#1 2026-08-12
DeepSeek V4-Flash vs Qwen3.8-Max (comparison), Kimi K3 vs V4-Flash (comparison)
Correction: Removed the unverified "V4-Flash-Max — the pricier April finetune at $2.80/$11.20" claim and every sentence deriving from it (variant bullet, pricing-table cell, "1.3x vs Max" and "20x variant" premises). Replaced with the canonical statement: "Max" is a max-effort evaluation setting of the single V4-Flash-0731 model; no separate Flash-Max variant exists.
Why: Live verification against official DeepSeek documentation: the API lists exactly two models (deepseek-v4-flash → DeepSeek-V4-Flash-0731, deepseek-v4-pro); Hugging Face hosts no Flash-Max repository; DeepSeek's own harness note identifies "Max" as an evaluation configuration. The $2.80/$11.20 figure could not be confirmed by any official source.
Canonical reference: /news/deepseek-v4-flash-0731-vs-max/
#2 2026-08-12
Best Open-Weight Coding Models 2026 (hub), DeepSeek V4-Flash vs Qwen3.8-Max (#1), Kimi K3 vs V4-Flash (#2)
Correction: Hub table cells corrected to official metadata: Kimi K3 "1.2T-class / MIT (modified)" → "104B of 2.8T / Kimi K3 license (custom)"; GLM-5.2 "355B total (48B active in INT4)" → "753B total (8 experts per token) / MIT / $1.40–$4.40 (Mistral-hosted)"; Llama 4 Maverick → "404B total / Llama 4 Community"; GLM "EU and US regional endpoints" → global + EU only. Comparison #1: Qwen "Official cards ~$2.00/$6.00" cell → "CNY 14.988/44.965 (≈$2.1/$6.3)". Comparison #2: license cell and both conclusion sentences rewritten — the "no terminal sweep" claim removed (the official card publishes GPQA 93.5 / DeepSWE 67.5 / TB2.1 88.3 on the Kimi Code harness). Qwen launch benchmark numbers reframed as vendor-reported: the official DeepSWE board measured qwen3-8-max at 57.5 ± 3 (xhigh), outside the vendor claim's confidence interval.
Why: Official pages fetched 2026-08-12: moonshotai/Kimi-K3 Hugging Face metadata (license tag "other", 2.78T params); deepswe.datacurve.ai official DeepSWE board (kimi-k3 68.5 max, qwen3-8-max 57.5 ±3, deepseek-v4-flash 53.3, glm-5-2 43.8); zai-org/GLM-5.2 card + config.json (753.3B total, 8 experts per token, MIT, 1M context); docs.mistral.ai GLM-5.2 rate card (global + EU, US pending); meta-llama/Llama-4-Maverick metadata (404B, llama4community license).
Canonical reference: /news/qwen3-8-max-vs-kimi-k3/ and /news/open-weight-llm-api-pricing-2026/
#3 2026-08-16
Claude Sonnet 5 Ends the Opus Monopoly (news)
Correction: Removed the "intro $2/$10 expires August 31, rising to $3/$15" claim and every sentence deriving from it (description, "pricing cliff" bullet, "50% jump" and "gets re-rated" premises, "September 1 — whether pricing holds" watch item). Replaced with the canonical statement: Anthropic made Sonnet 5's $2/$10 introductory pricing permanent on August 10, 2026; the scheduled September 1 increase to $3/$15 no longer applies. Title updated ("Pricing Sunsets Too" → "$2/$10 Now Permanent").
Why: Official Anthropic sources, verified 2026-08-16: the announcement changelog ("Edit August 10, 2026: Sonnet 5's introductory pricing of $2 ... is now permanent. The standard pricing of $3 / $15 previously set to take effect September 1 no longer applies.") and the Claude Platform release notes (August 10, 2026: "the previously scheduled increase to $3 / $15 per MTok on September 1, 2026 will not occur"). The ai-brief-2026-08-11 entry was already correct.
Canonical reference: /news/ai-brief-2026-08-11/