The Week "Aggressive Pricing" Stopped Meaning Anything

Between July 8 and July 9, Meta, xAI, and OpenAI each shipped a new model, and each one led with a pricing pitch. Meta called Muse Spark 1.1 ultra-low cost. Elon Musk called Grok 4.5 "Opus-class, but cheaper." OpenAI's own coverage described GPT-5.6 Sol's pricing as midrange. Put the three rate cards next to each other and the adjectives stop agreeing with the numbers, or with each other.

Three companies, three definitions of cheap

Here's what actually shipped, in dollars per million tokens, input then output:

• Meta Muse Spark 1.1 (July 9): $1.25 input / $4.25 output, pitched at coding agents, $20 in free credits for new API accounts
• Grok 4.5 (July 8, xAI): $2 input / $6 output, marketed as undercutting Claude Opus 4.7's $5 input / $25 output by roughly three-quarters on the output rate
• GPT-5.6 Sol (July 9): $5 input / $30 output, the top tier of a three-model family that also includes Terra at $2.50 / $15

Sol is the one worth sitting with. Several outlets described its pricing as midrange, but $30 per million output tokens is *more* than Opus 4.7 charges today, not less. A launch article can call a price whatever it wants. The rate card doesn't care what the headline says.

Opus-class model, but faster, more token-efficient and lower cost — Elon Musk, describing Grok 4.5

The reasoning-token catch nobody put in the headline

Muse Spark 1.1 is a reasoning model, and Meta bills its internal thinking tokens at the output rate. That detail didn't make most of the launch coverage, but it matters more than the sticker price. A model that reasons for two thousand tokens before it writes ten lines of code costs you for those two thousand tokens, at $4.25 per million, whether you asked for the reasoning or not.

We've made this point before in does thinking effort actually matter: effort is a dial, not a fixed cost, and turning it up quietly moves the bill regardless of which logo is on the model card. A cheap per-token rate attached to a model that reasons heavily on every prompt can still lose to a pricier model that reasons only when the task calls for it.

What the arena shows instead of a press release

This is the whole reason the coding arena runs the same one-shot prompt across models and shows the actual output, not a vendor's benchmark slide. Blind votes tell you which completion a reader preferred without knowing which company made it. Token counts from the real run tell you what that preference actually cost, instead of what the launch post claimed it would cost.

We covered a version of this arithmetic in Claude vs DeepSeek: what $50 per million tokens buys, and the conclusion holds again this week: the rate card is the start of the calculation, not the end of it. Muse Spark, Grok 4.5, and Sol are all new enough that we're still building out clean head-to-head runs for them; once they're live, the receipts will say more than any of this week's press releases did. Readers following the same launches in Polish can find the daily version of this beat at nowosci.ai.

Adjectives aren't units

"Aggressive," "midrange," and "ultra-low cost" are marketing categories. They describe how a company wants you to feel about a number, not the number itself. Three launches in 48 hours used three different words to describe pricing that ranges from $1.25 to $30 per million tokens depending on which side of the ledger and which tier you're reading, and the model called midrange turned out to be the most expensive of the three on output.

Read the two numbers, then go watch what a real prompt costs to finish. Everything in between is copywriting.