v3.5.0

Released September 25, 2026 · View on GitHub →

TL;DR — Claude Opus 5.5 is the new Anthropic default, Grok 4.7 and GPT-6 Sol and Luna arrive, DeepSeek V4 Pro is back, and saved settings on retired models now move to their replacements.

A full review of every provider's models on 2026-09-23, each change checked against the provider's own pages.

Upgrade notes

lower than Opus 5 did; set /thinking high for the old depth.

Codeep uses for OpenAI (Chat Completions), GPT-6 can call tools only that way, so agent turns send reasoning_effort: "none" whatever /thinking says; /thinking tells you so. Plain chat keeps your tier. Support for OpenAI's Responses API, which lifts this, is being built.

connection at all. Saved settings and custom bots on Astra move to GPT-6 Sol. Astra is still available through OpenRouter.

everyone.** Earlier migrations only reached configs created before 2026-08-15, so some of them run for you for the first time: - GLM Coding Plan: GLM-5.2, 5.1 and 5 move to GLM-5.3, GLM-5 Turbo to 5.3 Flash — both plans now accept only those two models. GLM-5 Turbo also moves to 5.3 Flash on Z.AI international pay-per-use. - Qwen: the Token Plan's retired Qwen3.8-Max Preview moves to Qwen3.8-Max; Qwen3-Max and the Qwen3 coders (retiring 2026-10-10) move to Qwen 3.7 Plus or Max, whichever the plan offers. - Gemini 3 previews move to Gemini 3.6 Flash and 3.1 Pro. - ModelScope's old fallback, Qwen3-Coder-480B, is no longer served; it moves to Qwen3.5-397B. - Custom bots, /rewind checkpoints and editor settings pinned to a retired model now run on its replacement instead of failing.

Added

1M context, cache reads at 5% of input. Codeep leaves it at least 32K reply tokens (64K at Max), because its thinking shares that limit. Opus 5 stays in the picker.

Build 0.1, which costs half as much.

to Flash in 3.3.0 stay on Flash and can pick V4 Pro again.

pay-per-use and on the Token Plan.

context).

Changed

4.6 and 4.7. Through OpenRouter, Max goes as high as each model supports instead of stopping at high.

Fixed

instead of not at all; cache reads use each model's own rate for Opus 5.5, Kimi, every Grok model and GLM (per region); Qwen3.6-Plus is repriced.

calls no longer fail on their second step.

HTTP 401; the editor integration said "No API key configured". It now says the plan may not include that model or limit, and quotes Kimi's message.

Downloads

Install with npm install -g codeep@3.5.0.