Three vendors repriced agent intelligence in a single day
OpenAI laddered GPT-5.6 across three price points, Grok 4.5 went public at $2 and $6, and Meta metered its first API at a quarter of rival rates. July 9 reset the market.
- GPT-5.6 went GA July 9 as a three-tier ladder: Sol at $5 input / $30 output per million tokens, Terra at $2.50 / $15, Luna at $1 / $6.
- Grok 4.5 opened to the public the same day at $2 / $6, with a 500K context window and cached input at $0.50 per million.
- Meta launched Muse Spark 1.1, its first metered API, at $1.25 / $4.25 with a 1M-token context, roughly a quarter of rival rates.
- Three price books rewritten in 24 hours, all aimed at the same buyer: whoever pays the token bill under an agent.
The price of the intelligence underneath the agent economy was rewritten three times on July 9, by three different vendors, in three different directions. Taken together it is the fastest one-day repricing this market has seen.
The ladder
OpenAI's GPT-5.6 general availability came as a family: Sol for the hardest work at $5 per million input tokens and $30 per million output, Terra pitched as 5.5-class performance at half the cost at $2.50 and $15, and Luna for high-volume pipelines at $1 and $6. An Ultra mode runs four agents in parallel by default, a spend-more lever built directly into the product. The ladder formalizes what builders were already doing by hand: routing cheap steps to cheap models and saving the expensive one for the steps that matter.
The floor
Below the ladder, the floor dropped twice. Grok 4.5 went public at $2 input and $6 output with a 500K context window, cached input at 50 cents, and a configurable reasoning-effort dial. Elon Musk's pitch was aimed squarely at Anthropic: "It is an Opus-class model, but faster, more token-efficient and lower cost."
Then Meta arrived. Muse Spark 1.1, per CNBC, is the company's first metered API, priced at $1.25 input and $4.25 output with a 1M-token context window, closed weights, and an explicit agentic-coding focus. Meta AI chief Alexandr Wang called it the company's "strongest model for agentic and coding work yet." A fourth hyperscaler just entered the market on price, and the entry point is a quarter of what the incumbents charge.
What it does to agent margins
Every agent-as-a-service business is a spread between what the customer pays per outcome and what the vendor pays per token. We watched that spread split in two directions last week; July 9 compressed it from three sides at once. For vendors selling outcomes, like HubSpot's 50 cents per resolution, every one of these price cuts is margin. The buyers who benefit most are the ones with routers already in place. The ones who suffer are anyone who signed an annual commitment at last month's prices.