Kimi K2.7 Code is generally available in GitHub Copilot
GitHub Copilot has made Kimi K2.7 Code, an open-weight model, generally available as a selectable, lower-cost option. While this offers more choice and potential cost savings, the announcement has largely been overshadowed by widespread user dissatisfaction with Copilot's recent shift to token-based billing and perceived price hikes. The community's discussion reveals a growing exodus to alternative, often self-hosted or more transparently priced, AI coding tools.
The Lowdown
GitHub Copilot has officially announced the general availability of Kimi K2.7 Code, an open-weight model, within its ecosystem. This marks a significant move by offering users a selectable, lower-cost option for their coding workflows, hosted directly by GitHub on Microsoft Azure.
- Kimi K2.7 Code is the first open-weight model integrated into the Copilot model picker, providing greater flexibility.
- Billing for this model is usage-based, aligning with provider list pricing, which GitHub highlights as a lower-cost alternative.
- The rollout is gradual, starting with Copilot Pro, Pro+, and Max plans, and will eventually extend to Business and Enterprise subscriptions.
- For Business and Enterprise users, Kimi K2.7 Code is off by default, requiring administrators to explicitly enable it after reviewing security and compliance implications.
This release aims to diversify model options for Copilot users, addressing demands for more transparent and potentially cost-effective AI assistance, although its impact is evaluated within the context of recent, controversial pricing changes to the broader Copilot service.
The Gossip
Copilot's Costly Conundrum
Many users expressed significant frustration over GitHub Copilot's recent pricing changes, particularly the switch to token-based billing which led to perceived massive price hikes. What was once considered great value, often exhausted within days or resulting in costs many times higher than previous subscription models, has prompted a mass migration to alternative services like Claude Code or self-hosted solutions. While some comments acknowledge the necessity for sustainable pricing or argue Microsoft is merely passing on costs, the overwhelming sentiment is one of disappointment and feeling priced out.
The Rise of Resident Robots
A strong theme emerged around users abandoning cloud-based AI due to fluctuating prices and unreliable service, instead embracing self-hosted or open-weight models. Many are setting up local AI rigs, often using models like Qwen3.6, for greater control, predictable costs, and the ability to customize. The availability of Kimi K2.7 Code is seen by some as a positive step towards offering open-weight models from a 'trusted provider,' but the broader discussion highlights a growing preference for local, stable, and extensible AI solutions.
Harnessing the Haze of AI
The discussion delves into the importance of the 'harness' or integration layer that wraps around AI models, affecting their perceived performance and usability. Users debate whether Copilot's harness is inferior to alternatives like Claude Code or custom setups, leading to 'dumber' results even when using the same underlying models. Some suggest issues with system prompts or tooling, while others defend Copilot's CLI and multi-model orchestration capabilities. The complexity of tool calls and contextual understanding within different harnesses is a key point of contention.
Benchmarking Brains and Billing
A curious side discussion surfaced regarding model performance using unconventional benchmarks, specifically tic-tac-toe. Users illustrated how even expensive models like GPT 5.4 mini and Kimi K2.7 played 'stupidly' at simple games, demonstrating the often-poor performance of LLMs in such tasks. This led to a broader conversation about the actual value for money, comparing model costs against their ability to solve tasks, and exploring whether cheaper models, despite initial errors, could become more cost-effective if they eventually succeed or are used strategically within multi-model workflows.