AI Coding News: July 25, 2026 — Claude Opus 5 Becomes the Default in Claude Code and Lands in GitHub Copilot

Model launches usually take a week to reach the tools developers actually type into. Claude Opus 5 took about a day. Anthropic shipped it on July 24, 2026, Claude Code made it the default Opus model in the same 24 hours, and GitHub Copilot put it in the model picker before the day was out. Today's follow-up release, Claude Code v2.1.220, is the quiet cleanup pass behind it.

Underneath the headline, though, are two behavior changes that will silently alter how your agents run — and one that will return a hard error if you ignore it. Here's what shipped, and what you need to check before your next deploy.

Claude Opus 5 Is Now the Default Opus Model in Claude Code

Claude Code v2.1.219 landed on July 24, 2026 and led with the model swap: claude-opus-5 is now the default Opus model, carrying a 1M-token context window, with fast mode priced at $10/$50 per million tokens. Opus 4.7 was dropped from fast mode entirely — /fast now applies to Opus 5 and Opus 4.8 only.

On the API side, the Opus 5 release notes spell out the specs: 1M tokens is both the default and the maximum context (there is no smaller variant), 128k max output tokens, and pricing held at $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8, and half the cost of Claude Fable 5.

GitHub Copilot Added Opus 5 the Same Day

GitHub announced Claude Opus 5 availability on July 24 as well, for Copilot Pro+, Max, Business, and Enterprise plans. It is selectable from the model picker across an unusually wide surface area:

• Visual Studio Code, Visual Studio, JetBrains, Xcode, and Eclipse
• Copilot CLI and the GitHub Copilot cloud agent
• github.com, the GitHub Copilot app, and GitHub Mobile on iOS and Android

Two caveats worth noting. Rollout is gradual, so an empty model picker is expected rather than broken. And Business and Enterprise administrators have to explicitly enable the Claude Opus 5 policy in Copilot settings before anyone on the org sees it — the model does not appear on an opt-out basis. GitHub also flags enhanced safeguards around cyber-related content, which can block some security-adjacent requests that other models would answer.

The Behavior Changes That Can Break Your Integration

This is the part worth reading twice if you build on the API rather than just typing into a CLI. Migrating to Opus 5 is a one-line model ID change, but three behaviors shift underneath it.

1. Thinking is on by default

On Opus 4.8, requests ran without thinking unless you opted in with thinking: {"type": "adaptive"}. On Opus 5, the identical request thinks by default and the model decides how much. Because max_tokens is a hard ceiling on thinking plus response text combined, any workload that was tuned against a no-thinking Opus 4.8 baseline needs its token budget revisited — otherwise responses start truncating for reasons that will not be obvious from the error.

2. Disabling thinking at high effort is now a 400 error

This is the genuine breaking change. thinking: {"type": "disabled"} is accepted only when the effort level is high or below. Pair it with xhigh or max and the request returns a 400. On Opus 4.8 the two settings were independent, so this is a live regression risk for anyone who disabled thinking for latency and separately turned effort up for quality. Fix it by choosing one: keep thinking disabled and drop effort to high, or keep the effort level and remove the thinking field entirely.

Anthropic's own guidance is to keep thinking on and control cost with lower effort instead. With thinking off, Opus 5 can occasionally write a tool call into its text output rather than emitting a proper tool_use block, or leak internal XML tags into the visible response — an ugly failure mode for anything parsing structured output.

3. Effort now carries real weight

The full ladder is low, medium, high (the default), xhigh, and max. Anthropic's claim is that Opus 5 converts extra effort into better output more reliably than any earlier Opus, which makes the setting a real tuning dial rather than a rounding error. The practical advice: start at the default and move in both directions against your own evals — step down where quality holds to save tokens and latency, step up for the hardest work. If you run at xhigh or max, set a generous max_tokens so the model has room to think across subagents and tool calls.

One more behavioral note that costs nothing to act on: Opus 5 verifies its own work without being asked. If your prompts carry instructions like "include a final verification step" or "use a subagent to verify" from earlier models, delete them — they now cause over-verification.

Two Quieter Opus 5 Wins: Cheaper Caching and Mid-Conversation Tools

The minimum cacheable prompt length on Opus 5 drops to 512 tokens, down from 1,024 on Opus 4.8. Prompt caching now applies to short system prompts and compact tool definitions that previously fell below the threshold — a free cost reduction with no code changes for anyone running high-volume, small-prompt workloads.

Separately, mid-conversation tool changes entered beta behind the mid-conversation-tool-changes-2026-07-01 header. You can add or remove tools between turns while preserving the prompt cache, instead of resending a fixed tool list for the life of a session. For agent frameworks that swap toolsets per phase — research, then edit, then verify — this removes a real source of cache invalidation.

What Else Shipped in Claude Code v2.1.219

The model swap overshadowed a genuinely large release. The changes most likely to affect how you work:

• Nested subagents now go three levels deep by default, up from one. Subagents can spawn their own subagents, which changes the cost profile of any workflow you thought was flat. Set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to restore the old behavior.
• A new sandbox.network.strictAllowlist setting denies non-allowlisted hosts for sandboxed commands without prompting — useful for CI and unattended runs where an interactive prompt just means a hung job.
• A DirectoryAdded hook fires after /add-dir or the SDK's register_repo_root registers a new working directory mid-session — the missing piece for anyone scripting per-repo setup.
• MCP debugging got noticeably less painful. claude mcp list and /mcp now surface HTTP status codes and error text when a server fails to connect, plus a warning for config values carrying hidden leading or trailing whitespace. A new mcp_server_errors field in the headless stream-json init event lists --mcp-config entries skipped by validation.
• Dynamic workflows now default to a medium size guideline (aim for fewer than 15 agents), configurable via the new workflowSizeGuideline settings key or the Dynamic workflow size row in /config.
• Windows and accessibility fixes. CLAUDE_CODE_GIT_BASH_PATH is now ignored with a warning instead of exiting when the path isn't a bash or sh binary, and screen-reader mode stopped rewriting the entire input line on every keystroke.

Today's Release: v2.1.220

Claude Code v2.1.220 shipped today, July 25, 2026, with a single changelog line: bug fixes and reliability improvements. After a release as large as v2.1.219, a quiet stabilization pass is exactly what you want to see — and a reasonable signal that upgrading now is lower risk than it was yesterday.

Elsewhere in AI Coding

OpenAI Codex — the Codex changelog shows release 26.715 on July 23 bringing ChatGPT Voice powered by GPT-Live into Chat, Work, and Codex, with a macOS Screen context option that shares a screenshot of your frontmost window. The same release added multi-folder support to local projects, with primary and secondary folder designation. On the CLI side, 0.145.0 from July 21 remains the newest stable build — 0.146.0 is still cycling through alphas.

Cursor — the July 22 Cursor Router release added an Auto mode that routes each request to an appropriate model across three optimization profiles: Intelligence, Balance, and Cost. Admins can enable the router per team or group, restrict which modes members use, set the default, and allow or block underlying models. It is on by default for Teams plans and available across desktop, web, iOS, CLI, and SDK. That is a different bet than Anthropic's: Cursor is abstracting model choice away, while Opus 5 asks you to tune effort deliberately.

OpenCode — v1.18.4 (July 20) is still the current release, after a dense run of eleven builds earlier in the month. It added adaptive thinking controls for Kimi models on Anthropic-compatible providers, restored Azure Cognitive Services endpoint support, and rewrote the v2 prompt input for more reliable command, context, shell, attachment, and history handling.

Gemini CLI — v0.50.0 from July 8 remains the latest stable release, with tool registry discovery for automatic tool detection and registration. The larger story is migration: Gemini CLI has been folded into Google's unified multi-agent platform, Antigravity, and since June 18, 2026 the Gemini Code Assist IDE extensions and Gemini CLI stopped serving requests for the individual, Google AI Pro, and Google AI Ultra tiers. If you are still on those tiers, the move to Antigravity CLI is not optional.

GitHub Copilot — beyond Opus 5, the Copilot changelog had a busy week: the Linear cloud agent reached general availability and the GitHub MCP Server gained support for the next MCP specification version (both July 23), agent automation controls for Issues entered public preview, and a new metrics impact dashboard shipped July 22 for org owners tracking adoption depth and usage cohorts.

What to Actually Do This Week

• Grep your codebase for "type": "disabled" alongside any xhigh or max effort setting. That combination is now a 400, and it will fail at runtime rather than at build time.
• Revisit max_tokens on anything that ran without thinking on Opus 4.8. Thinking tokens now share that budget.
• Strip "verify your work" instructions out of prompts inherited from earlier models — Opus 5 does it unprompted, and the duplication costs tokens.
• If you run Claude Code subagent workflows, decide deliberately whether depth-3 nesting is what you want before your next long session, and set the spawn-depth variable if not.
• Copilot Business and Enterprise admins: the Opus 5 policy needs enabling before your team sees the model at all.

The Bigger Picture

The interesting detail in this launch is not the benchmark story — it is the distribution speed. A frontier model reached the default slot in one vendor's CLI and the model picker of a competitor's IDE ecosystem inside 24 hours, at unchanged pricing. Model swaps have become routine plumbing rather than events.

Which makes the breaking change the real news. When a model can arrive as a silent default across your entire toolchain in a day, the failure mode shifts from "we did not adopt fast enough" to "we adopted before we read the migration guide." The 400 error on disabled thinking at high effort is the cheap version of that lesson. Read the behavior changes before the upgrade reads them to you.

References and Further Reading

• Claude Code changelog — full release notes for v2.1.219 and v2.1.220
• What's new in Claude Opus 5 — specs, behavior changes, and pricing
• Claude models migration guide — step-by-step Opus 4.8 to Opus 5 migration
• Prompting Claude Opus 5 — patterns for verbosity, delegation, and self-verification
• Effort parameter documentation — the low-to-max reasoning ladder
• Fast mode — access, supported models, and the $10/$50 pricing
• Claude Opus 5 is now available in GitHub Copilot — plans, surfaces, and admin policy
• OpenAI Codex changelog — release 26.715 and Codex CLI notes
• Cursor changelog — Cursor Router and v3.11
• Codex CLI releases — stable and pre-release builds on GitHub
• OpenCode releases — v1.18.4 and the July build series
• Gemini CLI releases — v0.50.0 and the Antigravity migration
• GitHub Copilot changelog — the full July release stream