OpenAI and Anthropic cut frontier model API prices in same-day releases

OpenAI and Anthropic shipped new frontier models within an hour of each other on 22 September 2026, and both arrived with deep API price cuts — OpenAI's new models at roughly half of their predecessors' rates, and Anthropic's flagship down 20%. According to dev.to, the near-simultaneous timing has led Simon Willison to describe the situation as a price war among frontier-model vendors, and it lands directly on developers' invoices.

OpenAI's GPT-6 line-up

GPT-6 Sol and GPT-6 Luna launched at about half the price of their GPT-5.6 counterparts. Luna is now $0.10 per million input tokens and $0.50 per million output tokens, a level dev.to describes as among the cheapest OpenAI has ever shipped, with only the smaller Nano-tier models going lower. Sol came in at $2 and $10 per million input and output tokens respectively — the same figure GPT-5.6 Terra previously carried, which per Willison removes any remaining reason to stick with Terra.

The apparent discount is larger than the headline numbers suggest. Willison notes that GPT-5.6 pricing was due to rise by 25% in November, so GPT-6's rates are being measured against a baseline that was about to climb.

Anthropic's answer

Claude Opus 5.5 arrived with a 20% reduction versus Opus 5.0 through 4.8, dropping from $5/$25 to $4/$20 per million input and output tokens. Cache-read pricing fell by 60%, a change dev.to flags as particularly relevant for agentic workloads: Willison estimates that more than 90% of input tokens in longer conversations are served from cache, so cache rates increasingly define real-world spend.

Anthropic also said cheaper Sonnet 5.5 and Haiku 5.5 models are on the way, and according to the report specifically pointed to Haiku's need to compete with GPT-6 Luna on price.

Grok sits in between

Grok 4.7, released the day before, now occupies the middle of the field. It undercut GPT-5.6 Sol substantially at launch, but GPT-6 Sol's new rates have roughly erased that lead on input pricing.

Benchmark results were mixed

Willison's standard pelican-drawing test produced a cautionary data point. Claude Opus 5.5 running at its "max" reasoning setting failed twice, each time hitting the 128,000-token output cap while still deliberating over a simple SVG request. Every failed attempt consumed about 20 minutes and $2.56. Despite that, he has since moved his own default coding tools to a combination of GPT-6 Sol and Claude Opus 5.5, and upgraded a public Datasette demo to GPT-6 Luna.

Why it matters

Per-token pricing at the frontier has quietly become the main competitive axis. When two leading labs cut costs on the same day, every application built on their APIs gets cheaper to run, and procurement decisions settled a week ago reopen overnight. The 60% cache-read reduction matters most for agents: as token-heavy, multi-turn workflows become the default, cached input — which Willison says can exceed 90% of tokens in long conversations — increasingly determines the actual bill.

The same-day timing also shows how quickly this market now moves. Anthropic is already positioning an unreleased Haiku model against Luna's price point, and Grok 4.7's launch advantage lasted barely a day. For developers, though, the pelican benchmark is a useful reminder that unit price is only half the equation: a model that burns its entire output budget reasoning about a trivial request can end up costing more than a pricier one that answers straight away.