@blockrun/franklin 3.35.3 → 3.35.5
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/agent/optimize.js +24 -5
- package/package.json +1 -1
package/dist/agent/optimize.js
CHANGED
|
@@ -32,17 +32,36 @@ const MODEL_MAX_OUTPUT = {
|
|
|
32
32
|
'anthropic/claude-opus-4.6': 32_000,
|
|
33
33
|
'anthropic/claude-opus-4.5': 32_000,
|
|
34
34
|
'anthropic/claude-sonnet-5': 128_000,
|
|
35
|
-
'anthropic/claude-sonnet-4.6':
|
|
35
|
+
'anthropic/claude-sonnet-4.6': 128_000, // Anthropic-documented; gateway corrected 2026-07-21
|
|
36
36
|
'anthropic/claude-sonnet-4.5': 64_000,
|
|
37
|
-
|
|
37
|
+
// Anthropic documents 64000 for Haiku 4.5. Held at 16384 until 2026-07-21
|
|
38
|
+
// because the gateway clamped to 8192 — asking for more only inflated the
|
|
39
|
+
// price quote (quotedOutputTokens scales with the ceiling) while the reply
|
|
40
|
+
// stayed capped. blockrun#266 corrected the gateway and it is live, so the
|
|
41
|
+
// real ceiling is now reachable.
|
|
42
|
+
'anthropic/claude-haiku-4.5': 64_000,
|
|
38
43
|
'openai/gpt-5.6-sol': 128_000,
|
|
39
44
|
'openai/gpt-5.6-terra': 128_000,
|
|
40
45
|
'openai/gpt-5.6-luna': 128_000,
|
|
41
|
-
|
|
42
|
-
|
|
46
|
+
// 128000 is measured, not read off the catalog. The earlier probe that
|
|
47
|
+
// seemed to refute it was invalid — both SDKs rejected max_tokens > 100000
|
|
48
|
+
// client-side, so nothing reached a provider. Re-probed 2026-07-21 with that
|
|
49
|
+
// guard bypassed: both accept 128000, as did every model advertising a
|
|
50
|
+
// ceiling above 100000. The catalog was right; the measurement was broken.
|
|
51
|
+
//
|
|
52
|
+
// Franklin does not run that SDK guard on either request path — the agent
|
|
53
|
+
// loop and the proxy both use raw fetch and import only payment helpers from
|
|
54
|
+
// @blockrun/llm — so these values are independent of whether the SDK-side
|
|
55
|
+
// fix lands. (blockrun-llm#27 / blockrun-llm-ts#15, open at time of writing.)
|
|
56
|
+
'openai/gpt-5.5': 128_000,
|
|
57
|
+
'openai/gpt-5.4': 128_000,
|
|
43
58
|
'openai/gpt-5.4-mini': 128_000,
|
|
44
59
|
'openai/gpt-5.4-nano': 32_768,
|
|
45
|
-
|
|
60
|
+
// Probed accepted at 65536 through the live gateway (2026-07-21) — a real
|
|
61
|
+
// request, not a catalog reading. Note this is a FLOOR, not a proven
|
|
62
|
+
// ceiling: 65536 also happens to be ESCALATED_MAX_TOKENS, so nothing here
|
|
63
|
+
// could distinguish a higher true limit even if one exists.
|
|
64
|
+
'openai/gpt-5-mini': 65_536,
|
|
46
65
|
'google/gemini-2.5-pro': 65_536,
|
|
47
66
|
'google/gemini-2.5-flash': 65_536,
|
|
48
67
|
'google/gemini-3.5-flash': 65_536,
|
package/package.json
CHANGED