OpenAI's cheap tier becomes a routing problem: Terra $2/$12, Luna $0.20/$1.20, seats untouched

Run the percentages backwards and the old price card appears. A 20% reduction that lands on $2 and $12 implies Terra was $2.50 and $15 ^[1]. An 80% reduction that lands on $0.20 and $1.20 implies Luna was $1 and $6 ^[2]. OpenAI published one percentage per model ^[3], so that reconstruction assumes the same cut applied to input and output. If it did not, the input side moved by something other than the published figure, and that is the first thing to check against a July invoice.

The more useful number is the ratio inside each card. Terra and Luna both bill output at six times input ^[3], which means output becomes the larger half of the bill once it reaches roughly 14% of the tokens in a run ^[4]. Agent work sits well above that threshold. ChatGPT Work and Codex search through context, call tools, write files, run tests and revise their output, and every one of those steps consumes tokens ^[8]. The rate that governs a long agent run is the output rate, not the input rate the headline leads with.

Which is why the line that matters runs between tiers, not between last month and this one. Terra costs ten times Luna on both input and output ^[5]. OpenAI says Luna can use tools and complete multi-step workflows ^[6], and pointed to document analysis, classification and routine implementation as the work that becomes economical at greater volume ^[12]. Nothing in the pricing announcement specifies an architecture; the pattern of reserving the top tier for planning and handing implementation or testing to a cheaper one is runtimewire's inference from the tier descriptions ^[11]. It is also the only way to collect the tenfold gap, and it is a classifier plus a step budget that somebody on staff writes and then maintains.

The two headline rates are not the whole card either. The API documentation lists separate rates for cached input, long-context requests and other processing modes ^[10], and a multi-hour run touches several of them.

For subscription customers the arithmetic is not yet available at all. Pro, Plus and Business usage is unchanged ^[4], while eligible Work and Codex plans are getting lower credit consumption for Terra and Luna ^[5] with no per-plan conversion rate published in the announcement ^[9]. API buyers can re-model this week. Credit-plan buyers have been handed an improvement they cannot size until they inspect their own workspace configuration.

Either way the discount only shows up for teams that already count tokens per step. A workflow whose output share per run is unknown cannot be tested against these numbers, and its owner has no way to say whether 20%, 80% or nothing arrived ^[4].