AUGUST 3, 2026 — KIMI SUNSET · ANTHROPIC OPEN-WEIGHTS · DEADLINE CALENDAR
- Kimi API sunset August 31: kimi-k2.5 and moonshot-v1 series confirmed discontinued August 31. Already blocked for new users since K3 launch July 16. Migrate to kimi-k3 (general) or kimi-for-coding (coding) now. K2.5 open weights on Hugging Face unaffected. US Treasury sanctions warning on Moonshot still unresolved — verify before expanding API commitments. Full migration guide →
- Anthropic open-weights position (July 27): Published the same day Kimi K3's 2.8T weights went live. Not a ban call — argues open weights are beneficial at current capability levels, but the risk/benefit ratio changes near WMD-uplift thresholds. Commits to conditional release framework. Explicitly endorses transparency, accessibility, security research, and economic access benefits. Full analysis →
- August deadline calendar — three dates to track:
• August 5 — Grok Voice Think Fast 2.0 auto-migration (pin grok-voice-think-fast-1.0 before then to stay on v1.0)
• August 31 — Kimi kimi-k2.5 / moonshot-v1 full sunset
• August 31 — Claude Sonnet 5 intro pricing ends ($2→$3/M + tokenizer adds ~35% tokens)
Story 1 — Kimi Legacy API: Migrate Before August 31
Per Moonshot's official model documentation, kimi-k2.5 and the moonshot-v1 family (8k, 32k, 128k, auto) will be fully discontinued August 31, 2026. Both strings are already unavailable to newly registered users. Teams on legacy Kimi strings have 28 days to migrate. The recommended path: search your codebase for moonshot-v1 and kimi-k2.5, replace with kimi-k3 for general tasks or kimi-for-coding for coding-specific workloads, and test on staging before the deadline. K2.5 open weights on Hugging Face are unaffected. Full migration guide →
Story 2 — Anthropic's Open-Weights Position: What It Means
Anthropic's July 27 position paper is the most detailed statement any frontier lab has published on when open-weight release is appropriate and when it is not. The core argument: open weights are beneficial (transparency, accessibility, security research, economic access) at current capability levels. The risk/benefit ratio changes as models approach thresholds for serious uplift in weapons of mass destruction or autonomous cyberattacks. Anthropic says it is not there yet but commits to a conditional framework rather than an unconditional position in either direction. This puts Anthropic between Meta (unconditional open weights, always) and OpenAI (closed API only, never). The practical impact for developers: no Anthropic frontier model weights will be released. The impact for policy: Anthropic has now publicly committed to the framework it will use when it decides the answer is no — before it gets there. Full analysis →
August Deadline Calendar — Three Dates Every API Team Needs
August 5 — Grok Voice Think Fast 2.0 auto-migration
grok-voice-latest auto-upgrades to Think Fast 2.0 on August 5. To stay on v1.0: pin grok-voice-think-fast-1.0 before then. Price change: $0.05→$0.08/min.
August 31 — Kimi legacy API sunset
kimi-k2.5 and moonshot-v1 series (8k/32k/128k/auto) stop responding for all users. Migrate to kimi-k3 or kimi-for-coding now.
August 31 — Claude Sonnet 5 intro pricing ends
$2/$10/M → $3/$15/M (+50%). New tokenizer adds 10-35% more tokens per equivalent text. Front-load batch jobs now. Enable prompt caching.
October 24 — DeepSeek legacy strings sunset
deepseek-chat and deepseek-reasoner discontinued. Migrate to deepseek-v4-flash now — 82.7% Terminal-Bench, $0.14/M.
Story 3 — Qwen3.8-Max: What Alibaba Confirmed, What It Did Not
Several AI news summaries on August 3 described Qwen3.8-Max as a new launch with $2/$6/M pricing, open weights arriving "next week," and benchmark results proving it leads on coding. None of those claims are accurate. Alibaba previewed qwen3.8-max-preview on July 19, 2026 at the World AI Conference in Shanghai — 15 days ago. As MarkTechPost documented on launch day, Alibaba shipped the announcement with no benchmark table, no model card, no active parameter count, and no per-token API price. The $2/$6/M figure circulating on X is Grok 4.5's price — not confirmed for Qwen3.8-Max. Alibaba's own predecessor Qwen3.7-Max is priced at $2.50/$7.50/M, which is the only confirmed reference point. Open weights are promised "soon" with no date, no license, and no HuggingFace repository as of August 3. Alibaba's performance claim — "second only to Fable 5" — is a self-assessment with no published benchmark to support it. The preview is accessible via Token Plan at ~10% of standard pricing and is worth testing. Do not migrate production workloads until Alibaba publishes a benchmark table, active parameter count, per-token pricing, and open-weight checkpoint. Full fact-check →