Roster change: Claude Opus 5.5 and GPT-6 Astra replace three models in October

What changes

The commercial models measured here change in October. The new models are each provider’s current flagship, and the models they replace are no longer the ones most people get by default.

Model	Role	Week
gpt-6-astra	new series, odd weeks	first week 2026-W41
gpt-5.5	series ends	last week 2026-W41
claude-opus-5-5	new series, even weeks	first week 2026-W42
claude-opus-4-8	series ends	last week 2026-W42
claude-opus-5	series ends	last week 2026-W42

Weeks are run labels. A run is labelled with the ISO week that ended the day before it starts, so 2026-W41 is the run of Monday 12 October 2026 and 2026-W42 the run of Monday 19 October 2026.

From 2026-W43 on, the commercial roster is GPT-6 Astra on odd weeks and Claude Opus 5.5 on even weeks. The local Llama baseline still runs every week, and the stance classifier (Claude Haiku 4.5) is unchanged.

One week of overlap

A new model is not a continuation of the old one. A drift record whose subject silently changes is not a drift record, so each new model starts its own series rather than inheriting the old model’s history.

To make the handover readable anyway, each retiring model runs in the same week as its successor, once:

• 2026-W41: GPT-5.5 and GPT-6 Astra answer the same corpus in the same week.
• 2026-W42: Claude Opus 4.8, Claude Opus 5 and Claude Opus 5.5 answer the same corpus in the same week.

Same week means same prompts and the same state of the world, so differences between the old and new model in that week are differences between the models, not between weeks. That is the comparison to use when asking “did the switch change the picture?” Comparing the new model’s later weeks with the old model’s earlier weeks mixes the change of model with ordinary drift.

The retired series stay on the site. Their model pages, weekly snapshots and raw data remain where they are and keep their links. They simply stop gaining new weeks.

Things to know when reading the new series

Claude Opus 5.5 thinks less by default than Claude Opus 5 did. We measure every model at its provider defaults and do not set a reasoning effort level. Anthropic’s default effort for Claude Opus 5.5 is “medium”, where Claude Opus 5’s default was “high”. Part of any difference between the two in the 2026-W42 overlap may therefore come from that default, and not only from the new model. We record it rather than correct for it, because the default is what people using the model get.

Claude Opus 5.5 always thinks. Its internal reasoning cannot be switched off, and it does not accept a temperature setting. As with the other thinking-by-default models (see the methodology page), it is sampled 20 times per prompt at the default setting, with no temperature-0 samples.

GPT-6 Astra is treated like GPT-5.5. OpenAI has not documented which request settings it accepts. We send it exactly what we sent GPT-5.5: no temperature setting (so, again, 20 samples per prompt and no temperature-0 samples) and the same 8192-token response limit, which leaves room for its reasoning so that it is not cut off before it answers.

No substitute model ever answers. Some providers can be asked to hand a request to a different model when the first one declines it. That option is never enabled here. Every response stored under a model’s name came from that model, and a refusal is recorded as a refusal, not replaced by another model’s answer.

Cost

The overlap weeks cost more because they run two or three frontier models at once. Our pre-run estimates are $61.11 for 2026-W41 and $42.73 for 2026-W42, against $38.20 for an ordinary odd week (GPT-6 Astra) and $15.28 for an ordinary even week (Claude Opus 5.5) from 2026-W43 on. The run’s spending ceiling now scales with each week’s estimate instead of being a single fixed number, with a hard maximum, so the overlap weeks can run without lifting the cap on ordinary ones. Details are in meridian/BUDGET.md.

@misc{meridian20261006rostersuccession,
  title        = {{Roster change: Claude Opus 5.5 and GPT-6 Astra replace three models in October}},
  author       = {Meridian},
  year         = {2026},
  howpublished = {\url{https://meridianaudit.org/reports/2026-10-06-roster-succession/}},
  note         = {Accessed: YYYY-MM-DD}
}