Opus 5 Arrives, and Sonnet 5 Was Never That Expensive

August 16, 2026 · InfraScout Team

The model catalog you pick from when configuring an AI provider had drifted behind what the providers actually offer. This release brings it current, retires the picks that no longer earn their place, and corrects a cost estimate that had been quietly overstating your Sonnet 5 spend by half.

New in the catalog

Claude Opus 5 is now selectable — the current Opus flagship, at the same list price as Opus 4.8 and strictly stronger. GPT-5.6 arrives in three tiers: Sol at the top, Terra in the middle, and Luna as the fast, inexpensive option. Each carries its own pricing, so a cost estimate reflects the tier you actually chose rather than the family's headline rate.

New connections now default to Claude Sonnet 5 rather than Sonnet 4.6.

What was retired, and what deliberately stayed

Opus 4.7, Opus 4.6 and Sonnet 4.6 left the picker, along with the GPT-5.4 family, GPT-5 and GPT-5 mini. In every case the replacement costs the same or less and performs better, so keeping them offered a choice with no upside.

Two survivors are deliberate:

• Opus 4.8 stays. It is the documented fallback when Opus 5's stricter security classifiers decline a request — and InfraScout conversations are, by definition, security work. Having a tier to fall back to matters more here than in general use.
• GPT-5 nano stays as the absolute budget floor.

Retiring a model from the picker does not touch a connection already using it. Existing connections keep working and keep costing exactly what they always did. What is new is that you can now see them: the AI Providers view carries a banner counting connections on superseded models, and each affected card shows an Outdated model badge whose tooltip names the recommended replacement. Opus 4.8 and GPT-5 nano are never flagged, since staying on them is the right call.

A cost estimate that was fifty percent too high

This is the part worth reading carefully.

Sonnet 5 launched at $2 / $10 per million input and output tokens, described as an introductory rate through the end of August 2026. InfraScout's cost estimates deliberately carried the higher $3 / $15 figure so that estimates would stay correct once the introductory window closed.

That window is not closing. Anthropic has confirmed $2 / $10 as the standard price and cancelled the increase. So the premise was void, and every component of the estimate was exactly 1.5× too high.

The practical consequence is bigger than the arithmetic. It inverted the comparison between models. On corrected numbers Sonnet 5 is roughly seventeen percent cheaper per message than Sonnet 4.6 — not more expensive, as the portal had been reporting. Sonnet 5 does carry more tokens per message, but that is a tokenizer change rather than more work being done, and the rate cut more than covers it.

Past conversations keep their old figures

A conversation's cost is recorded when the message is written, so this corrects future estimates only. Conversations you ran before this release still display the inflated number. Repricing historical records is a separate, deliberate decision rather than something to do quietly in the background — so if you are comparing spend across the release boundary, the older Sonnet 5 figures need dividing by 1.5 to be comparable.

Reasoning effort defaults were realigned at the same time. Fable 5 and Opus 4.8 now start at high rather than extra-high, matching each model's own documented guidance; Opus 5 and Opus 4.7 keep extra-high, which is the recommended setting for the heavy tool use that assessment work involves.

Conversations that cost less to run

Several changes reduce what a long assessment actually spends.

The AI used to be told about a context cleanup that had not happened. A detector was comparing the wrong two numbers and flagging a cleanup on essentially every request with a warm cache — hundreds of phantom events against a couple of hundred real ones. The model reacted to what looked like vanishing context by writing long "essential context to preserve" recaps into the conversation and re-running work it had already completed. The detector is gone, replaced by the exact record the provider reports, and the AI now receives explicit guidance that cleanup is platform-managed and recallable, so the right response is to recall rather than to hoard or repeat.

Cleanup thresholds are now model-aware. A model with a million-token window no longer clears at a threshold sized for a much smaller one, and an unrecognized model falls back to the conservative setting rather than the optimistic one.

Two more mechanisms were costing real money. The tenant's memory index sat ahead of the conversation history and was re-rendered whenever the AI wrote a memory file, which invalidated the cached prefix and re-billed the entire history. It is now held steady for the life of a conversation. And the agent tools a conversation reaches for first — listing agents, reading a host summary, running a command, querying an event log — are now loaded up front instead of being discovered through a search, so a pivot from cloud data to a host takes one step rather than several expensive ones.

Scheduled agentic workflows got the same treatment from the other direction. A workflow turn generates for several minutes and then spends more executing tools, which was long enough for its cached context to expire before the next turn could read it — so every turn re-billed the full prefix. Those runs now hold their cache across the gap. Task Agents deliberately keep the short-lived setting, because their turns fire seconds apart and the longer hold would buy nothing.

An OpenAI conversation could brick itself permanently

One failure mode deserves naming, because if you hit it there was no way out.

Past a certain length, a conversation on an OpenAI-dialect provider could pick up an internal record of the wrong kind and replay it on every subsequent turn. The provider rejected it every time, so the conversation failed permanently — not one turn, every turn, forever. Reloading did not help.

Both halves are fixed: the record is no longer captured, and a conversation already carrying one now skips it on replay. That second half matters most to you — an affected conversation heals itself as soon as the release lands, with nothing to migrate and nothing to recreate.

See AI Provider Connections for the full connection surface, and Chat for where the cost footer appears.

Questions or feedback? Reach us at info@infrascout.cloud.