GPT-6 Astra and Claude Fable 5.1 join the model pickers

Two models joined the pickers this week, alongside a reasoning-effort fix for Claude Sonnet 5 and Opus 5 and clearer pricing data in the model lists.

In this entry

1. GPT-6 Astra · OpenAI's new flagship, with one behavior worth knowing
2. Claude Fable 5.1 · cheaper cache reads, stricter data handling
3. Reasoning Effort on Claude Sonnet 5 and Opus 5
4. Pricing and retirement dates now show in the pickers

---

1 · GPT-6 Astra

OpenAI's new flagship, above GPT-5.6 Sol, is available as both a Workbench Model and a Target Model at $10 per million input tokens and $50 per million output.

One behavior is worth knowing before you switch to it: Astra has no reasoning-off mode. Its effort ladder starts at Low. If your saved effort preference is None and you select Astra, the control moves you up to Low and tells you it did · the one case where a model switch adjusts your setting upward rather than down. That is the model's constraint, not a choice we made.

Astra is the second model in the pickers to work this way. Gemini 3.7 Flash has the same shape, and its effort control has always started at Low for the same reason.

OpenAI reports Astra reaching stronger results while spending fewer output tokens than its predecessors, so the real cost gap against Sol is narrower than the headline rates suggest. Worth measuring on your own prompts rather than taking on faith.

2 · Claude Fable 5.1

Available as a Target Model, like Fable 5 before it. Its turns can run for many minutes to hours, which does not fit the live streaming panels, but you can author, lint, and critique prompts written for it.

Two properties set it apart from every other Claude model:

• Cache reads cost a quarter of what they do elsewhere. Fable 5.1 bills cached input at a rate four times better than the rest of the Claude line. On a workload that reuses a long prefix across many calls, that is a real change in the economics, not a rounding difference.
• It requires 30-day data retention. Fable 5.1 is not available under zero-data-retention agreements without specific authorization from Anthropic. If your organization runs under a ZDR arrangement, this model is not a drop-in option, and that is worth checking before you write prompts against it.

3 · Reasoning Effort on Claude Sonnet 5 and Opus 5

Those two models reason by default, unlike earlier Claude models. Setting Reasoning Effort to None now sends an explicit off for them rather than relying on the default.

If you run either at None, expect shorter responses and lower per-call cost. Every other Claude model is unaffected.

4 · Pricing and retirement dates now show in the pickers

Model names in the Target and Workbench pickers now carry their time-sensitive facts inline. GPT-4.1 Nano reads GPT-4.1 Nano · retires 2026-10-23, and GPT-5.6 Sol reads GPT-5.6 Sol · promo price until 2026-11-21. Previously both dates lived only in provider documentation.

GPT-5.6 Sol's current $4 / $20 is promotional. It holds at least through 2026-11-21, then returns to $5 / $30 · a 25% increase on input and 50% on output.

Cost estimates account for the expiry. The figures shown today remain the promotional ones, because that is what you are billed.

Improvements & fixes

• The token budget bar now uses GPT-5.6 Sol's real input ceiling. Sol accepts 922,000 input tokens, matching Terra and Luna.
• Deprecation status re-checked across both providers. Nothing new is retiring. GPT-4.1 Nano still shuts down on 2026-10-23 and stays selectable until then. Claude Sonnet 4.5 and Haiku 4.5 have no retirement date · dates published against them are earliest-possible floors, not schedules, and Anthropic commits to at least 60 days' notice.
• Claude Sonnet 5's $2 / $10 is now permanent. A scheduled increase to $3 / $15 was cancelled, so it stays cheaper than the Sonnet 4.6 it replaces.