VynarisEarly beta Kimi K3Get your API key

Blog · 2026-07-31 · Vynaris Team

OpenAI cut GPT-5.6 Luna 80% and Terra 20% (Sol held): the re-route math

OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 and Terra 20% to $2/$12 on 2026-07-30; Sol held at $5/$30. A real sticker cut, not a tier re-route: the re-route math on a 1,000-task/day workhorse and who should move. Prices verified 2026-07-31.

OpenAI cut GPT-5.6 Luna from $1/$6 to $0.20/$1.20 per million input/output tokens on 2026-07-30, an 80% drop on both sides. Terra fell 20% to $2/$12. The flagship Sol held at $5/$30. This is a real sticker cut, not a tier down-route you have to opt into. Prices verified 2026-07-31 off OpenAI's live rate card. Luna is now the cheapest frontier-lab workhorse, and it undercuts the budget tiers most teams run on.

That last line is the story. Luna's new $0.20 input matches OpenAI's own gpt-5.4-nano sticker ($0.20/$1.25) while being far more capable, and against Claude Haiku 4.5 ($1/$5) it runs 5x cheaper on input and 4.2x on output. If you set your cheap tier weeks ago and never re-checked, the floor moved under you today.

What actually changed (verified live 2026-07-31)

Model          Old $/1M        New $/1M        Change
-------------  --------------  --------------  ---------------
GPT-5.6 Luna   $1.00 / $6.00   $0.20 / $1.20   −80% in and out
GPT-5.6 Terra  $2.50 / $15.00  $2.00 / $12.00  −20% in and out
GPT-5.6 Sol    $5.00 / $30.00  $5.00 / $30.00  held

Luna's cache read dropped from $0.10 to $0.02 per million alongside the base cut. Two structural notes from the same live read: Sol picked up a fast mode at $10/$60, a flat 2x on both sides for roughly 2.5x speed, and the long-context (>272k) tiers still double to Luna $0.40/$1.80, Terra $4/$18, Sol $10/$45.

The re-route table: one workhorse shape, five budget models

Put a real shape on it: a 1,000-task/day classification agent at 2,000 input tokens and 300 output tokens per task. Monthly cost per token times volume, prices verified 2026-07-31:

Model                    $/1M in  $/1M out  $/1,000 tasks  $/month
-----------------------  -------  --------  -------------  -------
DeepSeek v4-flash        $0.14    $0.28     $0.36          $10.92
GPT-5.6 Luna (post-cut)  $0.20    $1.20     $0.76          $22.80
Gemini 3.5 Flash-Lite    $0.30    $2.50     $1.35          $40.50
gpt-5.4-mini             $0.75    $4.50     $2.85          $85.50
Claude Haiku 4.5         $1.00    $5.00     $3.50          $105.00

On this shape Luna's own bill fell from $114/month before the cut to $22.80 after, the 80% carried straight through. Claude Haiku 4.5 now costs 4.6x Luna for the same job, Gemini 3.5 Flash-Lite 1.8x, and gpt-5.4-mini 3.75x.

Monthly cost of a 1,000-task/day agent (2,000 in / 300 out) across six budget options; Luna drops from $114 pre-cut to $23 post-cut, past Haiku, gpt-5.4-mini and Flash-Lite, with DeepSeek at $11
Cost per month on a 1,000-task/day workhorse (2,000 in / 300 out). Faded bar = Luna pre-cut. Log scale. Prices verified 2026-07-31.

The honest catch: sticker is not cost-per-task, and DeepSeek still undercuts

Two caveats, because there always are. First, DeepSeek v4-flash ($0.14/$0.28) is still cheaper than Luna on this exact shape, $10.92 against $22.80 a month, and it wins pure output-heavy work by more, because its output price sits 4.3x below Luna's. Luna is the new floor among frontier labs, not the absolute floor. If raw sticker is your only axis and an open-weight model clears your quality bar, DeepSeek stays cheapest.

Second, the number that decides a route is cost-per-task, not cost-per-token. A model that needs one retry, or emits 30% more tokens for the same answer, erases a sticker gap. Verify token-efficiency on your own traffic before you migrate a routing table on a launch-day headline. Run your own in/out split through the calculator first.

Who should re-route this week

Your workhorse tier today              The move                                    What it saves on the shape above
-------------------------------------  ------------------------------------------  ----------------------------------------------------------------------------
Haiku 4.5 for cheap classification     A/B Luna at matched quality                 4.6x, about $82/mo per 1k-task/day agent
Gemini 3.5 Flash-Lite                  Price Luna against it                       1.8x, about $18/mo
gpt-5.4-mini                           Repoint to Luna (same provider)             3.75x, about $63/mo
Already on DeepSeek v4-flash           Stay, unless you want frontier-lab quality  Luna costs 2.1x more here
On Sol/Terra for down-routable volume  Cascade cheap-first to Luna                 A [routing](https://vynaris.com/glossary/model-routing) win, not a price cut

Sol holding at $5/$30 while Luna fell 80% says OpenAI is defending frontier margin and buying the high-volume workhorse market with the cheap tier. That is the pattern Google ran when Gemini 3.6 Flash cut output while hiking Flash-Lite: the cheap tier is where the price war is fought. The cut landed, and it landed on Luna, not Sol.

What to do

For the durable four-model verdict across extraction, classification, and short-generation shapes, see the budget-tier comparison companion. For the wider rate sheet, the July 2026 LLM price list. For how fast mode and other multipliers stack on any sticker, the price-multipliers breakdown.

FAQ

Is the GPT-5.6 Luna cut a real price drop or a tier re-route? A real sticker cut. Luna's list price fell from $1/$6 to $0.20/$1.20 per million tokens on the live rate card, verified 2026-07-31. The same call costs 80% less. DeepSeek v4-flash ($0.14/$0.28) is still cheaper on raw sticker, so Luna is the new floor among frontier labs, not the absolute floor.

Sources

Related: budget-tier comparison · July 2026 LLM price list · Gemini 3.6 Flash re-route math