Windsurf — Six Weeks of Changelog from Adaptive to Devin Local

https://windsurf.com/changelog

Why This Matters

Six weeks of Windsurf shipping in one place. The story across the window — April 6 through May 12, 2026 — is the Cognition-era reshaping of Windsurf into something more than a Cursor-shaped IDE. Three threads run through it:
  1. Adaptive model routing ships as a first-class quota-extending strategy, then gets polished and bug-fixed across two releases.
  2. Windsurf 2.0 lands with Devin (the cloud agent from the Cognition acquisition) integrated directly into the editor, plus an Agent Command Center.
  3. Devin Local (terminal CLI agent, Rust-implemented) ships shortly after, then graduates into a same-harness in-IDE agent that's ~30% more token-efficient than Cascade.
If you're tracking the consolidation of "AI IDE" and "agentic dev tooling" into a single product, this six-week window is the consolidation happening in public.

The Timeline

April 6, 2026 — v1.9600.38: Introducing Adaptive

The Adaptive model router debuts as a new entry in the model picker. The premise: intelligently select the optimal model per task and draw quota at a fixed per-token rate rather than per-model. Available immediately to Pro, Max, and Teams self-serve. Promotional pricing for the first two weeks — $0.50 / 1M input, $2.00 / 1M output, $0.10 / 1M cache-read.
Same release: the model picker gets inline token pricing, a prompt-cache timer in the context-window indicator, and per-response token counts on response cards. This is the cost-transparency push that the whole AI-IDE category has been moving toward — Windsurf is the first to put per-model rates and cache state directly in the picker.

April 6, 2026 — v1.9600.40: Adaptive Model Visibility

Same day, second release. Adaptive's display in the model picker gets a UI polish — a small change, but worth noting because the team shipped a same-day follow-up specifically to make the new routing option more discoverable. Read that as a signal of how much they want users to actually try Adaptive rather than stay on a default Claude/GPT pin.

April 7, 2026 — v1.9600.41: Adaptive Fix

A bug in the adaptive router was preventing model switching after the initial request — i.e., Adaptive was picking once and sticking. Fixed within 24 hours. Affected users had quota reset and overage restored. That's the right way to handle a billing-adjacent bug.

April 15, 2026 — v2.0.44: Windsurf 2.0

The major release of the window. Three structural changes:
  • Devin in Windsurf. The Devin cloud agent is integrated into the editor for all self-serve plans. One-click delegation from local session → Devin Cloud with a dedicated VM. Review changes and test results without leaving the editor. First Devin Cloud session gets up to $50 credit. Enterprise is disabled by default; admins enable per-org.
  • Agent Command Center. Kanban-style view of all local and cloud agent sessions by status. Sessions, PRs, files, and context group into Spaces — task-level containers you switch between.
  • Refined Windsurf Browser with toolbar integration and a Cascade tool that can read page content directly.
Plus accelerated Cascade sidebar load, improved .gitignore/.codeiumignore handling, remote-extension stability (WSL, SSH, Dev Containers), and typing performance in large active diff zones.

April 16, 2026 — Claude Opus 4.7

Opus 4.7 lands in Windsurf the day after 2.0 ships.

April 16 / April 17 / April 20 / April 21, 2026 — v2.0.50 / 2.0.61 / 2.0.63 / 2.0.67

Stabilization sequence after 2.0. Windows Terminal spawning fix, OAuth regression on certain MCP servers patched and re-patched, Devin Cloud connection reliability improvements. The cadence (four point releases in six days) tells you 2.0 shipped with real edges and the team chose visibility over polish.

April 24, 2026 — GPT-5.5

GPT-5.5 lands in Windsurf.

April 28, 2026 — v2.1.29: Devin for Terminal + Devin Local in Windsurf

The second structural shift of the window:
  • Devin for Terminal. A new CLI agent running locally on the user's machine, Rust-implemented, with seamless handoff to Devin Cloud (dedicated VM, testing). Multi-model support across Opus 4.7, GPT-5.5, and SWE-1.6. All Windsurf users get it with existing subscriptions.
  • Devin Local in Windsurf. Same agent harness as the terminal CLI, sharing session access with it. Up to 30% greater token efficiency than the existing Cascade agent.
This is the more important release of the two. It collapses the gap between the IDE agent and the terminal agent into a single substrate, and the token-efficiency number is the kind of claim that, if it holds, changes the per-task cost math.
Plus: streaming search subtitle layout, drag-and-drop fix in Cascade tabs, Ctrl+Shift+G keybinding fix on Windows/Linux, server-driven extension deny lists.

April 29, 2026 — v2.1.32

Bug-fix release. Crash on switching between Cascade conversations fixed. Auth issue preventing Devin Cloud session startup fixed. Agent-question response routing fixed. All three are usability-blocking; getting them out the door within 24 hours of v2.1.29 is appropriate.

May 6, 2026 — v2.2.17: Devin Review + Quick Review

Devin Review and Quick Review become available to all Windsurf IDE users on existing subscriptions. Self-serve users get a two-week free trial on Devin Review specifically. Enterprise users need a Cognition platform agreement to get Devin Review — the line where "Windsurf plan" stops being enough.
Also: list display option for agent inbox, sessions sidebar sorting/filtering, performance improvements for session loading and switching. A Windows update bug that was blocking upgrades got patched — workaround was killing devin.exe processes before updating.

May 12, 2026 — Claude Opus 4.7 (Fast Mode)

Opus 4.7's fast mode lands in Windsurf. "Full Opus 4.7 intelligence with approximately 2.5x faster output speeds." For an IDE where latency on inline edits matters a lot, this is a meaningful daily-driver improvement, not just a model addition.

What The Shape Tells You

Three observations across the six weeks:
  1. Adaptive is the strategic feature. Cost-per-task is the line that AI IDEs are starting to compete on, and a routing layer that picks among Opus / GPT / SWE-1.6 / whatever-comes-next is a more durable position than betting on any single provider being cheapest. The same-day visibility patch on April 6 is the giveaway — they want this used.
  2. Cognition's Devin is being absorbed methodically. April 15 → Devin Cloud in the IDE. April 28 → Devin terminal CLI + Devin Local in-IDE sharing the same harness. May 6 → Devin Review and Quick Review. That's the acquisition payoff sequencing: cloud → local → review surfaces, each two weeks apart.
  3. The token-efficiency claim on Devin Local (~30% better than Cascade) is the line in this window most worth verifying in real use. If it holds, Cascade has a short shelf life as the in-IDE agent. If it doesn't, this is a marketing number.

Our Take

The interesting thing about the Windsurf release stream is that it's no longer a Cursor-style "IDE with AI bolted on" — it's converging on an agent platform that happens to ship with an editor and a terminal client. The Spaces / Agent Command Center / shared-harness work is the structural change that signals where the product is going. Worth a fresh look for anyone who evaluated Windsurf earlier in the year and made a Cursor-vs-Windsurf-vs-Claude-Code decision based on the Cascade-era product.