52 lines
5.2 KiB
JSON
52 lines
5.2 KiB
JSON
{
|
|
"_meta": {
|
|
"purpose": "exp15 pre-reg freeze — vendor facts (prices/cache/deprecation), official-doc verified 2026-07-17",
|
|
"method": "12-agent web workflow (6 research + 6 adversarial verify); verdicts recorded per topic",
|
|
"two_directions_rule": "vendor-doc (this file) + live /models listing (preflight_models.json) both required for slug verdicts",
|
|
"fetch_date": "2026-07-17",
|
|
"note_prices": "per 1,000,000 tokens, USD. cache-hit = cached-input read price. Gemini/grok tiered at 200k prompt."
|
|
},
|
|
"deprecation": {
|
|
"deepseek_legacy_retire_utc": "2026-07-24 15:59",
|
|
"deepseek_legacy_slugs": ["deepseek-chat", "deepseek-reasoner"],
|
|
"mapping": "BOTH route to deepseek-v4-flash (chat=non-thinking, reasoner=thinking). Reseller claim reasoner->v4-pro is WRONG per official.",
|
|
"arm_config_impact": "NONE — eval prod config uses deepseek-v4-flash (present in /models); v4-flash/v4-pro are the migration targets, unaffected.",
|
|
"verdict": "CONFIRMED",
|
|
"sources": ["https://api-docs.deepseek.com/news/news260424/", "https://api-docs.deepseek.com/quick_start/pricing/", "https://api-docs.deepseek.com/updates/"]
|
|
},
|
|
"cache_verified": {
|
|
"deepseek": {"type": "automatic/implicit", "usage_fields": ["prompt_cache_hit_tokens", "prompt_cache_miss_tokens"], "hit_frac_of_miss": {"v4-flash": 0.020, "v4-pro": 0.0083}, "verdict": "CONFIRMED"},
|
|
"zai_glm": {"type": "automatic/implicit", "usage_field": "usage.prompt_tokens_details.cached_tokens", "hit_frac_of_input": 0.20, "base_url": "https://api.z.ai/api/paas/v4", "ttl": "unspecified (reasonable time limits)", "min_prefix": "not documented", "verdict": "CONFIRMED — E.8-bis gate PASSES; Q2c may report cache-normalized column"},
|
|
"openai": {"type": "automatic", "min_tokens": 1024, "hit_frac_of_input": 0.10, "ttl_min": 30, "verdict": "CONFIRMED"},
|
|
"gemini": {"type": "implicit default (3.x)", "min_tokens_31pro": 4096, "hit_frac_of_input": 0.10, "verdict": "CONFIRMED"}
|
|
},
|
|
"prices": {
|
|
"deepseek-v4-flash": {"in": 0.14, "cache_hit": 0.0028, "out": 0.28, "role": "A0 draft", "notes": "thinking ON mandatory (echo mine); max_tokens>=8000; peak_surge_risk"},
|
|
"deepseek-v4-pro": {"in": 0.435, "cache_hit": 0.003625, "out": 0.87, "role": "Q4a strong-draft candidate", "notes": "peak_surge_risk"},
|
|
"glm-5": {"in": 1.00, "cache_hit": 0.20, "out": 3.20, "role": "A0 editor (D30.1 bilingual v3)", "ctx": "200K/128K", "notes": "PAYG live 2026-06-16 — verify key can call PAYG"},
|
|
"glm-4.6": {"in": 0.60, "cache_hit": 0.11, "out": 2.20, "role": "editor fallback"},
|
|
"glm-4.5-air": {"in": 0.20, "cache_hit": 0.03, "out": 1.10, "role": "cheap"},
|
|
"gpt-5.4": {"in": 2.50, "cache_hit": 0.25, "out": 15.00, "role": "judge candidate (expensive) / Q4a ceiling diag", "slug": "gpt-5.4-2026-03-05", "notes": "+10% data-residency uplift clause possible; reasoning_effort:low for long edit (quirks)"},
|
|
"gemini-3.1-pro-preview": {"in": [2.00, 4.00], "cache_hit": [0.20, 0.40], "out": [12.00, 18.00], "tier_at": 200000, "role": "judge candidate (apex, EXPENSIVE)", "notes": "thinking billed as output, additive to total; mandatory thinking (no budget:0)"},
|
|
"grok-4.3": {"in": [1.25, 2.50], "cache_hit": [0.20, 0.40], "out": [2.50, 5.00], "tier_at": 200000, "role": "judge candidate (cheap, cross-family)", "notes": "reasoning additive to output; >=200k ALL tokens 2x; judge uses reasoning ON"},
|
|
"mistral-large-latest": {"in": 0.50, "out": 1.50, "role": "Q4a strong-draft candidate / cheap judge candidate", "maps_to": "Mistral Large 3", "notes": "clean on archaic zh (no echo); cross-family"},
|
|
"mistral-medium-latest": {"in": 1.50, "out": 7.50, "maps_to": "Mistral Medium 3.5", "notes": "pricier than large now"},
|
|
"kimi-k2.6": {"in": 0.95, "cache_hit": 0.16, "out": 4.00, "role": "miner candidate / judge candidate", "notes": "temperature MUST be 1; verbose reasoning ~11k tok/para billed as output; max_tokens>=16000"}
|
|
},
|
|
"risks": {
|
|
"deepseek_peak_surge": {
|
|
"claim": "DeepSeek announced (SCMP + TheNextWeb, 2026-06-30) a peak-hour surcharge doubling V4 prices during Beijing business hours 09:00-12:00 & 14:00-18:00 (UTC+8) = 01:00-04:00 & 06:00-10:00 UTC.",
|
|
"status": "announced but NOT reflected on flat official docs pricing table as of 2026-07-17 (docs-lag or not-yet-effective)",
|
|
"verdict": "PARTIAL (reputable press, not on official table)",
|
|
"mitigation": "run paid DeepSeek arms OUTSIDE 01:00-04:00 & 06:00-10:00 UTC, or budget 2x for cache-miss input + output; re-check pricing page near 2026-07-24 cutover"
|
|
},
|
|
"gemini_thoughts_accounting": {
|
|
"claim": "official ref: totalTokenCount = prompt+thoughts+candidates (thoughts NOT in candidatesTokenCount); some forums claim API folds thoughts into candidates",
|
|
"mitigation": "verify empirically on a live gemini-3.1-pro-preview usageMetadata response before freezing judge-cost telemetry math (persist raw usage)"
|
|
},
|
|
"glm5_payg_access": {
|
|
"claim": "GLM-5 model doc gates to Coding Plan Pro/Max but PAYG per-token rates listed and PAYG went live 2026-06-16; glm-5 present in our /models",
|
|
"mitigation": "1-token live probe of glm-5 with our ZAI key at first paid gate confirms callability before A0 editor spend"
|
|
}
|
|
}
|
|
}
|