What's rising in the AI coding ecosystem?

GPT-5.6 is released with three tiers: Sol for agentic coding, Terra for balanced performance, and Luna for speed and low cost ^1056,1310. Sol is designed for long-horizon coding and agentic work ¹⁰⁵⁷, and Sol on Medium is the preferred setting for coding with GPT-5.6 ^1255,1312. ChatGPT Work, built on GPT-5.6 and Codex, transforms ChatGPT into a general work agent that completes projects across apps ¹⁰⁶⁹. GPT-5.6 Sol outperforms Claude and Gemini for solo builders due to benchmark leadership and the integrated Work/Codex interface ¹⁴²⁰. Sol is also favored over Fable 5 for autonomous code building, reviewing, and screen testing because of lower token usage and better autonomy ¹⁷⁹². ChatGPT Work is preferred over Claude for automating the middle steps of office work ¹⁴¹⁷. GPT-5.6 Sol demonstrates real-time performance in Blender at 750 tokens/s without acceleration (1972), and is preferred over Claude Code for 3D modeling because it can generate textures directly (1978).[11]

Claude Code's new desktop launch signals a shift from copilot to autonomous software engineer, with the IDE becoming the operating system for development ¹²⁷⁰. Claude Code originated from Anthropic's internal 'clide' tool ⁸⁵⁵. Version 2.1.203 adds background session auto-recovery and blocks unauthorized destructive actions ⁷⁹⁷. Version 2.1.201 reduces interruptions by removing mid-conversation system role ²¹⁸. The Artifacts feature turns sessions into live, interactive web pages with version history ²²³. Claude Cowork enables cross-device task continuation with background processing when the laptop is closed ⁷⁹⁴. Claude Code skills are evolving from static instructions to self-improving, benchmarked software (1976). A hidden TeamMateTool in v2.1.19 reveals a leader-worker system, pointing to built-in multi-agent capabilities (1987). The new Home tab unifies Chat and Cowork and allows scheduling Skills to run autonomously ¹⁰⁰⁹. Anthropic is turning internal tools into products, similar to Google's evolution ¹⁰⁰³.[10]

Multi-agent collaboration is on the rise: codex-plugin-cc lets Claude Code delegate tasks to Codex in the same terminal ¹⁴¹. A new tool enables monitoring and managing multiple coding agents from mobile TUI and web ⁴⁸⁴. Self-improving agentic routines where Claude prompts Claude are gaining traction ¹¹⁷². AI programming shifts from individual code completion to multi-agent collaboration, where Claude in Slack can fix bugs, write PRs, and send daily reports, making task decomposition and context management the new core skills ¹¹⁰⁸. Running multiple Claude Code sessions in TMux with agent teams and Git worktree isolation enables adversarial code reviews and parallel feature testing ³⁷⁶. Hermes Agent 0.18.0 Runtime introduces /moa and /learn commands ⁵²⁰, and Hermes Agent can learn and autonomously execute repetitive workflows after a single demonstration ³²¹.[7]

Cursor remains the best all-rounder for AI coding, while Claude Code excels at design and research, Codex offers useful plugins, and Opencode provides the best value ¹²⁶⁵. The Cursor $20 plan is the top value pick ¹³²². Interest in Cursor alternatives is rising following the SpaceX acquisition ¹⁰³⁰. Cursor's decision to restrict Sonnet 5 and GPT 5.5 to Max-only plans signals a shift toward higher-tier monetization, causing user backlash ⁸⁶⁰. Grok 4.5 is now available in Cursor at $2/M input and $6/M output tokens ⁹³⁰. OpenAI's Sites feature allows turning an idea into a live site that can be published and shared ¹¹⁷¹. The Browser Company's new desktop app introduces a fully functional in-app browser, a cloud browser for agents, and side chat integrating ChatGPT/Codex ¹¹⁷⁰. Cursor's iOS app enables building from anywhere with always-on cloud agents ³⁷⁹. Continue v1.2.24 shifts away from Hub slugs toward explicit model definitions in config templates ⁶²².[9]

Cost efficiency is driving model choices: GLM-5.2 and MiniMax-M3 are preferred over Opus 4.8 for cost-sensitive tasks, offering similar intelligence at 1/5 the price ³⁷³. Running a 671B model locally on a Mac Studio is now feasible, signaling a shift toward on-device inference for large reasoning models ¹⁵⁵⁵. The NVIDIA DGX Spark, with 128GB unified memory and 1 petaflop AI compute, enables running 200B parameter models locally without quantization, paying for itself in 10 months vs cloud subscriptions ²²⁹. To use Claude Fable 5 cost-effectively in Claude Code, delegate roles: Fable for overall design, Opus for heavy reasoning, Sonnet for execution and task management ¹²⁵. Claude Sonnet 5 is supported across the Anthropic, Bedrock, Vertex, Claude Code, SAP AI Core, OpenRouter, and Vercel AI Gateway providers ³⁶⁷. Anthropic's free workshops and official courses provide a structured path to building autonomous agents with Claude ^727,573.[7]

Anthropic's session on self-learning agents with memory and dreaming shows significant gains: Rakuten cut first-pass mistakes by 90%, Harvey saw 6x gain on legal benchmark ²³²⁷. Fable 5 is the best model for creative thinking, but GPT-5.6 Sol is almost as good and much cheaper, making it the better value ^{2315,2149,2059,1969}. Sol on Medium is preferred over GPT-5.5 xhigh for coding work due to being an upgrade ^{2312,2140,1959,1636,1474,1701,1549,1408,1312}. GPT 5.6 Sol is preferred over Claude Fable 5 for building, reviewing, and testing after initial decisions are made ^2255,2219. HyperFrames is preferred over custom Claude Code-built video editing tools because it is free, requires no asset imports, and can mimic 80% of editing from reference videos ²¹⁴⁴. Anthropic's model naming scheme (Haiku < Sonnet < Opus < Fable < Mythos) is ambiguous, with Fable's position relative to Opus uncertain ²¹⁴². Anthropic senior engineers released a 90-minute course on building a custom agentic harness with harness + loop architecture, now used by 90% of their engineers (2071).[19]

Local inference of a 671B parameter model on a single desktop is now feasible, bridging the gap between local and server-grade AI ²¹⁵⁵. The open-sourcing of obsidian-skills signals a trend where note-taking apps become agent-accessible file systems ¹⁸⁷⁰. Vibe coded projects fail for the same reasons as hand-coded ones: lack of value and marketing, not scaling issues ^1635,1550. The real AI acceleration is the feedback loop where models help build better models, not just raw intelligence gains ¹⁴⁸³. AI-enhanced browsers are coming to a close due to unsolvable security/privacy issues; AI should use its own separate browser ¹³¹⁵. Today's LLMs are less likely to default to building everything in React compared to last year, reducing the need to explicitly request no React ¹⁰⁹. As AI agents take on longer-running work, engineering focus moves to setting direction, reviewing work, and designing better systems ¹⁰⁴. The 'understand to participate' framing is a useful way to describe the cognitive debt problem when working with coding agents ²⁶⁰.

Opus 4.8 is only useful for front-end work, as GPT-5.6 Sol and Fable 5 are preferred for most tasks ¹⁸⁷⁹. GLM-5.1 references are being removed ⁹⁹⁸. AI-enhanced browsers are coming to a close due to unsolvable security/privacy issues ¹³¹⁵. The trend of defaulting to React for everything is declining ¹⁰⁹.[4]