Two frontier labs ship a generation at once: GPT-6 and Claude Opus 5.5

Two new model families land in the same week

OpenAI released GPT-6 Sol and GPT-6 Luna as reasoning models that accept text and image inputs, priced far below their GPT-5.6 predecessors: Sol at $2 input / $10 output per million tokens, Luna at $0.10 / $0.50 for high-volume work. The launch is positioned as frontier intelligence for everyday work, and the same pairing rolls into Codex and ChatGPT Work — Sol for complex agentic coding, Luna for focused bulk tasks — reaching Plus, Pro, Business, Enterprise, and Edu, with Enterprise admins required to enable them. Prompt caching got its own treatment alongside the launch, with higher cache hit rates, explicit breakpoints, and new diagnostics to cut latency and cost.

Anthropic answered with Claude Opus 5.5, the first model in the new Claude 5.5 family, performing at Fable 5.1's level on most work while costing 40% less to run than Opus 5. Anthropic frames it as matching Fable 5.1 intelligence at higher speed, writing more clearly for long sessions, with raised usage limits on Pro, Max, and Team. On the platform side it ships with a 1M token context window by default, always-on adaptive thinking, and $4/$20 per MTok across the Claude API, Bedrock, Google Cloud, and Microsoft Foundry — a release carrying real breaking changes, since thinking can no longer be disabled and some tool_choice values now return 400 errors.

Codex becomes a cross-platform surface

Codex's week was mostly about catching up to the new models. The CLI added GPT-6 Sol and Luna with Amazon Bedrock support and migration prompts, turned fullscreen transcripts on by default, and fixed proxy routing, network restrictions across redirects, and tmux mouse handling. Before that, the models became selectable from the model picker, with the rate-limit switch prompt now recommending Luna. Codex also arrived on iOS as ChatGPT for iOS with a redesigned home screen and an iPad split view that keeps the task list beside the open task, plus nested Git repository browsing and side chats that carry over attachments and review comments.

SDKs race to expose the new models

The official SDKs moved in lockstep. Python added GPT-6 Sol and Luna model identifiers, then GCP external storage and the GPT-Rosalind research model alongside fixes that had clients retrying only replayable content and async methods using asyncio.get_running_loop; earlier in the week it added external storage configuration and safety case retrieval. Two patches followed: one preserving single-pass chat tool iterables and merging HTTP headers case-insensitively, another preserving single files for fallback extraction. Go added the same Sol and Luna identifiers, then GPT-Rosalind and GCP external storage, after fixing legacy model choices no longer being preserved. Node added Sol and Luna identifiers, GCP external storage and GPT-Rosalind, and session environment reset events.

Claude Code rides the new default

Claude Code made Opus 5.5 the default Opus model with 1M context and $4/$20 pricing, and tightened write approval so a write through a symlinked path is judged by where it lands — no more auto-approved writes landing outside the tree. It also hardened resume and proxy stream handling, stopping rm -rf "$(pwd)"-style commands from running unprompted in auto mode and fixing truncated gateway streams being shown as complete. A maxProseWidth setting caps prose in wide terminals, with fixes for 400 errors from undecryptable web search history. Admins got model governance settings and a /doctor prompt-audit command to block specific models and flag configs written for older ones. Elsewhere, Anthropic's life sciences group reported that roughly 950 Claude agents identified a novel CRISPR-like enzyme system, the Claude API added inline tool definitions in mid-conversation messages, and ChatGPT Voice gained plugin support across web, iOS, and Android.