Your AI Coding Bill Tripled. Here’s Where Developers Went.

For two years the deal was simple: rent the smartest model in the world by the token and let it write your code. In June 2026 that deal is quietly falling apart. As flagship coding APIs raise prices faster than they raise capability, a growing group of developers is doing something that sounded absurd a year ago, moving their AI agents off the cloud and onto their own laptops. A widely shared June 15 field guide on agentic coding with local models lays out exactly how, and why the economics have flipped.

The math that broke: why cloud coding got expensive

The trigger is pricing, not ideology. GitHub Copilot has moved from a predictable credit model to usage-based billing, and the models that were previously bundled for free are no longer free. Because GitHub effectively resells tokens from the underlying model providers, every upstream price increase lands on developers with an added markup, so the squeeze feels sharper there than at the source.

The upstream numbers are the real story. Google’s Gemini Flash 3.5 costs roughly three times its Flash 2.5 predecessor. OpenAI’s GPT-5.5 runs about three times the price of GPT-5. Anthropic’s Claude is the exception that proves the rule: it was already priced so high that the latest move was actually downward. Across the board, the frontier labs are shipping models whose performance improvements are not keeping up with their price increases, which breaks the long-standing assumption that next year’s model is both better and cheaper.

The reseller premium

For a solo developer running an agent that loops through dozens of tool calls per task, usage-based pricing turns a fixed monthly bill into an unpredictable variable cost. That uncertainty, more than any single price tag, is what is driving experimentation with alternatives. When the agent is the one spending your tokens, every retry, every long context window, and every verbose system prompt has a dollar value attached.

The local stack: what running AI on your laptop really means

Running a coding agent locally is not a single download. It is a four-part stack. The harness is the deterministic code that wraps the model, the part you already know as VS Code Copilot, the Copilot CLI, or newer agents like Pi. The model is the neural network weights file, where quantization settings such as Q8 or Q4 trade memory for fidelity, much like image resolution, in formats like GGUF or MLX. The runtime is the inference engine, most commonly Llama.cpp, with Apple’s MLX, browser-friendly ONNX Runtime, and the server-grade vLLM as alternatives. Finally the model manager, typically Ollama, LM Studio, or the open-source Jan, loads and serves the weights through an OpenAI-compatible API that the rest of your tools already speak.

The most important idea in the whole approach is the deterministic harness. A weaker local model left to its own devices will hallucinate APIs and leave tasks half-finished. But wrap it in good tooling, linting, automated tests, an AGENTS.md instruction file, and explicit tool definitions, and its effective coding quality can improve by as much as 6x. In other words, much of the gap between a small open model and a frontier model is closed not by a bigger brain but by better guardrails around a smaller one.

The VRAM tricks that make it usable

The practical bottleneck is memory and speed. A long context window eats VRAM fast: at the maximum 262,144 tokens a typical setup needs about 25.74GB, while the author’s preferred 150,000-token window needs 22.45GB. A neat trick is KV cache quantization, setting the key cache to Q8_0 and the value cache to Q4_0, which keeps keys at higher resolution than values and cuts the GPU memory requirement from 28.75GB down to 22.45GB. The catch is responsiveness: cold starts add 10 to 30 seconds on the first request, the very first agent prompt can take 2 to 5 minutes as the harness sends its large system prompt, and anything under 10 tokens per second is too slow to tolerate for real coding.

Who actually wins: Gemma 4 versus the Chinese open models

On the open-weight side, Google’s Gemma 4 family is the workhorse. It spans an E2B edge model at 2B parameters that runs almost anywhere but risks hallucinations, an E4B that doubles it and makes a cheap starting point, a 12B that natively understands images for front-end and visual coding, a 26B-A4B mixture-of-experts model that activates only 4B of its 26B parameters and fits comfortably on an 8 to 12GB-VRAM card, and a 31B that is Google’s largest open weight but is not MoE and crawls at 1 to 2 tokens per second on a thin AMD APU. Quantization-aware-training variants squeeze memory further while holding quality. For most developers the 26B-A4B is the sweet spot.

The geopolitical twist

Here the field guide makes a pointed observation: what U.S. labs release for free is not their best. OpenAI’s gpt-oss-20b is dated, and Anthropic publishes no open weights at all, leaving Gemma 4 as the only serious American open model. The strongest open coding models on the HuggingFace leaderboards now come from Chinese labs, Qwen, Kimi, GLM, and DeepSeek among them. The author’s own June 9 update is the clearest sign of where this is heading: after months on local models, they adopted DeepSeek V4 Pro, which they describe as nearly matching Claude Opus 4.8 while offering roughly 5x the context window at an estimated 17 to 86 times lower cost.

That does not mean local wins outright. The honest conclusion is hybrid: reach for a hosted model like DeepSeek V4 Pro on hard, complex tasks, and keep a local model for simpler work, offline use, and anything privacy-sensitive. The frontier still wins on raw intelligence; the open stack wins on cost, control, and not leaking your codebase to a third party.

What this means for teams and the people paying the bill

For engineering leaders, the takeaway is that “which model do we use” is becoming “how do we route.” The emerging default is a tiered setup: a cheap or free model, possibly through a unified router like OpenRouter with a hard 1 dollar per month guardrail and new API keys capped at zero max credit, handles routine work, while a premium model is reserved for the genuinely hard problems. That structure caps the unpredictable spending that usage-based pricing introduced, and it keeps a fallback ready for the day a provider raises prices or deprecates a free tier.

There is also a quieter argument in the guide worth sitting with as a manager. Working with a weaker model forces the developer to stay engaged, to read the output, to keep judgment in the loop rather than rubber-stamping a confident answer. In knowledge work, the author argues, slow is fast, and the skills you build wrestling a small model on hard mode transfer directly to using the big tools more effectively. For a CEO weighing AI spend, the cheapest line item may also be the one that keeps your team sharp.

Related

• The Startup Betting Your AI Lab Will Betray You
• The Open-Weight Model That Coded Alone for 24 Hours
• Your AI Agents Keep Failing. This $24M Startup Knows Why.
• Microsoft Just Put AI Inside Windows Itself

Sources

1. Alex Ewerlof Notes – Local LLMs for agentic coding
2. GeekNews – Using local LLMs for agentic coding (2026-06-15)
3. TechCrunch – Can tech companies learn to love cheaper AI models?