feat: support Claude Opus 5

cmd/root/eval.go

@@ -18,7 +18,7 @@ import (
 	"github.com/docker/docker-agent/pkg/telemetry"
 )
 
-const defaultJudgeModel = "anthropic/claude-opus-4-5-20251101"
+const defaultJudgeModel = "anthropic/claude-opus-5"
 
 type evalFlags struct {
 	evaluation.Config

docs/concepts/models/index.md

@@ -73,7 +73,7 @@ for details.
 | Provider            | Key              | Example Models                       | API Key Env Var                     |
 | ------------------- | ---------------- | ------------------------------------ | ----------------------------------- |
 | OpenAI              | `openai`         | gpt-5, gpt-5-mini, gpt-4o            | `OPENAI_API_KEY`                    |
-| Anthropic           | `anthropic`      | claude-sonnet-4-5, claude-opus-4-7   | `ANTHROPIC_API_KEY`                 |
+| Anthropic           | `anthropic`      | claude-sonnet-4-5, claude-opus-5     | `ANTHROPIC_API_KEY`                 |
 | Google              | `google`         | gemini-3.5-flash, gemini-3-pro       | `GOOGLE_API_KEY` / `GEMINI_API_KEY` |
 | AWS Bedrock         | `amazon-bedrock` | Claude, Nova, Llama models           | AWS credentials                     |
 | Docker Model Runner | `dmr`            | ai/qwen3, ai/llama3.2                | None (local)                        |

docs/features/cli/index.md

@@ -455,7 +455,7 @@ $ docker agent eval <agent-file>|<registry-ref> [<eval-dir>|./evals] [flags]
 | Flag                | Default                              | Description                                                                |
 | ------------------- | ------------------------------------ | -------------------------------------------------------------------------- |
 | `-c, --concurrency` | num CPUs                             | Number of concurrent evaluation runs                                       |
-| `--judge-model`     | `anthropic/claude-opus-4-5-20251101` | Model for LLM-as-a-judge relevance scoring (format: `provider/model`)      |
+| `--judge-model`     | `anthropic/claude-opus-5` | Model for LLM-as-a-judge relevance scoring (format: `provider/model`)      |
 | `--output <dir>`    | `<eval-dir>/results`                 | Directory for results, logs, and session databases                         |
 | `--only <pattern>`  | (all)                                | Only run evals with file names matching these patterns (repeatable)        |
 | `--base-image`      | (default)                            | Custom base Docker image for eval containers                               |

docs/features/evaluation/index.md

@@ -157,7 +157,7 @@ $ docker agent eval <agent-file>|<registry-ref> [<eval-dir>|./evals]
 | Flag                | Default                     | Description                                                       |
 | ------------------- | --------------------------- | ----------------------------------------------------------------- |
 | `-c, --concurrency` | num CPUs                    | Number of concurrent evaluation runs                              |
-| `--judge-model`     | `anthropic/claude-opus-4-5-20251101` | Model for LLM-as-a-judge relevance scoring                        |
+| `--judge-model`     | `anthropic/claude-opus-5` | Model for LLM-as-a-judge relevance scoring                        |
 | `--output`          | `<eval-dir>/results`  | Directory for results, logs, and session databases                |
 | `--only`            | (all)                       | Only run evals with file names matching these patterns            |
 | `--base-image`      | (default)                   | Custom base Docker image for eval containers (see [Custom Base Images](#custom-base-images)) |

docs/providers/anthropic/index.md

@@ -98,7 +98,8 @@ models:
 
 | Model ID            | Description                                         |
 | ------------------- | --------------------------------------------------- |
-| `claude-opus-4-7`   | Highest-capability Opus model; supports task budget |
+| `claude-opus-5`     | Highest-capability Opus model; full effort ladder (low–max) |
+| `claude-opus-4-7`   | Previous Opus flagship; supports task budget        |
 | `claude-sonnet-4-5` | Most capable Sonnet; supports extended thinking     |
 | `claude-sonnet-4-0` | Previous Sonnet generation, still supported         |
 | `claude-haiku-4-5`  | Fast and inexpensive, good for tight loops          |

pkg/modelinfo/modelinfo.go

@@ -3,7 +3,7 @@
 //
 // Some providers must specialize their behavior depending on the underlying
 // model: pick OpenAI's Responses API for o-series and gpt-5, switch Claude
-// Opus 4.6/4.7/4.8 to adaptive thinking, use level-based thinking for Gemini 3+,
+// Opus 4.6/4.7/4.8/5 to adaptive thinking, use level-based thinking for Gemini 3+,
 // auto-enable interleaved thinking for any Claude model regardless of host
 // (Anthropic, Bedrock, Vertex AI Model Garden), decide which attachment MIME
 // types can be forwarded natively, and so on.
@@ -113,23 +113,24 @@ func AlwaysReasons(modelID string) bool {
 	return isOSeries(normalizeOpenAI(modelID))
 }
 
-// claudeOpus46To48Prefixes lists the bare Claude Opus model families that
+// adaptiveOnlyOpusPrefixes lists the bare Claude Opus model families that
 // reject token-based thinking ([RejectsTokenThinking]) and instead require
 // adaptive thinking. This is an API behavior quirk that models.dev does not
 // describe, so it stays hard-coded here (unlike context windows, which come
 // from the catalogue).
-var claudeOpus46To48Prefixes = []string{"claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8"}
-
-// isClaudeOpus46To48 reports whether modelID names a Claude Opus 4.6, 4.7 or
-// 4.8 model (or a dated variant like claude-opus-4-7-20251101). Bedrock-style
-// identifiers such as "global.anthropic.claude-opus-4-8" are recognised by
-// stripping the inference-profile prefix first.
-func isClaudeOpus46To48(modelID string) bool {
+var adaptiveOnlyOpusPrefixes = []string{"claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-5"}
+
+// isAdaptiveOnlyOpus reports whether modelID names a Claude Opus model that
+// rejects token-based thinking (e.g. claude-opus-4-6, claude-opus-5, or a
+// dated variant like claude-opus-4-7-20251101). Bedrock-style identifiers
+// such as "global.anthropic.claude-opus-5" are recognised by stripping the
+// inference-profile prefix first.
+func isAdaptiveOnlyOpus(modelID string) bool {
 	m := normalize(modelID)
 	if bare, ok := bedrockClaudeModelName(m); ok {
 		m = bare
 	}
-	for _, prefix := range claudeOpus46To48Prefixes {
+	for _, prefix := range adaptiveOnlyOpusPrefixes {
 		if m == prefix || strings.HasPrefix(m, prefix+"-") {
 			return true
 		}
@@ -141,15 +142,15 @@ func isClaudeOpus46To48(modelID string) bool {
 // `thinking.type=enabled` (token-based extended thinking) and instead requires
 // `thinking.type=adaptive`.
 //
-// Currently Claude Opus 4.6, 4.7 and 4.8 (and dated variants like
-// claude-opus-4-7-20251101). Bedrock-style identifiers such as
-// "global.anthropic.claude-opus-4-8" are recognised too.
+// Applies to Claude Opus 4.6, 4.7, 4.8, 5, and their dated variants (e.g.
+// claude-opus-4-7-20251101, claude-opus-5-20260724). Bedrock-style identifiers
+// such as "global.anthropic.claude-opus-5" are recognised too.
 // For these models the agent transparently switches a token-based budget to
 // adaptive thinking.
 //
 // See https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking
 func RejectsTokenThinking(modelID string) bool {
-	return isClaudeOpus46To48(modelID)
+	return isAdaptiveOnlyOpus(modelID)
 }
 
 // SupportsAdaptiveThinking reports whether an Anthropic Claude model accepts
@@ -161,10 +162,10 @@ func RejectsTokenThinking(modelID string) bool {
 // Claude 3.x) reject `thinking.type=adaptive`/`output_config.effort` with a 400
 // and must use token-based extended thinking (`thinking.type=enabled`) instead.
 //
-// Supported: Opus 4.6/4.7/4.8 (which additionally reject token budgets, see
-// [RejectsTokenThinking]), Sonnet 4.6, the Claude 5 families (e.g. Sonnet 5),
-// and the codenamed frontier models (Fable, Mythos). Bedrock-style identifiers
-// such as "global.anthropic.claude-sonnet-4-6" are recognised too.
+// Supported: Opus 4.6/4.7/4.8/5 (which additionally reject token budgets, see
+// [RejectsTokenThinking]), Sonnet 4.6, the Claude 5 families (e.g. Sonnet 5,
+// Opus 5), and the codenamed frontier models (Fable, Mythos). Bedrock-style
+// identifiers such as "global.anthropic.claude-sonnet-4-6" are recognised too.
 //
 // The set is a superset of [RejectsTokenThinking]: a model that rejects token
 // budgets must accept adaptive thinking.

pkg/modelinfo/modelinfo_test.go

@@ -96,6 +96,7 @@ func TestSupportsDeferredTools(t *testing.T) {
 		model    string
 		want     bool
 	}{
+		{"anthropic", "claude-opus-5", true},
 		{"anthropic", "claude-opus-4-5", true},
 		{"anthropic", "claude-opus-4-5-20251101", true},
 		{"anthropic", "claude-sonnet-4-5", true},
@@ -300,6 +301,10 @@ func TestRejectsTokenThinking(t *testing.T) {
 		{"claude-opus-4-60", false}, // must not match
 		{"claude-opus-4-70", false},
 		{"claude-opus-4-80", false},
+		// Claude Opus 5 rejects token-based thinking (adaptive-only generation).
+		{"claude-opus-5", true},
+		{"claude-opus-5-20260724", true},
+		{"global.anthropic.claude-opus-5-20260724-v1:0", true},
 		{"", false},
 		// An "openai/" qualifier must NOT be stripped here: normalize() is
 		// generic (not OpenAI-specific), so a qualified id naming a Claude
@@ -388,9 +393,11 @@ func TestSupportsAdaptiveThinkingSupersetOfRejects(t *testing.T) {
 
 	models := []string{
 		"claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8",
-		"claude-opus-4-8-20260601", "claude-sonnet-4-5", "claude-sonnet-4-6",
+		"claude-opus-4-8-20260601", "claude-opus-5", "claude-opus-5-20260724",
+		"claude-sonnet-4-5", "claude-sonnet-4-6",
 		"claude-sonnet-5", "claude-haiku-4-5", "claude-fable-5",
 		"anthropic.claude-opus-4-8-v1:0", "global.anthropic.claude-opus-4-6-v1:0",
+		"global.anthropic.claude-opus-5-20260724-v1:0",
 	}
 	for _, m := range models {
 		t.Run(m, func(t *testing.T) {
@@ -414,10 +421,11 @@ func TestSupportsFullThinkingDisplay(t *testing.T) {
 		{"claude-sonnet-4-5", true},
 		{"claude-haiku-4-5", true},
 		{"claude-opus-4-5", true},
-		// The adaptive generation (4.6+) only accepts summarized/omitted.
+		// The adaptive generation (4.6+, and Opus 5) only accepts summarized/omitted.
 		{"claude-opus-4-6", false},
 		{"claude-opus-4-7", false},
 		{"claude-opus-4-8", false},
+		{"claude-opus-5", false},
 		{"claude-sonnet-4-6", false},
 		{"claude-sonnet-5", false},
 		{"claude-fable-5", false},

pkg/modelinfo/thinking_levels.go

@@ -122,6 +122,12 @@ func anthropicTopTierSupport(m string) (hasXHigh, hasMax bool) {
 		// release adds xhigh while the preview tops out at max.
 		return !strings.Contains(m, "preview"), true
 	}
+	// Claude Opus 5+ ships the full effort ladder (xhigh and max), matching Opus 4.7+.
+	if strings.Contains(m, "opus") {
+		if major, _, ok := claudeOpusSonnetVersion(m); ok && major >= 5 {
+			return true, true
+		}
+	}
 	family, minor, ok := claudeFamilyMinor(m)
 	if !ok {
 		return false, false

pkg/modelinfo/thinking_levels_test.go

@@ -89,6 +89,18 @@ func TestSupportedThinkingLevels(t *testing.T) {
 			modelID:  "anthropic.claude-sonnet-4-5-20250929-v1:0",
 			want:     []effort.Level{effort.None, effort.Low, effort.Medium, effort.High},
 		},
+		{
+			name:     "claude opus 5 gets xhigh and max",
+			provider: "anthropic",
+			modelID:  "claude-opus-5",
+			want:     []effort.Level{effort.None, effort.Low, effort.Medium, effort.High, effort.XHigh, effort.Max},
+		},
+		{
+			name:     "bedrock claude opus 5 gets xhigh and max",
+			provider: "amazon-bedrock",
+			modelID:  "us.anthropic.claude-opus-5-20260724",
+			want:     []effort.Level{effort.None, effort.Low, effort.Medium, effort.High, effort.XHigh, effort.Max},
+		},
 		{
 			name:     "claude fable gets xhigh and max",
 			provider: "anthropic",
@@ -419,6 +431,9 @@ func TestAnthropicTopEfforts(t *testing.T) {
 		// Both xhigh and max.
 		{"claude-opus-4-7", []effort.Level{effort.XHigh, effort.Max}},
 		{"claude-opus-4-8", []effort.Level{effort.XHigh, effort.Max}},
+		{"claude-opus-5", []effort.Level{effort.XHigh, effort.Max}},
+		{"claude-opus-5-20260724", []effort.Level{effort.XHigh, effort.Max}},
+		{"us.anthropic.claude-opus-5", []effort.Level{effort.XHigh, effort.Max}},
 		{"claude-fable-5", []effort.Level{effort.XHigh, effort.Max}},
 		{"claude-mythos-5", []effort.Level{effort.XHigh, effort.Max}},
 		// Bedrock-style identifiers with regional prefixes: the prefix is

pkg/modelsdev/snapshot_date.txt

@@ -1 +1 @@
-2026-07-20T07:08:23Z
+2026-07-24T19:29:40Z

pkg/modelsdev/snapshot_test.go

@@ -37,6 +37,27 @@ func TestEmbeddedSnapshotIncludesGPT56Family(t *testing.T) {
 	}
 }
 
+// TestEmbeddedSnapshotIncludesClaudeOpus5 is the regression test for the
+// Claude Opus 5 snapshot refresh (released 2026-07-24): it guards against a
+// stale embedded snapshot silently dropping offline model resolution /
+// pricing / context limits, the same way TestEmbeddedSnapshotIncludesGPT56Family
+// spot-checks the gpt-5.6 family.
+func TestEmbeddedSnapshotIncludesClaudeOpus5(t *testing.T) {
+	t.Parallel()
+
+	db := embeddedSnapshot()
+	require.NotNil(t, db)
+	anthropic, ok := db.Providers["anthropic"]
+	require.True(t, ok, "embedded snapshot must contain the anthropic provider")
+
+	m, ok := anthropic.Models["claude-opus-5"]
+	require.True(t, ok, "embedded snapshot must contain anthropic/claude-opus-5")
+	assert.Equal(t, 1000000, m.Limit.Context, "claude-opus-5 must report the 1M context window")
+	assert.Equal(t, int64(128000), m.Limit.Output, "claude-opus-5 must report the 128k max output")
+	assert.True(t, m.Reasoning, "claude-opus-5 must be marked as a reasoning model")
+	assert.Equal(t, "claude-opus", m.Family, "claude-opus-5 must belong to the claude-opus family")
+}
+
 // TestEmbeddedSnapshotParses ensures the snapshot baked into the binary is
 // valid JSON and carries a non-trivial catalog. A broken snapshot would
 // silently degrade every offline lookup, so we guard it at build time.