Coding agent harnesses proliferated in the first half of 2026, with terminal CLIs, IDE extensions, and cloud platforms all shipping agentic capabilities. These tools converged on the same protocols, splitting the coding agent into two distinct layers: the harness that manages the workflow, and the model that generates the code. That split changed which layer holds pricing power.

User’s choice

More than thirty coding agent harnesses shipped in the first half of 2026. They converged on a shared interface specification. Three protocol layers standardized in parallel: MCP for tool connections, SKILL.md for project-level context files, and OpenAI-compatible API endpoints for model access. A model served through any compliant endpoint works in all of them. OpenCode, the most-starred open-source agent on GitHub, illustrates the result. The MIT-licensed harness crossed 180,000 stars and 7.5 million monthly developers by connecting to more than 75 model providers through a single interface, treating the model as a swappable variable.

The leading harnesses now differentiate on autonomy philosophy, since model access has standardized. Claude Code gates every file edit on developer approval. Codex CLI runs unattended for hours in its Goal mode, delegating entire tasks to a cloud sandbox. OpenCode offers a plan-then-build split with full provider neutrality, feeding Language Server Protocol diagnostics back into the model loop after each edit. All three accept the same model connections. A developer who switches from GPT-5.5 to a LongCat or local Qwen instance changes a configuration line.

The no-name undercut

On June 30, Meituan revealed that LongCat-2.0, a model the Chinese delivery and local-services company had built for agentic coding, had been running on OpenRouter under the alias “Owl Alpha” since late April. Developers had driven the anonymous model to first place on the Hermes Agent workspace, second on Claude Code deployments, and third across OpenClaw environments, all ranked by monthly call volume, with no brand recognition attached.

LongCat-2.0 carries 1.6 trillion total parameters in a Mixture-of-Experts layout that activates roughly 48 billion per token, with a native one-million-token context window sustained through a sparse attention mechanism that reduces scaling from quadratic to linear. Meituan claims that the full training run, spanning more than 30 trillion tokens on a 50,000-card cluster of domestically produced accelerators, completed without NVIDIA hardware. Standard API pricing sits at $0.75 per million input tokens and $2.95 per million output tokens, with cached context reads free. GPT-5.5 charges $5 and $30 for the same units, and Claude Sonnet 5 charges $2 and $10.

When the harness abstracts the model, switching cost drops to near zero, and the remaining differentiator between a $30-per-million-output-token provider and a $2.95 one narrows to measured output quality on the task at hand. LongCat-2.0 posted a SWE-bench Pro score of 59.5 against GPT-5.5’s 58.6, a gap of less than a point, with both figures vendor-reported. For high-volume agent workloads that run thousands of coding sessions per month, a tenfold pricing gap at near-equivalent benchmark performance becomes difficult to justify.

Going cold turkey

NVIDIA’s DGX Spark, shipping since June, places a petaflop of AI compute and 128GB of unified memory on a desktop for $4,679. The system runs open-weight models to 200 billion parameters locally, and at cloud inference rates, a small development team recoups the hardware cost within a quarter, converting a recurring per-token expense into a one-time purchase with no API dependency.

Google’s Gemma 4 12B runs on a laptop with 16GB of RAM, combining vision and voice in a single local architecture. The same open-weight ecosystem that supplies the cheap cloud alternatives, including Qwen, DeepSeek, and LongCat, also supplies the models that this local hardware runs. Frontier providers now face simultaneous pricing pressure from cloud-hosted Chinese open-weight models that compress margins and from local inference hardware that eliminates the API call entirely, with harness agnosticism making both substitutions trivial to execute.

Separation of concerns

Frontier model pricing carried an implicit integration premium for as long as the model and the harness shipped as a single product, because switching providers meant rebuilding a workflow. Three developments dissolved that premium in the first half of 2026: the harness layer standardized around shared protocols, Chinese labs released competitive open weights at a fraction of the cost, and desktop hardware capable of running those weights shipped to developers. The coding agent remains a high-value product, but the value has migrated to the harness layer, where autonomy design, context management, and developer trust provide the differentiation that model access no longer can.

Subscribe now

AI Central

The Former Monopoly of Convenience