Five Frontier Models in One Week: What Actually Changes in Your Agent Stack

Last Updated: September 28, 2026

What shipped, and when

This was not a coincidence week. It was four labs landing inside 72 hours of each other, which tells you something about how release calendars are now set.

Pat McGuinness's roundup of the week is the best single read on the shape of it, and his framing is right: four of the five sit on the Pareto intelligence-cost frontier, and all five have a use case. But roundups optimise for "which model is best." That is the wrong question if you operate agents. The right question is which line items on your invoice just moved.

The real headline is the cache line

Everyone quotes input and output prices. In a production agent loop, neither is the dominant cost. Cached context reads are — and Anthropic said so explicitly, describing cache reads as making up the majority of agentic and coding work costs before dropping them from $0.50 to $0.20 per million, a 60% cut.

OpenAI went further than a discount. Two changes in the GPT-6 caching update are architectural, not commercial: you can now change reasoning effort mid-conversation and enable or disable tools without invalidating the cached prefix, and explicit breakpoints let you choose where a cached prefix ends.

If you have ever built an escalation ladder — run cheap, detect low confidence, re-run harder — you know why that matters. Until now, bumping effort or swapping the tool list rewrote the prefix and you paid full freight for context you had already sent. Dynamic effort escalation was theoretically obvious and economically stupid. It is now neither. That single change probably alters more production architectures this quarter than any benchmark on any chart.

Cost per token is not cost per task

Here is where the price war gets slippery. Grok 4.7 lists at $2/$6 — cheaper per token than GPT-6 Sol's $2/$10. McGuinness's read is that Sol is more token-efficient and ends up cheaper on balance, and that Grok 4.7 burns more tokens than 4.6 for little gain. Anthropic makes the same argument in its own favour: Opus 5.5 costs less per token and uses fewer tokens per task, which is how a 20% list cut becomes a claimed 40% workload cut.

What this breaks in a stack you are already running

At these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences.

— Anthropic, Claude Opus 5.5 announcement

The fallback rate nobody is pricing

This is the most operationally important detail of the week, and it is buried in footnotes on two different vendor pages.

Anthropic notes that Opus 5.5 ships with safeguards similar to Fable 5.1's, and that in its own benchmark runs, when safeguards intervened, cybersecurity tasks were completed by Opus 4.8 and biology tasks by Opus 5. It also notes that Zapier's AutomationBench runs were performed without fallback models, so safeguard interventions were considered failures. OpenAI, comparing on the same benchmark, observes that a competitor datapoint understates real cost because Opus 5 fallbacks occurred on roughly 40% of tasks.

Strip away the vendor sniping and the lesson is the same for both: at frontier tier, a non-trivial fraction of agent steps may not be served by the model you selected. That has consequences a benchmark table cannot express — different latency, different token accounting, different behaviour under the same prompt, and an audit trail that says one model while another did the work. If you run agents in a regulated context, your model-attestation logging needs to record what actually served the request, not what your config requested.

What I'd do this week

We run this pattern for clients as a standing quarterly exercise rather than a reaction to release weeks — see how we structure it under automation and model selection.

The honest caveats

Three things this week is carefully not saying. First, the cost claims are workload-dependent: Anthropic's 40% figure is measured at default settings on typical workloads, and McGuinness reports that real user metrics suggest a cost closer to Opus 5. Your mileage depends entirely on your cache hit rate and effort settings.

Second, availability is not uniform. OpenAI shipped Sol and Luna to ChatGPT Work, Codex and the API, but not yet to the standard Chat experience. Opus 5.5 is available across AWS, Google Cloud and Azure on day one, with Sonnet 5.5 and Haiku 5.5 still to come — meaning the cheap end of the Claude family has not yet been repriced.

Third, open weights are not free inference. MiMo-V2.6-Pro is genuinely impressive and genuinely open, with a published technical report and weights on Hugging Face. It is also described in user reports as token-hungry and slow to think. Self-hosting a 1T-parameter MoE to save on a workload that GPT-6 Luna handles for $0.10 per million input tokens is a decision that needs a spreadsheet, not an ideology.

The frontier moved this week. Your unit economics moved further. If you want a second pair of eyes on the routing table before the invoice arrives, start here.