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

[sciencebased]

Summary

• Escalation run 4753e843 (2026-05-30) flagged that the dataset tracked a non-existent Cohere SKU command-r-plus-2025 at $3/$15.
• Cohere's pricing page lists no "2025" Command R+. The canonical current SKU is Command R+ 08-2024 at $2.50 / $10; the older 04-2024 is $3/$15 (which the stale entry's price actually matched).
• Migrates the entry to command-r-plus-08-2024 (id, name, release 2024-08-30, pricing) and re-points the benchmark row and two use-cases. Adds a dated, source-cited research report.

Source: https://cohere.com/pricing

Scope notes

• No new provider added; no invented IDs (command-r-plus-08-2024 is the real Cohere SKU).
• The escalation's Anthropic (Claude Opus 4.8) and xAI (grok-4 → grok-4.3) findings were folded into the existing open PRs Data refresh: Anthropic Claude model context/output (2026-05-28) #7 and Data refresh: xAI grok-4 pricing (2026-05-29) #8 to avoid duplicate PRs.
• Google Gemini pricing re-verified and matches the dataset (no change). Meta Llama 4 Maverick metadata held back — no clean primary numeric source (llama.com robots-blocked; Azure Foundry lacks token figures).

Testing

• npm run typecheck passes
• npm run build passes (build artifacts reverted, not committed)
• No dangling references to the old command-r-plus-2025 id

🤖 Generated with Claude Code

[sciencebased]

Escalation 4753e843 re-verification (2026-06-01) — no data changes; documentation only. The Cohere correction on this branch is unaffected.

This branch's report is the central escalation report, so the remaining cross-cutting held-back items are recorded here:

Held back — DeepSeek V4 Pro post-promo pricing. The pricing page (https://api-docs.deepseek.com/quick_start/pricing) says the 75% promo for deepseek-v4-pro ends 2026-05-31 15:59 UTC and pricing "will be adjusted to 1/4 of the original price." It does not list explicit post-promo dollar amounts, and "75% off" and "1/4 of original" are the same ×0.25 multiplier — so the snippet is self-contradictory about whether the rate even changes. Per the conservative policy we do not infer an unstated rate: dataset left unchanged, and no DeepSeek PR opened (no existing PR touches DeepSeek; the bar of explicit published rates isn't met). Follow-up flagged: once DeepSeek publishes explicit standard rates, update the price and the model blurb (still reads "discounted 75% through 2026-05-31").

Held back — Grok Build 0.1 (see PR #8): specialty xAI SKU outside the locked provider/flagship strategy and missing schema-required metadata.

Added a dated addendum to research/research_report_2026-05-30.md; tsc -b --noEmit passes.

[sciencebased]

Mistral Large 3 repricing folded into this PR — 2026-06-03 (escalation 4753e843)

Change: mistral-large-3 pricing corrected from $2.00 in / $6.00 out → $0.5 in / $1.5 out per 1M tokens (batch updated to $0.25, derived as 50% of new input).

Source (official, primary): https://mistral.ai/pricing/ — verified 2026-06-03. Page lists Mistral Large 3 ("Open-weight, general-purpose, flagship multimodal and multilingual model") at $0.5 input / $1.5 output per 1M tokens.

Schema fit: Maps cleanly to the existing Pricing type and the existing mistral-large-3 model id. No new id, no schema change.

Batch discount caveat: Mistral states a general 50% batch discount but does not publish a per-model batch line for Large 3. The schema's optional batchInputPerMTokUsd was set to $0.25 (50% of the new $0.5 input), consistent with how the prior entry carried batch = 50% of input. This is a derived value from the stated general discount, not an explicit Mistral per-model figure — documented in the research report addendum.

Checks run:

• npm run typecheck (tsc -b --noEmit) — pass
• npm run build (tsc -b && vite build) — pass; build artifacts (dist/, *.tsbuildinfo) reverted, not committed.

Files changed:

• src/data/pricing.ts — mistral-large-3 entry
• research/research_report_2026-05-30.md — appended 2026-06-03 addendum with source citation

No new PR opened: per the daily-refresh duplicate-avoidance strategy, this correction was folded into existing open PR #9. PRs #6/#7/#8 left unmodified.

Commit: 1bf18a3