OpenCode 2 v2.0.27: summarized Claude thinking, a GPT Astra prompt#

Published October 14, 2026 by Alex Sorokoletov

Part of the OpenCode 2 Version Tracker series. | OpenCode on GitHub | OpenCode 2 docs

OpenCode 2 v2.0.27 follows v2.0.26 with two changes to how it talks to models: Claude reasoning is now requested as summarized adaptive thinking (#48269), and OpenAI's GPT Astra models get their own system prompt (#48057). Two fixes land alongside, one for editor integrations over ACP and one for remote authentication in the terminal.

Model prompts and reasoning#

OpenCode selects a system prompt per model family. GPT, Claude, Kimi, Trinity and Meta models each resolve to their own prompt file, and anything unmatched falls back to a default. This release adds a dedicated prompt for OpenAI's GPT "Astra" models: any model whose id contains astra, such as gpt-6-astra and gpt-6-astra-fast reached through the ChatGPT subscription provider, now resolves to that prompt instead of the standard GPT one (#48057).

Claude reasoning requests changed. For Claude models that are not pinned to a manual thinking budget, OpenCode now sends thinking as an adaptive mode with a summarized display, and layers the reasoning effort (low through max) on top of it rather than requesting a fixed token budget (#48269). The same adaptive setting applies to Claude reached directly, through Amazon Bedrock, and through SAP AI Core; models configured for manual thinking, and Opus 4.5, keep the manual path.

Editor and terminal fixes#

New environment variable#

The rest of the work in this range sits in model handling, editor integration, the terminal client, and the project's own website.

What this tells us#

The two model changes are provider-interface work. Per-family system prompts give each model its own instructions, and the adaptive-thinking request tracks the reasoning controls Anthropic exposes, with the effort level carried through the direct, Bedrock and SAP AI Core paths. That is the same provider-breadth maintenance the 1.x line has carried, now running in the v2 core rather than beside it.

The ACP and remote-auth fixes fit v2's server-backed shape. The ACP agent is an editor's attach point to that service, and remote authentication is the terminal's. Restoring the session options an editor sees and reporting a failed remote login at startup both harden that attach surface. It is a server-backed footprint: a background service the terminal, desktop, browser UI, and an external editor all attach to.

The website moved too. Install links and desktop downloads now resolve to v2 (#49874, #49931, #49984), and the 1.x documentation carries a v2 announcement banner. With v2 tagging new builds nearly every day, the download and documentation surfaces now lead to v2 rather than the 1.x line it still ships beside.

This analysis is based on publicly available open-source code and release metadata, conducted for educational and research purposes. All trademarks and software referenced belong to their respective owners.

Sources#

1. GPT Astra system prompt: #48057
2. Summarized adaptive thinking: #48269
3. ACP session options and reasoning: #48225
4. Remote auth startup errors: #49016
5. v2 install links and desktop downloads: #49874, #49931, #49984
6. Parallel web search tool: OpenCode 2 docs