GPT-5.6 resets the agent stack (2026)

Coverage window: July 3, 2026, 10:29 a.m. UTC-8 through July 10, 2026, 10:00 a.m. UTC-8.

GPT-5.6 became the default story for AI coding tools this week because OpenAI released Sol, Terra, and Luna for general availability on July 9, and the family landed the same day across OpenAI API, ChatGPT, Codex, GitHub Copilot, Devin, and Microsoft 365 Copilot. 1 2 3 The practical question for engineering leads is no longer which model won a launch-week benchmark. The question is which agent runtime, permission model, cost control, and migration path each team is willing to support.

Fast triage

Area	What changed	Follow-up for tooling teams
GPT-5.6 rollout	OpenAI released GPT-5.6 Sol, Terra, and Luna on July 9; Sol scored 53.6 on Agents' Last Exam and 80 on Artificial Analysis Coding Agent Index, while API pricing is $5/$30, $2.50/$15, and $1/$6 per 1M input/output tokens for Sol, Terra, and Luna. 1	Re-run internal agent benchmarks on Sol, Terra, and Luna before changing defaults. GitHub says all three models are available in Copilot across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, CLI, Cloud Agent, Copilot App, GitHub.com, and Mobile. 2
Cursor	Cursor v3.11 shipped Side Chats, conversation search, redesigned project/repo pickers, and cloud-agent hooks on July 10. 4 Cursor also released Grok 4.5 on July 8 as a mixture-of-experts model trained with SpaceXAI, with base pricing of $2/$6 per 1M input/output tokens and fast pricing of $4/$18. 5	Test whether Side Chats reduce context pollution in long-running agents. Review Grok 4.5 separately for non-coding knowledge work because Cursor positions it for software engineering, data science, finance, and legal work. 5
Copilot	GitHub added Codex as a public-preview agent provider in JetBrains IDEs on July 7, expanded agent customizations with hooks and MCP server management, and made Inline Chat generally available in JetBrains. 6 GitHub also expanded Kimi K2.7 Code to Copilot Business and Enterprise on July 7. 7	Update model policies and JetBrains rollout docs. Kimi K2.7 Code is the first open-weight model in Copilot's model picker, and GitHub hosts it on Microsoft Azure. 7
CLI agents	Claude Code shipped six releases from v2.1.201 through v2.1.206 in the window, and v2.1.206 explicitly mentions `claude-opus-4-8` for `/code-review` findings quality. 8 Codex CLI jumped from 0.142.5 stable to 0.144.1 stable on July 9. 9	Treat the CLI layer as production infrastructure. Claude Code, Codex CLI, Grok Build, Kimi Code, MiMo Code, ZCode, and CodeGraph all shipped operational changes this week. 8 10 11
Migration risk	Continue.dev's final v2.1.0-vscode release shipped on June 19 with a Cursor co-authored commit, and reporting says Continue cloud data will be permanently deleted on July 15. 12 13 Devin Desktop also continued the Windsurf rebrand path after Cascade reached end-of-life on July 1. 14 13	Notify Continue users before July 15 if they relied on cloud sync. Search internal docs, scripts, and onboarding material for Cascade references and update them to Devin Desktop or Devin Local where appropriate.

Model routing became the platform question

GPT-5.6 matters because it arrived as a model family and as a distribution event. OpenAI described Sol as the flagship model, Terra as the balanced model, and Luna as the cost-efficient model; the same announcement introduced Programmatic Tool Calling, a multi-agent beta, and ultra mode with four parallel agents by default and up to 16 configurable agents. 1 Devin made the three GPT-5.6 models available in Devin Desktop and Devin CLI on July 9, with Devin Cloud preview access through July 16. 3

The buyer-side implication is model policy drift. GitHub Copilot now exposes GPT-5.6 Sol, Terra, and Luna across its IDE, CLI, web, cloud-agent, mobile, and app surfaces. 2 GitHub also added Kimi K2.7 Code to Business and Enterprise plans, which means Copilot administrators now have to decide how an open-weight model fits into enterprise policy rather than treating it as a consumer-plan experiment. 7

Cursor is moving in a different direction: model supply is becoming part of the IDE vendor's own product surface. Cursor released Grok 4.5 on July 8 as its first general-purpose model for domains beyond software engineering, and the company said the model was trained on trillions of tokens of Cursor data from codebase interactions and developer-agent interactions. 5 Cursor also said CursorBench was excluded from the Grok 4.5 evaluation because an earlier snapshot of the Cursor codebase was accidentally included in training data. 5 That exclusion is the right instinct: teams should ask whether benchmark data is clean before accepting any vendor's coding-agent score.

Claude Code and Codex CLI show the same routing problem from the terminal side. Claude Code v2.1.206 includes a release-note item for improved /code-review findings quality on claude-opus-4-8, which confirms Opus 4.8 is an available Claude Code model option. 8 Codex CLI 0.144.0 added clearer Amazon Bedrock model names for the GPT-5.6 family and variants, which points to GPT-5.6 catalog preparation outside OpenAI's own hosted surfaces. 9

Agent control surfaces got more operational

Cursor v3.11 is the cleanest example of agent work becoming a managed workflow. Side Chats let users open /side or /btw to run a durable agent conversation alongside the main chat, with context from the main conversation and the ability to at-mention the side chat back into the main thread. 4 Cursor also added Cmd+K transcript search in the Agents Window, Cmd+F within a conversation, redesigned project/repo pickers with scoped search across local, cloud, and remote-machine contexts, and five cloud-agent hooks: beforeSubmitPrompt, afterAgentResponse, afterAgentThought, stop, and subagentStart. 4

GitHub's JetBrains update is similar in shape. Codex can now be selected as an agent provider from the Copilot Chat panel in JetBrains IDEs, while agent customizations now include hooks, MCP Marketplace browsing, HTTP and command MCP server types, and workspace-level MCP configuration through .github/mcp.json. 6 GitHub also added permission modes for Claude agent sessions in JetBrains and approval settings for Copilot CLI sessions, including a preview Autopilot mode that auto-approves everything, including clarifying questions. 6

Claude Code's releases were less headline-driven, but they changed the operating surface. Version 2.1.206 added /cd directory path suggestions, a /doctor checkup that can trim CLAUDE.md, /commit-push-pr push automation, Anthropic public gateway login support, external worktree confirmation, and automatic background-agent updates after a Claude Code upgrade. 8 Version 2.1.205 blocked auto mode from modifying the session transcript file, improved agent-view status display, and reserved the Claude Browser MCP server name. 8

Codex CLI's 0.144.0 release also points toward stricter control. The release added a writes app-approval mode, interactive MCP authentication without an experimental opt-in, a reset-credit redemption picker for usage limits, and a warning when Ultra reasoning starts high-concurrency multi-agent work. 9 Grok Build kept a faster cadence, shipping 11 releases from v0.2.84 through v0.2.94; the window added Agent Dashboard via /sessions, a /goal command, Linux voice dictation, /auto classifier permission mode, background subagents by default, and local-directory plugin marketplace support. 10

Kimi Code, ZCode, and MiMo Code filled in the same pattern from different directions. Kimi Code 0.23.0 through 0.23.5 added Preserved Thinking, archived sessions, progressive tool disclosure, Extra Usage display, image compression improvements, and better recovery from unsupported image formats. 11 ZCode shipped five releases from v3.3.0 through v3.3.4, adding hardware acceleration controls, team packages, MCP OAuth for local development, background tasks, SSH plugin sync, and Windows performance fixes. 15 MiMo Code v0.1.5 added 14 builtin skills, a Windows native installer, experimental Orchestrator mode, cron scheduling, vision-aware image paste, exact-string edit matching by default, and a second permission prompt for irreversible deletions. 16

Migration and monitoring items

Continue.dev is the most time-sensitive migration item. The final v2.1.0-vscode release shipped on June 19 with the commit message chore: vscode 2.1.0 Co-authored-by: Cursor, the repository is now read-only, and reporting says Continue's conversation history, saved configs, and team settings will be permanently deleted on July 15. 12 13 Cursor has not published an official acquisition post in the research window, so treat the July 15 data-deletion item as the operational priority rather than waiting for integration details. 13

The Windsurf-to-Devin transition also needs cleanup. Windsurf completed its rebrand to Devin Desktop on June 2, the bundled agent is now Devin Local, and Cascade reached end-of-life on July 1 according to third-party reporting. 13 Devin Desktop v3.4.22 on July 4 added cloud-session auto-reconnect, image copy from the chat context menu, default agent mode for new users, orphaned Devin ACP agent cleanup, autonomous-mode diffs, skill permissions: frontmatter, sandbox.excluded config, and ACU usage in /usage; v3.4.27 on July 7 fixed diff viewing in autonomous mode. 14

Some monitored tools had little or no product delta. v0.app had no July changelog entry in the window, with its latest changelog month still June 2026. 17 Aider remains at v0.86.0, with its latest release dated August 9, 2025. 18 JetBrains Junie still exposes the same public landing-page shape, with AI Pro at $100/year, AI Ultimate at $300/year, AI Enterprise at $720/year, and no public changelog or standalone CLI surfaced in the research window. 19

The measurement and context stack is filling in

CodeGraph is the most directly useful infrastructure signal for agent context. CodeGraph v1.3.0 added Nix and ArkTS/HarmonyOS support on July 7, v1.3.1 fixed large-codebase indexing failures on July 9, and v1.4.0 on July 10 cut one slow-storage Java-project indexing benchmark from more than 25 minutes to under 1 minute. 20 For teams that feed code graphs into agents, indexing reliability matters as much as model choice.

Weave Router is still small but active. The project reached 832 stars and shipped a v0.2.5 bump in the window, with commits for API-key search, policy-router candidate resolution, a $5 subscription overdraft cap, HMM router feedback forwarding, DeepInfra provider removal, Kimi K2.7 fallback binding, and telemetry for routing strategy and route ID. 21 Router behavior should be reviewed like any other production dependency because it decides which provider sees which prompt.

Two community signals point to the same management problem. Uvik's 2026 developer usage report says 90% of developers use at least one AI tool at work, with GitHub Copilot at 29% share, Cursor and Claude Code at 18% each, and Codex at 3% in the cited JetBrains January 2026 data. 22 Developers Digest describes DevOS as a Journi platform for measuring AI coding tool ROI across Claude Code, Cursor, and other agents, with utilization, impact, and cost dimensions. 23 Treat those figures as market signals, not neutral benchmarks, but the direction is clear enough: teams need their own measurements.

The skills ecosystem is also becoming a supply-chain surface. The tech-leads-club agent-skills repository reports 4.9k stars, 439 forks, 1,069 commits, support for 20-plus agents, an npm package, an MCP server, and Snyk-backed scanning claims for marketplace skills. 24 TrueFoundry's Skills Registry still lists 1,023 skills, 68 publishers, and 11 categories in this checkpoint. 25 Skills now need the same review path as MCP servers and editor extensions.

Actions before next Friday

Rebuild the model allowlist. Include GPT-5.6 Sol, Terra, and Luna; Kimi K2.7 Code in Copilot Business and Enterprise; Cursor Grok 4.5; Claude Opus 4.8 in Claude Code; and GPT-5.6 variants in Codex/Bedrock catalogs. 1 7 5 8 9
Test agent approval modes against real repositories. Copilot Autopilot, Codex writes, Claude Code manual/background-agent behavior, Cursor cloud-agent hooks, and Grok Build /auto all change where human approval enters the workflow. 6 9 8 4 10
Close migration loose ends. Export or retire Continue cloud data before July 15, replace Cascade references with Devin Desktop or Devin Local, and keep v0.app, Aider, and Junie on lightweight monitoring until they publish new release signals. 12 13 17 18 19
Put cost and ROI instrumentation next to model testing. GPT-5.6 pricing, Cursor Grok 4.5 pricing, Codex usage-limit reset credit redemption, Kimi Extra Usage display, and DevOS-style utilization/impact/cost reporting all point to the same admin requirement: coding-agent rollout needs a measurement plan before usage scales. 1 5 9 11 23

Cover image: image from OpenAI.

GPT-5.6 resets the agent stack