diff --git a/app/services/levelcode/model_catalog.rb b/app/services/levelcode/model_catalog.rb
index 84f92b62..54f62e71 100644
--- a/app/services/levelcode/model_catalog.rb
+++ b/app/services/levelcode/model_catalog.rb
@@ -98,6 +98,20 @@ module ModelCatalog
         # gated to Max/Ultra so it never feels broken (rec #3).
         input: 10.00, cached_input: 1.00, output: 50.00,
         context: 1_000_000, min_tier: :max, status: :confirmed
+      },
+      # Sits immediately after Fable 5 because the rate card is IDENTICAL ($10/$1/$50) — an equal
+      # neighbour, exactly as Opus 5 follows Opus 4.8, so the monotonic-multiplier invariant holds.
+      "openai/gpt-6-astra" => {
+        label: "GPT-6 Astra", provider: "openrouter",
+        # Confirmed vs the OpenRouter models API (2026-09-06, listed 2026-09-04): $10/M in · $50/M out
+        # · $1/M cached read · 1.05M ctx (128K out). Image + file input. Its ENDPOINTS span $5/$25 (Flex)
+        # to $20/$100 (Fast) — the routing ceiling in OpenRouterAdapter is what keeps our cost at this row.
+        #
+        # Tier: pro_plus, not max. Same per-turn cost as Fable 5 (~13.3x), which buys ~39 turns on Pro+ —
+        # the number Pro already advertises for Opus 5 — and only ~19 on Pro, which is where rec #3 says a
+        # model starts to feel broken. Fable 5 stays at :max as previously decided; revisit together.
+        input: 10.00, cached_input: 1.00, output: 50.00,
+        context: 1_050_000, min_tier: :pro_plus, status: :confirmed
       }
     }.freeze
 
diff --git a/app/services/levelcode/open_router_adapter.rb b/app/services/levelcode/open_router_adapter.rb
index bccc77cc..486802bf 100644
--- a/app/services/levelcode/open_router_adapter.rb
+++ b/app/services/levelcode/open_router_adapter.rb
@@ -17,9 +17,9 @@ class OpenRouterAdapter
 
     # OpenRouter provider routing for the FREE model (gpt-oss). OpenRouter's "WandB" provider
     # mishandles the gpt-oss "harmony" format and 401s with "Unknown role: final", so pin the free
-    # model to reliable providers and exclude WandB. Scoped to the free model — paid models keep full
-    # routing. See atompp-internal/orbits-build/DECISIONS.md (2026-07-08). If WandB reappears, verify
-    # the `ignore` slug against OpenRouter's provider list / Activity log.
+    # model to reliable providers and exclude WandB. Scoped to the free model; paid models get a PRICE
+    # ceiling instead (with_provider_routing). If WandB reappears, verify the `ignore` slug against
+    # OpenRouter's provider list / Activity log.
     FREE_MODEL_PROVIDER = {
       "order" => %w[fireworks together deepinfra],
       "ignore" => %w[wandb],
@@ -46,7 +46,7 @@ def stream(body, on_chunk:)
       # final usage chunk otherwise, and the request would consume tokens unmetered
       # (cap bypass + under-billing). Preserve any client-supplied stream_options.
       stream_opts = (body["stream_options"] || body[:stream_options] || {}).merge("include_usage" => true)
-      request = build_request(uri, with_free_model_routing(body).merge("stream" => true, "stream_options" => stream_opts))
+      request = build_request(uri, with_provider_routing(body).merge("stream" => true, "stream_options" => stream_opts))
 
       usage = nil
       model = body["model"] || body[:model]
@@ -89,7 +89,7 @@ def stream(body, on_chunk:)
     # Non-streaming pass-through. Returns the parsed upstream JSON Hash.
     def complete(body)
       uri = URI(BASE_URL)
-      request = build_request(uri, with_free_model_routing(body).merge("stream" => false))
+      request = build_request(uri, with_provider_routing(body).merge("stream" => false))
       response = http(uri).request(request)
 
       unless response.code.to_i == 200
@@ -111,15 +111,54 @@ def initialize(status, body)
 
     private
 
-    # Inject the free-model provider routing (FREE_MODEL_PROVIDER) when the request targets the
-    # gpt-oss free model, unless the client already supplied its own `provider` preference. Every
-    # other model is returned unchanged so paid routing is untouched.
-    def with_free_model_routing(body)
+    # Shape the upstream `provider` preference for the routed model.
+    #
+    #   free model (gpt-oss) → FREE_MODEL_PROVIDER, unless the client sent its own preference.
+    #   any other catalog row → a PRICE CEILING at that row's catalog rate, merged ON TOP of whatever the
+    #                           client sent (a client may steer routing; it may not lift the ceiling).
+    #   off-catalog / Moonshot-native → untouched.
+    #
+    # WHY THE CEILING. Metering bills a request at the catalog rate (Levelcode.cost_micros) and never
+    # reads what OpenRouter actually charged. OpenRouter, left to route freely, spreads one model over
+    # endpoints priced up to 2x that rate — GPT-6 Astra runs $5/$25 (Flex) through $20/$100 (Fast), and
+    # Opus 4.8 has an Anthropic "fast" tier at $10/$50 against our $5/$25. Routed there, we under-bill by
+    # half and cannot see it. `max_price` makes wire cost <= billed cost by construction.
+    #
+    # It is a HARD filter: OpenRouter refuses the request outright if no endpoint fits, rather than
+    # quietly running it over price. That is the safe money direction — the same direction rate_for
+    # takes for an off-catalog id — and it turns a stale catalog price into a loud 4xx instead of a
+    # silent loss. Units line up without conversion: OpenRouter reads max_price in $/M tokens, and a
+    # catalog rate is micro-$/token, which is the same number.
+    #
+    # The FREE model is deliberately NOT ceilinged: its catalog row ($0.03/$0.15) sits below its listed
+    # endpoints ($0.04/$0.17), so a ceiling would exclude every endpoint and take the free tier down.
+    def with_provider_routing(body)
       model = (body["model"] || body[:model]).to_s
-      return body unless model.include?("gpt-oss")
-      return body if body.key?("provider") || body.key?(:provider)
+      if model.include?("gpt-oss")
+        return body if body.key?("provider") || body.key?(:provider)
+
+        return body.merge("provider" => FREE_MODEL_PROVIDER)
+      end
+
+      ceiling = price_ceiling(model)
+      return body unless ceiling
+
+      # Exactly one string-keyed "provider" goes out, whatever the caller sent. A symbol-keyed body
+      # would otherwise serialise :provider beside "provider" (and :max_price beside "max_price"), and
+      # which of the two OpenRouter honours would be down to its parser — the ceiling would hold or
+      # not by accident. A preference that is not an object is dropped rather than raised on: the
+      # ceiling is the part that has to survive, and a 500 here would be ours, not upstream's.
+      client = body["provider"] || body[:provider]
+      client = client.is_a?(Hash) ? client.transform_keys(&:to_s).except("max_price") : {}
+      body.except(:provider).merge("provider" => client.merge("max_price" => ceiling))
+    end
+
+    # `{"prompt" => $/M, "completion" => $/M}` for a catalog model, nil for anything else.
+    def price_ceiling(model)
+      row = Levelcode::ModelCatalog.find(model)
+      return nil unless row
 
-      body.merge("provider" => FREE_MODEL_PROVIDER)
+      { "prompt" => row[:input].to_f, "completion" => row[:output].to_f }
     end
 
     def http(uri)
diff --git a/spec/models/levelcode_model_catalog_spec.rb b/spec/models/levelcode_model_catalog_spec.rb
index 81fb1c0c..99ef9c27 100644
--- a/spec/models/levelcode_model_catalog_spec.rb
+++ b/spec/models/levelcode_model_catalog_spec.rb
@@ -23,6 +23,7 @@
       expect(described_class.find("anthropic/claude-opus-5")[:context]).to eq(1_000_000)
       expect(described_class.find("moonshotai/kimi-k3")[:context]).to eq(1_048_576)
       expect(described_class.find("openai/gpt-oss-120b")[:context]).to eq(131_072)
+      expect(described_class.find("openai/gpt-6-astra")[:context]).to eq(1_050_000)
       # No row may go without one: a nil/0 window disables the clamp entirely.
       expect(described_class.all.values.map { |m| m[:context] }).to all(be_positive)
     end
@@ -49,7 +50,10 @@
         # Opus 5 TIES Opus 4.8 — identical rates ⇒ identical multiplier. That tie is exactly what lets
         # it sit beside 4.8 without disturbing the monotonic ordering asserted below.
         "anthropic/claude-opus-4-8" => 6.67, "anthropic/claude-opus-5" => 6.67,
-        "openai/gpt-5.5" => 7.40, "anthropic/claude-fable-5" => 13.35
+        "openai/gpt-5.5" => 7.40,
+        # GPT-6 Astra TIES Fable 5 — identical rate card ⇒ identical multiplier — so it sits after Fable
+        # without disturbing the monotonic ordering, the same way Opus 5 sits beside Opus 4.8.
+        "anthropic/claude-fable-5" => 13.35, "openai/gpt-6-astra" => 13.35
       }.each { |id, mult| expect(described_class.multiplier(id)).to be_within(0.02).of(mult) }
     end
 
@@ -65,14 +69,20 @@
       expect(described_class.entitled(:free)).to eq([ "openai/gpt-oss-120b" ])
     end
 
-    it "pro reaches the roster EXCEPT Fable (gated to Max/Ultra by UX, rec #3)" do
+    it "pro reaches the roster EXCEPT the ~13x models (Fable, Astra — ~19 Pro turns feels broken, rec #3)" do
       pro = described_class.entitled(:pro)
       expect(pro).to include("moonshotai/kimi-k2.7-code", "anthropic/claude-opus-4-8", "anthropic/claude-opus-5", "openai/gpt-5.5")
-      expect(pro).not_to include("anthropic/claude-fable-5")
+      expect(pro).not_to include("anthropic/claude-fable-5", "openai/gpt-6-astra")
     end
 
-    it "Max/Ultra reach the full roster incl. Fable" do
-      expect(described_class.entitled(:max)).to include("anthropic/claude-fable-5")
+    it "pro_plus adds GPT-6 Astra (~39 turns — the count Pro already advertises for Opus 5) but not Fable" do
+      pro_plus = described_class.entitled(:pro_plus)
+      expect(pro_plus).to include("openai/gpt-6-astra")
+      expect(pro_plus).not_to include("anthropic/claude-fable-5")
+    end
+
+    it "Max/Ultra reach the full roster incl. Fable and Astra" do
+      expect(described_class.entitled(:max)).to include("anthropic/claude-fable-5", "openai/gpt-6-astra")
       expect(described_class.entitled(:ultra)).to match_array(described_class.ids)
     end
   end
diff --git a/spec/models/levelcode_plans_spec.rb b/spec/models/levelcode_plans_spec.rb
index 5722138c..7d077452 100644
--- a/spec/models/levelcode_plans_spec.rb
+++ b/spec/models/levelcode_plans_spec.rb
@@ -117,7 +117,16 @@
       # Every Pro-tier engine is confirmed now — the frontier rows (Codex 5.3, GPT-5.5, Sonnet 5) went live.
       expect(Levelcode.allowed_models('orbits_pro')).to match_array([ Levelcode::DEFAULT_MODEL, Levelcode::FREE_MODEL, 'anthropic/claude-opus-4-8', 'anthropic/claude-opus-5', 'moonshotai/kimi-k3', 'openai/gpt-5.3-codex', 'openai/gpt-5.5', 'anthropic/claude-sonnet-5' ])
       # Fable 5 is confirmed too, but Max-tier — a Pro plan still can't reach it (entitlement, not price).
-      expect(Levelcode.allowed_models('orbits_pro')).not_to include('anthropic/claude-fable-5')
+      expect(Levelcode.allowed_models('orbits_pro')).not_to include('anthropic/claude-fable-5', 'openai/gpt-6-astra')
+    end
+
+    it 'Pro+ reaches GPT-6 Astra; Fable stays Max-tier; Max reaches both' do
+      expect(Levelcode.allowed_models('orbits_pro_plus')).to include('openai/gpt-6-astra')
+      expect(Levelcode.allowed_models('orbits_pro_plus')).not_to include('anthropic/claude-fable-5')
+      expect(Levelcode.allowed_models('orbits_max')).to include('openai/gpt-6-astra', 'anthropic/claude-fable-5')
+      # Entitlement, not price: a Pro user asking for Astra is routed to the plan default, never 402'd.
+      expect(Levelcode.gateway_model('orbits_pro', 'openai/gpt-6-astra')).to eq(Levelcode::DEFAULT_MODEL)
+      expect(Levelcode.gateway_model('orbits_pro_plus', 'openai/gpt-6-astra')).to eq('openai/gpt-6-astra')
     end
 
     it 'free CANNOT reach the flagship no matter what is requested' do
diff --git a/spec/services/levelcode/open_router_adapter_spec.rb b/spec/services/levelcode/open_router_adapter_spec.rb
index d9d7a6dd..ae5064b0 100644
--- a/spec/services/levelcode/open_router_adapter_spec.rb
+++ b/spec/services/levelcode/open_router_adapter_spec.rb
@@ -29,4 +29,103 @@ def status_of(payload) = adapter.send(:error_frame_status, payload)
       expect(status_of('{"usage":{"total_tokens":9}}')).to be_nil
     end
   end
+
+  # Metering bills at the catalog rate and never reads OpenRouter's own cost, so the only thing
+  # keeping wire cost <= billed cost is the price ceiling this helper injects. These pin its shape.
+  describe "#with_provider_routing (a private request-shaping helper)" do
+    def routed(body) = adapter.send(:with_provider_routing, body)
+
+    it "pins the free model to FREE_MODEL_PROVIDER and never ceilings it" do
+      out = routed("model" => Levelcode::FREE_MODEL, "messages" => [])
+      expect(out["provider"]).to eq(described_class::FREE_MODEL_PROVIDER)
+      expect(out["provider"]).not_to have_key("max_price")
+    end
+
+    it "leaves a client-supplied provider preference alone on the free model" do
+      pref = { "order" => %w[fireworks] }
+      expect(routed("model" => Levelcode::FREE_MODEL, "provider" => pref)["provider"]).to eq(pref)
+    end
+
+    it "ceilings a paid catalog model at exactly its catalog rate, in $/M" do
+      row = Levelcode::ModelCatalog.find("openai/gpt-6-astra")
+      out = routed("model" => "openai/gpt-6-astra", "messages" => [])
+      expect(out["provider"]).to eq("max_price" => { "prompt" => 10.0, "completion" => 50.0 })
+      expect(out["provider"]["max_price"]).to eq("prompt" => row[:input].to_f, "completion" => row[:output].to_f)
+    end
+
+    it "ceilings every confirmed paid row at its own rate (no model is left routable over price)" do
+      Levelcode::ModelCatalog.ids.reject { |id| id == Levelcode::FREE_MODEL }.each do |id|
+        row = Levelcode::ModelCatalog.find(id)
+        mp = routed("model" => id)["provider"]["max_price"]
+        expect(mp).to eq("prompt" => row[:input].to_f, "completion" => row[:output].to_f), id
+      end
+    end
+
+    it "merges the ceiling ON TOP of a client preference — steering is kept, the ceiling cannot be lifted" do
+      out = routed("model" => "anthropic/claude-opus-5",
+                   "provider" => { "sort" => "throughput", "max_price" => { "prompt" => 99.0, "completion" => 999.0 } })
+      expect(out["provider"]["sort"]).to eq("throughput")
+      expect(out["provider"]["max_price"]).to eq("prompt" => 5.0, "completion" => 25.0)
+    end
+
+    it "drops a provider preference that is not an object and still applies the ceiling (never a 500)" do
+      [ "openai", %w[openai azure], true, false, nil, 7 ].each do |bad|
+        out = routed("model" => "openai/gpt-6-astra", "provider" => bad)
+        expect(out["provider"]).to eq("max_price" => { "prompt" => 10.0, "completion" => 50.0 }), bad.inspect
+      end
+    end
+
+    it "normalises a symbol-keyed body to ONE string-keyed provider, so the ceiling survives serialisation" do
+      out = routed(model: "openai/gpt-6-astra", provider: { sort: "price", max_price: { prompt: 99.0, completion: 999.0 } })
+      expect(out).not_to have_key(:provider)
+      expect(out["provider"]).to eq("sort" => "price", "max_price" => { "prompt" => 10.0, "completion" => 50.0 })
+      json = JSON.generate(out)
+      expect(json.scan('"provider"').size).to eq(1)
+      expect(json.scan('"max_price"').size).to eq(1)
+    end
+
+    it "leaves an off-catalog model untouched (nothing to ceiling against)" do
+      body = { "model" => "vendor/not-in-catalog", "messages" => [] }
+      expect(routed(body)).to eq(body)
+    end
+
+    it "does not mutate the caller's body" do
+      body = { "model" => "openai/gpt-6-astra" }.freeze
+      expect { routed(body) }.not_to raise_error
+      expect(body).not_to have_key("provider")
+    end
+  end
+
+  # The helper above is only worth anything if #stream and #complete actually route through it. No
+  # spec drives those (the request specs double the router), so pin them at the adapter's own wire
+  # seam: build_request is what serialises the upstream body, so the body it receives IS the wire.
+  # A sentinel stops the call before any socket is opened.
+  describe "#stream / #complete put the provider routing on the wire" do
+    sentinel = Class.new(StandardError)
+    let(:sent) { {} }
+
+    before do
+      allow(adapter).to receive(:build_request) do |_uri, body|
+        sent.replace(body)
+        raise sentinel
+      end
+    end
+
+    it "#stream sends the price ceiling for a paid model" do
+      expect { adapter.stream({ "model" => "openai/gpt-6-astra", "messages" => [] }, on_chunk: ->(_) { }) }.to raise_error(sentinel)
+      expect(sent["provider"]).to eq("max_price" => { "prompt" => 10.0, "completion" => 50.0 })
+      expect(sent["stream"]).to be(true)
+    end
+
+    it "#complete sends the price ceiling for a paid model" do
+      expect { adapter.complete("model" => "anthropic/claude-opus-5", "messages" => []) }.to raise_error(sentinel)
+      expect(sent["provider"]).to eq("max_price" => { "prompt" => 5.0, "completion" => 25.0 })
+      expect(sent["stream"]).to be(false)
+    end
+
+    it "#stream still pins the free model to FREE_MODEL_PROVIDER" do
+      expect { adapter.stream({ "model" => Levelcode::FREE_MODEL, "messages" => [] }, on_chunk: ->(_) { }) }.to raise_error(sentinel)
+      expect(sent["provider"]).to eq(described_class::FREE_MODEL_PROVIDER)
+    end
+  end
 end