Add verified GPT-5.6 Bedrock cache-write pricing (1.25x input rate)

OpenAI's own docs confirm the gap flagged in the DUX-11214 follow-up
was real, not just unverified: "Cache writes have no additional fee
on models before the GPT-5.6 family. For GPT-5.6 models and later
model families, cache writes cost 1.25x the uncached input token
rate." (https://developers.openai.com/api/docs/guides/prompt-caching)

GPT-5.6 is the first OpenAI family billed for cache writes at all,
which is why every other OpenAI entry in models.json has no
cache_write_input_per_million field - that absence really is $0,
not an omission. Add the field to the three hand-added
openai.gpt-5.6-luna/-terra/-sol entries (7791056), computed as
1.25x each model's own input_per_million:
  Luna:  1.0   x 1.25 = 1.25
  Terra: 2.5   x 1.25 = 3.125
  Sol:   5.0   x 1.25 = 6.25

Update models_gpt_5_6_bedrock_spec.rb to assert the new rates and
replace the "intentionally omitted, unverified" comment with the
citation above.

Also document a second, separate, still-open concern (not resolved
here): OpenAI's developer community reported two GPT-5.6-specific
usage-accounting bugs in July 2026 - a cached/cache-write
double-counting bug OpenAI staff confirmed and refunded, and a
seemingly still-open report of ~9x inflated output_tokens billing.
We consume these models via Bedrock's mantle passthrough, not
OpenAI's own API, so whether Bedrock reproduces either bug is
unknown and unverified - flagged as an open risk to revisit.

Co-authored-by: Sam Boland <sam.boland@appfolio.com>

Author: appf-supernova

lib/ruby_llm/models.json

@@ -16814,7 +16814,8 @@
         "standard": {
           "input_per_million": 1.0,
           "output_per_million": 6.0,
-          "cache_read_input_per_million": 0.1
+          "cache_read_input_per_million": 0.1,
+          "cache_write_input_per_million": 1.25
         }
       }
     },
@@ -16841,7 +16842,8 @@
       "cost": {
         "input": 1.0,
         "output": 6.0,
-        "cache_read": 0.1
+        "cache_read": 0.1,
+        "cache_write": 1.25
       },
       "limit": {
         "context": 272000,
@@ -16880,7 +16882,8 @@
         "standard": {
           "input_per_million": 5.0,
           "output_per_million": 30.0,
-          "cache_read_input_per_million": 0.5
+          "cache_read_input_per_million": 0.5,
+          "cache_write_input_per_million": 6.25
         }
       }
     },
@@ -16907,7 +16910,8 @@
       "cost": {
         "input": 5.0,
         "output": 30.0,
-        "cache_read": 0.5
+        "cache_read": 0.5,
+        "cache_write": 6.25
       },
       "limit": {
         "context": 272000,
@@ -16946,7 +16950,8 @@
         "standard": {
           "input_per_million": 2.5,
           "output_per_million": 15.0,
-          "cache_read_input_per_million": 0.25
+          "cache_read_input_per_million": 0.25,
+          "cache_write_input_per_million": 3.125
         }
       }
     },
@@ -16973,7 +16978,8 @@
       "cost": {
         "input": 2.5,
         "output": 15.0,
-        "cache_read": 0.25
+        "cache_read": 0.25,
+        "cache_write": 3.125
       },
       "limit": {
         "context": 272000,

spec/ruby_llm/models_gpt_5_6_bedrock_spec.rb

@@ -5,19 +5,30 @@
 RSpec.describe RubyLLM::Models do
   include_context 'with configured RubyLLM'
 
-  # None of these three hand-added entries (see 7791056d) carry a
-  # cache_write_input_per_million rate, matching every other OpenAI-family
-  # model in models.json — no OpenAI model in this registry has that field
-  # modeled. This is consistent with, but NOT independently verified
-  # against, OpenAI/Bedrock pricing docs for GPT-5.6 specifically: whether
-  # Bedrock charges a distinct (non-zero) cache-write rate for these models,
-  # the way it does for Anthropic's cache_creation_input_per_million, is
-  # unknown. Revisit if models.dev adds real entries for these ids, or if
-  # OpenAI/AWS publish a documented cache-write rate for GPT-5.6.
+  # cache_write is verified, not a placeholder: per OpenAI's own docs
+  # (https://developers.openai.com/api/docs/guides/prompt-caching), "Cache
+  # writes have no additional fee on models before the GPT-5.6 family. For
+  # GPT-5.6 models and later model families, cache writes cost 1.25x the
+  # uncached input token rate." GPT-5.6 is the first OpenAI family to charge
+  # for cache writes at all, which is why it's the only OpenAI family in this
+  # registry carrying a cache_write_input_per_million value — every other
+  # OpenAI entry has none because the rate really is $0 for those models.
+  #
+  # Separately open (do NOT resolve here): OpenAI's developer community has
+  # reported two GPT-5.6-specific usage-accounting bugs since launch (July
+  # 2026) — one where cached_tokens + cache_write_tokens could nearly
+  # double-count against prompt_tokens (OpenAI staff andyw1 confirmed this
+  # and issued retroactive refunds), and a second, seemingly still-open one
+  # alleging usage.output_tokens can be inflated ~9x by a reasoning-token
+  # resummation bug, with the inflated figure being what's billed. We consume
+  # these models via AWS Bedrock's mantle passthrough, not OpenAI's own API,
+  # so it is unknown and unverified whether Bedrock's usage accounting
+  # reproduces either bug or computes usage independently. Treat billed
+  # Luna/Terra/Sol costs as an open risk until this is checked.
   {
-    'openai.gpt-5.6-sol' => { input: 5.0, output: 30.0, cache_read: 0.5 },
-    'openai.gpt-5.6-terra' => { input: 2.5, output: 15.0, cache_read: 0.25 },
-    'openai.gpt-5.6-luna' => { input: 1.0, output: 6.0, cache_read: 0.1 }
+    'openai.gpt-5.6-sol' => { input: 5.0, output: 30.0, cache_read: 0.5, cache_write: 6.25 },
+    'openai.gpt-5.6-terra' => { input: 2.5, output: 15.0, cache_read: 0.25, cache_write: 3.125 },
+    'openai.gpt-5.6-luna' => { input: 1.0, output: 6.0, cache_read: 0.1, cache_write: 1.25 }
   }.each do |id, cost|
     it "resolves #{id} from the bedrock provider with the documented effort values" do
       model = RubyLLM.models.find(id, :bedrock)