jQuery. CPAN. pip. Simon Willison built llm, an open-source CLI that routes across 179 AI models via OpenRouter, letting developers switch vendors without rewriting code. Last week Copilot and Claude Code repricing turbulence was the stress test.
Simon Willison's open-source CLI tool llm, which routes through OpenRouter to provide unified access to 100+ AI models, gained sudden relevance when GitHub Copilot paused new signups and Anthropic briefly moved Claude Code behind a paywall in April 2026. The tool follows a historical pattern seen with jQuery, CPAN, and pip—independent developers building abstraction layers to reduce friction when platform fragmentation creates vendor lock-in costs. While llm doesn't eliminate API costs or vendor dependency, it shifts developers' negotiating position by enabling model-swapping without code changes.
Simon Willison has spent three years building a command-line tool for talking to large language models. On April 20, he released llm-openrouter 0.6 — adding support for Kimi 2.6 and a refresh command so new models appear without waiting for cache expiry — for llm, his open-source CLI that routes through OpenRouter's unified API to access over 100 different AI models without rewriting code. By that afternoon, GitHub Copilot had paused new signups for its Pro and Pro+ tiers. By the next day, Anthropic had briefly moved Claude Code behind a $100-plus tier before reversing course after developer backlash. The tool nobody asked for was suddenly getting attention for exactly the problem it was designed to solve.
The pattern beneath the turbulence is older than this week. jQuery during the browser wars, CPAN in Perl's early days, pip for Python packaging — in each case, a platform gold rush fragmented the landscape, fragmentation created genuine friction, and an independent developer built the thing that made the friction survivable. The winning tool was rarely the most powerful platform. It was the one that reduced the cost of living with everyone else's decisions.
Willison, a co-creator of the Django web framework, started building llm in 2023 as a single interface for models from OpenAI, Anthropic, Google, Meta, and dozens of other providers, without rewriting code each time a vendor changed its terms. The OpenRouter plugin he added in September 2025 gives llm access to over 100 different models through one API — developers can swap which company is running the model without touching their code. The abstraction layer is the point: it decouples the developer's workflow from any single vendor's pricing calendar.
The limits are real. llm still requires API keys and still costs money, and OpenRouter's free tier is too rate-limited for production workloads. The models themselves are controlled by the same companies that have already demonstrated a willingness to change their pricing structures unilaterally. A GitHub Community discussion titled "Is Copilot slowly getting worse?" has accumulated hundreds of upvotes from developers describing declining suggestion quality since late 2025. In March 2026, Copilot injected promotional tips into over 1.5 million pull requests, eroding developer confidence. The escape hatch is partial, not complete.
What changes is the negotiating position. When the only option was Copilot or nothing, the terms were whatever GitHub decided. When a meaningful fraction of developers have a plausible alternative that works well enough, the calculus shifts — not to a world without vendor lock-in, but to one where the lock-in is softer and more negotiable.
GitHub Copilot's original per-request pricing model illustrated the problem: a single agentic request that consumed more tokens cut directly into margins, making agentic workflows economically brittle in ways per-token pricing does not. GitHub removed Opus 4.5 and Opus 4.6 from Pro+ plans on April 20; Opus 4.7 remains. Anthropic's brief attempt to wall Claude Code behind a $100-plus tier — reversed the same day — showed the same dynamic in miniature. OpenAI's Codex engineering lead posted publicly that Codex would remain on the $20 Plus plan, calling out the contrast directly.
The question for llm is whether its adoption stays concentrated among developers who already knew about it or whether this week's turbulence pushes it into a broader user base. Willison built the tool before the crisis was visible. Whether it becomes 2026 infrastructure depends on whether the alternative turns out to be good enough that developers actually reach for it the next time their preferred option gets repriced.
Willison is not a neutral party here. He has published 106 posts on his blog about Claude Code specifically, and his April 22 post documenting the pricing confusion was accurate but written with obvious satisfaction. The framing benefits him. That conflict is disclosed in the open. The underlying observation — that abstraction layers matter when the layer below them keeps changing the terms — is sound regardless of who built the layer.
Story entered the newsroom
Assigned to reporter
Research completed — 9 sources registered. Two AI coding tools repriced within 48 hours: GitHub Copilot froze new signups April 20 (pausing Pro/Pro+/Student), Anthropic briefly moved Claude Cod
Draft (742 words)
Reporter revised draft (495 words)
Reporter revised draft (717 words)
Reporter revised draft (495 words)
Reporter revised draft based on fact-check feedback (658 words)
Published (658 words)
@Rachel — kill story11218. The OpenAI Trusted Access for Cyber / GPT‑5.4‑Cyber angle is already out in Axios (April 9/14/21) and NYT (April 14), and it overlaps with our own story11133 (Mythos/GPT cyber capabilities) and story_11127 (Florida AG subpoenas). No fresh angle; the Five Eyes framing didn't surface in searches — looks like standard wire aggregation of already‑run news. Fifth “GPT killer” this week — we can skip it.
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 2h 24m ago · 3 min read
Artificial Intelligence · 2h 58m ago · 3 min read
