Data refresh: fix Cohere Command R+ SKU (2026-05-30)

research/research_report_2026-05-30.md

@@ -0,0 +1,70 @@
+# Daily AI vendor refresh — 2026-05-30 (escalation 4753e843)
+
+Locked providers: Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, xAI, Cohere.
+
+Escalation run 4753e843 flagged potential material updates across five vendors. Each candidate was re-verified against primary / authoritative official sources before any change. This report covers the **Cohere** correction made on this branch; the **Anthropic** (Opus 4.8) and **xAI** (grok-4 → grok-4.3) changes were folded into the existing open PRs (#7 and #8 respectively) to avoid duplicate PRs.
+
+## Diff & Freshness
+
+### Cohere — corrected (this branch)
+- The dataset tracked a non-existent SKU `command-r-plus-2025` ($3 / $15). Cohere's pricing page lists **no "2025" Command R+**. The real, current SKUs are **Command R+ 08-2024 ($2.50 in / $10 out)** and the older **Command R+ 04-2024 ($3 in / $15 out)**. Source: https://cohere.com/pricing
+- Decision: migrate the entry to the canonical latest **`command-r-plus-08-2024`** at **$2.50 / $10** (release 2024-08-30). Cross-references in benchmarks and use-cases updated to the new id. The old $3/$15 value matched the older 04-2024 SKU, confirming the stored id/price pairing was stale/inconsistent.
+
+### Anthropic — handled in PR #7 (data-refresh-2026-05-28)
+- Added **Claude Opus 4.8** (`claude-opus-4-8`): 1M ctx, 128K max output, $5/$25, cache read $0.50, batch $2.50/$12.50. It is now Anthropic's flagship ("NextOpus"); Opus 4.7 is legacy. Opus 4.6/4.5 (also $5/$25) are legacy SKUs not tracked — not added to avoid dataset bloat. Sources: https://platform.claude.com/docs/en/docs/about-claude/models , .../pricing
+
+### xAI — handled in PR #8 (data-refresh-2026-05-29)
+- `grok-4` was **retired 2026-05-15** and now redirects to **`grok-4.3`** at standard rates ($1.25 / $2.50, $0.20 cached, 1M ctx). The proposed $3/$15 change (from the x.ai/api marketing page) was rejected as a stale secondary source; the entry was migrated to the live `grok-4-3` id. Source: https://docs.x.ai/docs/models
+
+### Google — inconclusive earlier, now confirmed UNCHANGED
+- The escalation noted absent paid-tier prices in a prior fetch. Re-fetch of https://ai.google.dev/gemini-api/docs/pricing shows paid-tier prices that **match the current dataset exactly**: Gemini 2.5 Pro $1.25 in / $10 out (≤200k), Gemini 2.5 Flash $0.30 in / $2.50 out. **No change.**
+
+### Meta — held back (no clean primary metadata)
+- Official Meta (llama.com) remains robots-blocked. Azure AI Foundry's model catalog confirms Llama 4 Maverick exists as a natively-multimodal MoE but **does not publish numeric context window / max output** values. No reliable primary number found, so Llama 4 Maverick metadata is **left unchanged**. Source checked: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/models-featured
+
+## Models
+
+### Cohere
+- `command-r-plus-08-2024` (Command R+ 08-2024): 128K context; $2.50 input / $10 output per 1M tokens. Source: https://cohere.com/pricing
+
+## Pricing
+
+### Cohere
+- `command-r-plus-08-2024`: $2.50 input / $10.00 output per 1M tokens (list). Source: https://cohere.com/pricing
+
+## Benchmarks
+- No new benchmark source integrated; existing Command R+ row re-pointed to `command-r-plus-08-2024` (scores carried over).
+
+## Use-cases
+- Enterprise-RAG and regulated-EU use cases re-pointed from `command-r-plus-2025` to `command-r-plus-08-2024`.
+
+## Historic-usage
+- No dataset changes made today.
+
+---
+
+## Addendum — 2026-06-01 (escalation 4753e843 re-verification)
+
+Closing out the remaining escalation items. **No data changes** to this PR (Cohere correction stands as-is).
+
+### DeepSeek V4 Pro post-promo — held back (no explicit rate published)
+- The DeepSeek pricing page (https://api-docs.deepseek.com/quick_start/pricing) states the 75% promotional discount for `deepseek-v4-pro` ends **2026-05-31 15:59 UTC** and that pricing "will be officially adjusted to **1/4 of the original price** after the promotion ends."
+- The page does **not list explicit post-promo dollar amounts**. Critically, "75% discount" and "1/4 of the original price" are arithmetically the same multiplier (×0.25), so the snippet is self-contradictory about whether the rate actually changes. The current stored values ($0.435 in / $0.87 out / $0.003625 cache-hit) already equal 1/4 of the implied original.
+- **Decision: dataset left unchanged.** Per the conservative source-backed policy, we do not infer a new post-promo number that the primary source does not explicitly state. No DeepSeek PR was opened (no existing PR touches DeepSeek, and the bar for a new PR — primary docs explicitly listing current rates — is not met).
+- Follow-up: once the promo window has fully lapsed and DeepSeek publishes explicit standard rates, update `deepseek-v4-pro` pricing and refresh its model blurb (which still reads "discounted 75% through 2026-05-31"). Tracked as a held-back item, not actioned today.
+
+### Grok Build 0.1 — held back
+- Documented in PR #8's report (research/research_report_2026-05-29.md): specialty xAI coding SKU, outside the locked one-flagship-per-provider strategy and missing schema-required metadata. Not added.
+
+---
+
+## Addendum — 2026-06-03 (escalation 4753e843 — Mistral Large 3 repricing)
+
+### Mistral — corrected (this branch)
+- The dataset tracked `mistral-large-3` at **$2.00 in / $6.00 out** (batch $1.00). The official Mistral pricing page now lists **Mistral Large 3 at $0.5 input / $1.5 output per 1M tokens** ("Open-weight, general-purpose, flagship multimodal and multilingual model"). Source: https://mistral.ai/pricing/ (verified 2026-06-03).
+- **Decision: updated** `mistral-large-3` to **$0.5 / $1.5**. The change maps cleanly to the existing `Pricing` schema and model id — no new id, no schema change required.
+- **Batch discount:** Mistral's page states batch processing gets a general **50% discount** but does not enumerate a per-model batch line for Large 3. The schema has an optional `batchInputPerMTokUsd` field; the prior entry carried batch = 50% of input. To stay consistent and conservative, `batchInputPerMTokUsd` was updated to **$0.25** (50% of the new $0.5 input). This is a derived value from the stated general discount, not an explicit per-model figure from Mistral; noted here per the source-backed policy.
+- **No new PR opened.** Per the daily-refresh duplicate-avoidance strategy, this correction was folded into the existing open PR #9 (branch `data-refresh-2026-05-30`) rather than opening a new daily-refresh PR. PRs #6/#7/#8 left unmodified.
+
+### Other Mistral SKUs on the pricing page — not added
+- The page also lists Mistral Medium 3.5 ($1.5/$7.5), Small 4 ($0.1/$0.3), Devstral 2 ($0.4/$2), Codestral ($0.3/$0.9). These are outside the locked one-flagship-per-provider dataset strategy and were **not** added to avoid dataset bloat.

src/data/benchmarks.ts

@@ -181,7 +181,7 @@ export const benchmarks: Benchmark[] = [
     source: "⚠ vendor-reported. Reasoning-tier sibling of V4 Flash; pending independent eval.",
   },
   {
-    modelId: "command-r-plus-2025",
+    modelId: "command-r-plus-08-2024",
     reasoning: 76,
     coding: 70,
     math: 65,

src/data/models.ts

@@ -287,11 +287,11 @@ export const models: Model[] = [
       "DeepSeek's reasoning-tier V4. Currently discounted 75% through 2026-05-31 — pricing will step up afterwards.",
   },
   {
-    id: "command-r-plus-2025",
-    name: "Command R+ 2025",
+    id: "command-r-plus-08-2024",
+    name: "Command R+ 08-2024",
     provider: "Cohere",
     family: "Command R+",
-    releaseDate: "2025-06-12",
+    releaseDate: "2024-08-30",
     contextWindowTokens: 128_000,
     maxOutputTokens: 8_000,
     modalities: ["text", "tool-use"],

src/data/pricing.ts

@@ -85,9 +85,10 @@ export const pricing: Pricing[] = [
   },
   {
     modelId: "mistral-large-3",
-    inputPerMTokUsd: 2.0,
-    outputPerMTokUsd: 6.0,
-    batchInputPerMTokUsd: 1.0,
+    inputPerMTokUsd: 0.5,
+    outputPerMTokUsd: 1.5,
+    batchInputPerMTokUsd: 0.25,
+    notes: "Official Mistral pricing (mistral.ai/pricing, 2026-06-03): $0.5/$1.5 per 1M tokens; batch processing gets 50% discount.",
   },
     {
     modelId: "deepseek-v4-flash",
@@ -104,9 +105,9 @@ export const pricing: Pricing[] = [
     notes: "75% promotional discount through 2026-05-31 15:59 UTC (per DeepSeek API docs).",
   },
     {
-    modelId: "command-r-plus-2025",
-    inputPerMTokUsd: 3.0,
-    outputPerMTokUsd: 15.0,
-    notes: "Cohere Command R+ pricing (Cohere pricing page).",
+    modelId: "command-r-plus-08-2024",
+    inputPerMTokUsd: 2.5,
+    outputPerMTokUsd: 10.0,
+    notes: "Cohere Command R+ 08-2024 list pricing per cohere.com/pricing ($2.50 in / $10 out per 1M tokens).",
   },
 ];

src/data/use-cases.ts

@@ -25,7 +25,7 @@ export const useCases: UseCase[] = [
     id: "enterprise-rag",
     title: "Enterprise RAG with citations",
     category: "rag",
-    recommendedModels: ["command-r-plus-2025", "claude-sonnet-4-6", "gemini-2-5-pro"],
+    recommendedModels: ["command-r-plus-08-2024", "claude-sonnet-4-6", "gemini-2-5-pro"],
     why:
       "Cohere's Command R+ is purpose-built for grounded answers with citations. Gemini 2.5 Pro shines when the corpus is pasted into a single 1M+ token context. Sonnet is the safe middle.",
     costStrategy:
@@ -65,7 +65,7 @@ export const useCases: UseCase[] = [
     id: "regulated-eu-deployment",
     title: "Regulated EU / data-residency deployment",
     category: "support",
-    recommendedModels: ["mistral-large-3", "command-r-plus-2025"],
+    recommendedModels: ["mistral-large-3", "command-r-plus-08-2024"],
     why:
       "Mistral hosts in EU regions. Cohere offers private deployment options for regulated industries (finance, healthcare).",
     costStrategy: