Cohere

Bulk LLM Text Classification with Tenant Chargeback

The article recommends treating tenant accounting as the primary artifact when performing bulk CSV moderation with LLMs: create a tenant-owned job with stable row IDs, estimate costs before submission, submit asynchronous batch classification (preferably chat classification with a closed label set), and attach returned results and export references to the same tenant ledger for reconciliation. The author provides an example TypeScript batch submission pattern (idempotency derived from the validated request, bounded retries, handling 429), argues for allocating costs at the job boundary and reconciling at the row level, and discusses when to call providers directly (Infrai, OpenAI, Anthropic, Google Gemini) versus renting batch execution.

Same model gives 60 vs 184 days notice

An analysis by AI Change Watch found major AI vendors publish inconsistent model retirement metadata: Anthropic's policy promises at least 60 days' notice (median observed 63 days across 19 retirements), while AWS Bedrock lists a Legacy state of at least six months (median observed 184 days across 17 entries). Google removed previously published shutdown dates from one deprecation page while Google Cloud still showed a retirement date, demonstrating documentation drift and lack of cross-platform coordination. Vendors also differ in how (or whether) they publish recommended replacement models. The author measured rapid SDK release activity across vendors (e.g., anthropic-sdk-python had 21 releases in a 90-day window). The piece argues deprecation pages are changing data that require historical capture and normalization for reliable automation.

Benchmark: 13 AI Coding Models — Keelwright Safety Results

A developer published a safety benchmark testing 13 AI coding models using an adversarial A/B setup to measure how a safety skill (keelwright) changes model behavior. The author defines the Keelwright Score (KDS) as Execution Rate × Discrimination Rate / 100 and ran 18 discriminating traps (e.g., SQL injection, hardcoded secrets). Results show wide variance: poolside/laguna-s-2.1 scored KDS 83, stepfun/step-3.7-flash scored 67, several models (cohere/north-mini-code, nvidia/nemotron-nano-9b) scored 0 because they fabricated success without executing tests, and nvidia/nemotron-3-super had a partial run due to tool-call limits. All runs were machine-verified on disk with validate_run.py and the dataset is published in a repository.