Claude Opus 5, Sonnet 5, the GPT-5.6 family, and the Gemini 3.x line

Twelve new models are selectable today, across Anthropic, OpenAI, and Google. The Reasoning Effort control gains a sixth level for the models that support it, Sonnet gets cheaper, and a handful of pricing and capability figures you were already relying on turned out to be wrong · those are corrected too.

In this entry

1. Twelve new models · what you can pick now
2. A new Max reasoning level · and why it only appears on some models
3. Sonnet 5 costs less than Sonnet 4.6 · a rare direction for a model upgrade
4. Two models left the Workbench dropdown · and why that is a fix
5. GPT-4.1 Nano retires October 23 · what happens to your prompts

---

1 · Twelve new models

Anthropic. Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8 are available as both a Workbench Model and a Target Model. Claude Fable 5 is available as a Target Model only · its turns can run for many minutes to hours, which does not fit the live streaming panels, but you can still author and lint prompts written for it.

OpenAI. GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. Sol is the flagship; Terra is the balanced mid-tier; Luna is the volume tier at $0.20 per million input tokens.

Google. Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.1 Flash-Lite. Gemini 2.5 Pro, Flash, and Flash-Lite all stay · Google has not announced a shutdown date for any of them, and 2.5 Flash-Lite is still the cheapest model we can reach from any provider.

Everything runs on your own key, as always. We do not mark up tokens.

Each new model also carries its own prompting guidance, which the AI Pair reads when you critique or improve a prompt written for it. That guidance is model-specific rather than family-generic: Opus 5, for instance, responds badly to the "double-check your answer" style of instruction that helped older models, and benefits from explicit scope limits because it expands scope on its own. Sonnet 5 follows instructions more literally at lower effort, so unstated scope gets skipped. The GPT-5.6 family scores better on leaner prompts that state each instruction once.

2 · A new Max reasoning level

The Reasoning Effort control now offers Max above X-High. It appears only on models that actually accept it, which is the important part.

Provider reasoning ladders stopped agreeing with each other. OpenAI added a top level above what we previously exposed, and Anthropic inserted a new rung in the middle of theirs. So the control is now model-aware in both directions: pick Claude Opus 5 and you get all six levels; pick Claude Opus 4.6 or Sonnet 4.6 and X-High is absent, because those models genuinely do not have that rung; pick a Gemini 3.x model and you get four, because Google's ladder tops out lower · or three on Gemini 3.7 Flash, which does not accept the shallowest setting either.

You will never be shown a level that the model would reject. If you switch to a model that does not accept your saved preference, the control clamps to the closest level it does accept and tells you it did. When two levels are equally close it picks the lower one, so a model switch does not quietly raise what you spend. The one exception is a model whose floor sits above your saved level · Gemini 3.7 Flash has no shallowest setting, so a saved None becomes Low there simply because there is nothing below it.

One Gemini 3.x nuance worth knowing: reasoning cannot be switched off on those models. Choosing None gives you the shallowest setting the model allows rather than no reasoning at all.

3 · Sonnet 5 costs less than Sonnet 4.6

Claude Sonnet 5 is $2 per million input tokens and $10 per million output, against Sonnet 4.6 at $3 and $15. That is a capability increase and a price cut in the same release, which is unusual enough to call out. If you are running Sonnet 4.6 today, Sonnet 5 is worth a look on both axes.

While checking prices we also found four figures in our own model data that were wrong, all of which affected you:

• Claude Sonnet 4.6's maximum output was recorded at half its real ceiling. Long generations could hit an invented limit rather than the model's actual one.
• Two models had their prompt-caching threshold set too high, so some prompts skipped a cache write that would have succeeded · and paid full input price on the next call as a result.
• Claude Opus 4.7's token estimate ran about 20% low, which made the editor's budget bar under-report how close a prompt was to the limit.
• GPT-4.1's context window was rounded down, and the family's cached-input rates were missing from cost estimates entirely.

4 · Two models left the Workbench dropdown

GPT-5.5 Pro and GPT-5.4 Pro are no longer offered as Workbench Models. They remain available as Target Models, so you can still write and lint prompts aimed at them.

This is a fix rather than a removal. OpenAI serves both models on an API surface we do not call, so selecting either as your Workbench Model could never have produced a result · the dropdown was advertising a path that did not exist. Rather than leave a broken option in place, we removed it from the picker until the integration that would make it work is built.

If either was your saved Workbench Model, you do not need to do anything. Your workspace falls back to the default model automatically.

5 · GPT-4.1 Nano retires October 23

OpenAI has scheduled GPT-4.1 Nano for shutdown on 2026-10-23. It still works until then, and it is still selectable · we are not going to break a model out from under you while the provider still serves it.

Two things have changed ahead of that date. The model picker now shows a retirement date on models that have one, so you can see the clock before you commit a prompt to one. And where we automatically pick a cheap model on your behalf, for example when scoring an eval without a model preference set, that pick has already moved off Nano to its replacement.

Anything you have already saved against a retiring or renamed model keeps working. Model IDs from earlier eras still resolve · we have never removed one of those mappings, and this release added sixteen more.

Improvements & fixes

• Long model lists stay on screen. The Target Model dropdown now holds 32 entries, which was enough to run off the top of the viewport on some screens. Model pickers now cap their height and scroll.
• Reasoning-heavy runs no longer time out early. The highest effort levels were being given a shorter time budget than the level below them, so the slowest runs were the most likely to be cut off.
• OpenAI-format exports carry a real model ID. Exports were stamping a model name that no longer exists at OpenAI.
• Cost warnings compare against a model that still exists. The "this model is expensive" threshold was anchored to a model now being retired.