Simon Willison on the latest model price war
A practical snapshot of LLM API competition: frontier-ish models are getting cheaper fast, while expensive high-reasoning modes still need operational guardrails.
Logged at IST: 2026-09-23 09:51 IST
What it is: Simon Willison's early read on a cluster of model launches: Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol and GPT-6 Luna.
Gist: The most important part is pricing. Simon notes that GPT-6 Luna is half the price of GPT-5.6 Luna, at $0.10/M input and $0.50/M output, while GPT-6 Sol also halves the GPT-5.6 Sol tier. Claude Opus 5.5 gets a smaller but still meaningful cut from the earlier Opus price level, with input at $4/M and output at $20/M, plus much cheaper cache reads for long agentic conversations.
The qualitative picture is more mixed. Simon is moving Codex and Claude Code defaults to GPT-6 Sol and Claude Opus 5.5, and has upgraded his Datasette Agent demo to GPT-6 Luna. But his pelican SVG test exposed a failure mode for Opus 5.5 at max reasoning: it over-thought the task until it hit the 128,000 output-token limit without returning a finished SVG, twice.
The useful takeaway is not just that models are getting cheaper. It is that the operating envelope is moving quickly: cheap models may become good enough for more production agent paths, while expensive reasoning modes still need explicit cost, latency, and completion guardrails.
Newsletter angle: Good pricing/agent-ops item on how model defaults can change when capability moves down the cost curve, and why high-reasoning modes still need reliability controls.