v3.5.0

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

• Opus 5.5 defaults to medium effort. On /thinking auto it runs one level
  lower than Opus 5 did; set /thinking high for the old depth.
• GPT-6 Sol and Luna run agent turns with reasoning off. On the connection
  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.
• GPT-6 Astra is no longer offered on OpenAI: it cannot call tools on that
  connection at all. Saved settings and custom bots on Astra move to GPT-6 Sol.
  Astra is still available through OpenRouter.
• Saved settings on retired models now move to their replacements for
  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

• Claude Opus 5.5 (claude-opus-5-5), now the Anthropic default: $4/$20,
  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.
• Grok 4.7 (grok-4.7): $2/$6, 500K context. The Grok default stays Grok
  Build 0.1, which costs half as much.
• GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), 1.05M context.
• DeepSeek V4 Pro is back — DeepSeek cancelled its retirement. Users moved
  to Flash in 3.3.0 stay on Flash and can pick V4 Pro again.
• GLM-5.3 FlashX on Z.AI pay-per-use; Qwen 3.8 Max and Flash on Qwen China
  pay-per-use and on the Token Plan.
• /thinking on Kimi Code (kimi-for-coding, now K2.8 Preview with 1M
  context).

Changed

• Max sends max on GPT-5.6 and GPT-6 (it sent xhigh), and xhigh on Grok
  4.6 and 4.7. Through OpenRouter, Max goes as high as each model supports
  instead of stopping at high.

Fixed

• Cost estimates: OpenAI cache writes (GPT-5.6 on) count at 1.25x input
  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.
• Agent runs on Claude models that call tools without writing text between
  calls no longer fail on their second step.
• Kimi Code plan limits (for example K3 past 256K on a Plus plan) come back as
  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.
• The task planner no longer sends a temperature to models that reject it.