Subject: Opus 4.7 model picker collapsed from 5 effort rows to 1 between v3.0.12 (Linux) and v3.3.27 (Windows)
Bug summary
Between Cursor v3.0.12 (Linux, built 2026-04-04, commit a80ff7dfcaa45d7750f6e30be457261379c29b00) and Cursor v3.3.27 stable (Windows 10.0.26200, commit 80b138a7a0a948e1a798e9ed7867d76a1ba9a310) the model picker for Claude Opus 4.7 collapsed from five distinct selectable rows into a single consolidated row.
What v3.0.12 (Linux) shows
Five separately-selectable rows in the chat/agent model picker:
-
Opus 4.7
-
Opus 4.7 Max
-
Opus 4.7 Low
-
Opus 4.7 Medium
-
Opus 4.7 High
What v3.3.27 (Windows stable) shows
Only one Opus 4.7 row visible (labeled along the lines of “Opus 4.7 Extra High” / xhigh). Low / Medium / High / Max are no longer selectable from the picker. The current published Cursor models documentation describes a single “thinking variant”, which appears to match the new collapsed UI but does not match the prior UI that v3.0.12 still serves.
Why this matters
Anthropic’s published effort scale for Opus 4.7 (in their Claude API extended-thinking reference) explicitly defines five tiers — low, medium, high, xhigh, max — with materially different cost and latency profiles. Without per-effort selection in the Cursor picker, users cannot drive the cost/quality trade-off Anthropic exposes. We were actively using Opus 4.7 Medium for cost-sensitive runs and Opus 4.7 Max for hard architecture work; both are now unreachable from the Windows picker.
Repro
-
On Windows 10.0.26200, Cursor stable v3.3.27 (commit 80b138a), open the chat/agent model picker.
-
Locate Claude Opus 4.7. Observe only one row (no Low/Medium/High/Max).
-
On a Cursor v3.0.12 Linux build (commit a80ff7d) signed in to the same account, open the same picker. Observe five Opus 4.7 rows.
Notes
-
Linux v3.0.12 is dated 2026-04-04, which is before Anthropic shipped Opus 4.7 on 2026-04-16, so the per-effort rows must be coming from a Cursor backend response, not the client binary. That suggests this is a server-side rollout, not a client regression, and is therefore not addressable by downgrading the IDE.
-
I checked every page in the Cursor documentation site and the Cursor community forum that touches models, picker, settings.json, BYOK, Max Mode, subagents, enterprise model controls, and SDK / CLI. Nothing documents a way to re-expose the per-effort rows.
-
I checked every changelog entry between 2026-04-02 (3.0) and 2026-05-07 (3.3). No entry announces this consolidation.
-
I am happy to share links to two adjacent (non-duplicate) forum reports in a follow-up reply once new-user link limits allow.
Asks
-
Confirm whether the consolidation is intentional and, if so, link the changelog entry. If unintentional, treat as a UI regression.
-
If intentional, expose either:
-
(a) a per-row picker like v3.0.12 still serves, gated behind a Settings → Models toggle, or
-
(b) a documented setting (settings.json key, command, env var, or team admin model-config option) that pins effort = low / medium / high / xhigh / max for Opus 4.7.
-
-
Document the chosen path in the official Cursor models documentation and reflect it in the published per-model list the way the GPT-5 Fast variants are already enumerated.
Request IDs
-
Windows v3.3.27 reproduction: ea11c600-5502-489c-b4a4-9ee09585bc37
-
Linux v3.0.12 reproduction: 437fd93d-8b9a-440a-8b0d-e41f5b003579
Privacy mode was set to Privacy Mode at the time of each reproduction.
