog has a published SWE-bench Verified reading. It intentionally leaves out newer models whose available public scores use another suite, a different harness, or no comparable score at all.

The lower-left is where Auto earns its keep. DeepSeek V4 Flash 0731 is a good example of the shape we want from an efficient rung: a sub-credit floor in the managed catalog, a strong shared benchmark reading, and a context window large enough for realistic repository prompts. The upper-right is still important. Claude Sonnet 5 and Qwen3.8 Max are not “bad values” because they cost more; they are deliberate escalation or pinned-choice models when the expected cost of another failed loop is higher than the token bill. The point is to make that trade visible before the run starts.

Notice what the chart does not say. It does not rank Gemini 3.8 Flash below GLM-5.3 because one has a SWE-bench Pro row and the other has a Terminal-Bench row. It does not convert Kimi K3’s Terminal-Bench 2.1 score into a fake SWE-bench value. A benchmark visual is only useful when its axes share a definition. Missing data is a meaningful state, not an invitation to interpolate.

CritiqueCode app: All In cloud sessions and models

CritiqueCode is Critique’s author harness: an interactive coding agent that implements in a session, then enters the existing review and verified-repair machinery. All In is the hosted version of that harness with the cloud machine and model inference in one monthly plan. It gives the author an isolated worker, a repository branch, a live session, and a model selector whose routes match the published All In catalog.

All In tracks cloud sessions and inference dollars measured at provider cost. All In $10 includes 50 sessions and $20 of inference; All In $30 includes 100 sessions and $60 of inference. Founders Edition includes unlimited cloud sessions and $400 of inference for $150. On capped tiers, whichever balance reaches zero first pauses hosted use until the plan is upgraded or the monthly balance resets. The Dashboard shows both balances and the next renewal.

All In does not auto-charge past its plan. When a balance is empty, upgrade the subscription or wait for the monthly reset. Standalone Inference API credit remains a separate balance for local CritiqueCode and direct API calls; it does not refill the All In cloud-session allowance.

How to choose between Auto, a pinned route, and credits

For a normal change request, start with Auto and let the session stay coherent. For a benchmark or a regression investigation, pin a route so the comparison has a stable independent variable. For large review programs, use the managed model shelf and treat credits as a budget you can reason about per run. For hosted CritiqueCode, choose an All In tier based on the number of cloud workspaces you expect and the provider cost of the models you intend to use.

What this update means for model names and old configurations

The UI is intentionally name-first. “GLM-5.3 Flash” is useful to a person; a provider-prefixed route with a suffix is useful to a transport layer. Both can exist, but they should not compete for the same visual space. LobeHub provider marks now cover the refreshed catalog in the hosted selector, model pages, legacy Remedy surfaces, and the article components. Sakana has a small matching local mark because that provider is not present in the icon package; it follows the same sizing and accessibility contract as the other marks.

Old IDs remain aliases where a forward mapping is safe: Fable 5 resolves to Fable 5.1, Gemini 3.6 and 3.7 resolve to Gemini 3.8 Flash, Grok 4.3 and 4.5 resolve to Grok 4.6, GLM 5.1 and 5.2 resolve to GLM-5.3, and the previous DeepSeek Flash names resolve to V4 Flash 0731. The removed base MiMo v2.5 is not advertised as a current route; MiMo v2.5 Pro remains available. Aliasing protects stored policies without allowing stale names to leak back into the current product shelf.

A note on pricing, benchmarks, and responsibility

A price page answers “what will this route cost under this product’s billing rule?” A provider directory answers “what token rate was reported for this model at the time of the snapshot?” A benchmark card answers “how did this model score in this suite, under this harness, on this date?” Those are three different facts. We show them together because engineers need the comparison, but we do not collapse them into a single quality score. The most responsible choice is usually the model that completes the task with the fewest failed loops at a cost your team can defend.