VynarisEarly betaGet your API key

Blog · 2026-09-11 · Vynaris Team

GPT-5.6 promo expired: Sol reverts to $5/$30, Luna jumps 5x to $1/$6

GPT-5.6 promotional pricing expired. Sol reverts to $5/$30, Luna jumps 5x to $1/$6. An 8k/2k Sol task costs $0.10, up 38.9%. Prices verified 2026-09-11.

GPT-5.6 Sol is back to $5/$30 per million tokens. The promotional rates that cut Sol to $4/$20 on July 30 are gone from the live pricing page as of September 11. Terra reverted from $2/$12 to $2.5/$15. Luna jumped from $0.20/$1.20 to $1/$6, a 5x increase. Prices verified 2026-09-11.

TL;DR

Verdict table

Model          Promo rate (in/out per MTok)  Standard rate (in/out per MTok)  8k/2k per 1,000 (promo)  8k/2k per 1,000 (standard)  Change
-------------  ----------------------------  -------------------------------  -----------------------  --------------------------  -------
gpt-5.6-sol    $4 / $20                      $5 / $30                         $72.00                   $100.00                     +38.9%
gpt-5.6-terra  $2 / $12                      $2.5 / $15                       $40.00                   $50.00                      +25.0%
gpt-5.6-luna   $0.20 / $1.20                 $1 / $6                          $4.00                    $20.00                      +400.0%

These workloads are editable assumptions, not observed traffic. The Sol row is (8,000 × $5 + 2,000 × $30) / 1,000,000 × 1,000 = $100.

Source: OpenAI API pricing; derived fixed workloads. Prices verified 2026-09-11.

GPT-5.6 cost per 1,000 tasks: promotional vs standard rates
Source: OpenAI API pricing, verified 2026-09-11. 8k input / 2k output per task. Log scale.

The promo is gone from the live page

The prior article covered the promo launch on August 22, when Sol dropped from $5/$30 to $4/$20. That article noted OpenAI's statement that the promotional rate was available "at least through 2026-11-21."

The live pricing page now shows Standard rates for all three GPT-5.6 models. The @AGTPinsights post on X on July 30 described the discount as running "for the next three months." September 11 is 43 days into that window.

This article states what the live page shows. OpenAI has not published a primary source confirming an early end to the promotion. The page reflects Standard rates as of today.

Luna's 5x jump

Luna was the cheapest model in the GPT-5.6 lineup at $0.20/$1.20 per MTok. At standard rates, it costs $1/$6. Both components multiplied by 5.

A 10k-input/2k-output Luna call that cost $0.0044 now costs $0.022. Per 1,000 tasks, that is $4.40 to $22.00. The per-request cost shift is large in percentage terms but small in absolute dollars. A workload running 100,000 Luna tasks per month at 8k/2k sees its bill go from $400 to $2,000.

Luna's promotional rate was aggressive enough to compete with open-source alternatives on a per-token basis. At $1/$6, that comparison narrows. Workloads that chose Luna for price alone should re-evaluate. The 10k-input/2k-output workload tells the same story: $0.0044 per task during the promo, $0.022 at standard rates. Per 1,000 tasks, that is $4.40 to $22.00.

Batch, Flex and Priority all reverted

The service-tier ladder preserved its 0.5x/1x/2x structure. Every tier reverted to standard rates.

Tier      Sol input/output per MTok  8k/2k per 1,000  Change from promo
--------  -------------------------  ---------------  -----------------
Batch     $2.5 / $15                 $50.00           +38.9%
Flex      $2.5 / $15                 $50.00           +38.9%
Standard  $5 / $30                   $100.00          +38.9%
Priority  $10 / $60                  $200.00          +38.9%

"Fast" is now "Priority." The rate is $10/$60, exactly 2x Standard. Batch and Flex remain identical at $2.5/$15, both 0.5x Standard. Batch inference and Flex share a price but differ in request contract. Pick the service behavior first, then price the eligible tier.

Long-context rates reverted

OpenAI applies long-context pricing to the full request above 272,000 input tokens. The multiplier is 2x for input and 1.5x for output. At standard rates, that makes Sol $10/$45 for long-context requests. During the promo it was $8/$30.

A 300k-input/2k-output request now costs $3.00 + $0.09 = $3.09, or $3,090 per 1,000 tasks. During the promo it was $2.40 + $0.06 = $2.46, or $2,460 per 1,000. The increase is 25.6%, smaller than the balanced workload because input dominates at this scale. The threshold itself did not change. The long-context cost analysis covers the billing mechanics in detail.

Cache rates reverted too

Prompt caching reads for Sol fell from $0.50 to $0.40 per MTok during the promo. They are back to $0.50. Cache writes reverted from $5 to $6.25.

For 8,000 cached tokens plus 2,000 output tokens, the bill is now $4 + $60 = $64 per 1,000 tasks. During the promo it was $3.20 + $40 = $43.20. The prompt caching production guide covers the operational side. A lower cache rate cannot rescue a workload that misses its reusable prefix.

The routing crossover flipped

On the balanced 8k/2k workload, Claude Opus 5 costs $40 input + $50 output = $90 per 1,000 tasks at Anthropic's $5/$25 rates. Claude Sonnet 5 costs $24 + $30 = $54 at $3/$15 (Sonnet 5 had its own increase from $2/$10 on September 1).

Promotional Sol at $72 was 20% cheaper than Opus 5. Standard Sol at $100 is 11.1% more expensive. Any router that cached the promotional break-even is now routing to a more expensive model.

Use the cost calculator with your actual token counts. The crossover depends on your input-to-output ratio, not the headline rate. An input-heavy workload (20k input, 1k output) costs $130 per 1,000 at Sol standard rates. The same workload costs $125 at Opus 5 and $75 at Sonnet 5. The gap widens with output share: a 2k-input/8k-output workload costs $250 at Sol, $210 at Opus 5, and $126 at Sonnet 5.

Monthly budget impact

A workload running 100,000 Sol tasks per month at 8k/2k input/output sees its bill go from $7,200 to $10,000. That is $2,800 per month, a 38.9% increase, with no change in token volume or model selection.

Teams that built cost projections on the promotional rate face an immediate gap. The prior article's budgeting advice to keep the old $5/$30 row as a conservative forecast scenario was the right call. That scenario is now the live rate.

Honest tradeoff

Do not switch from Sol to Sonnet 5 because $54 is lower than $100. Model quality, retry rates, latency, and failed-task cost can reverse a $46 spread. A 20% retry rate turns $100 into $100 × 1.2 = $120 before failure penalties. A model that fails 30% of tasks at $100 per 1,000 successful attempts costs $100 / 0.7 = $142.86 per 1,000 attempts.

The other caveat is timing. The live page shows Standard rates today. OpenAI could introduce a new promotional rate tomorrow. Budget for the Standard rate and treat any future discount as upside, not a baseline.

Sources