JULY 20, 2026 — THE FIRST MONDAY AFTER FABLE 5 WENT CREDIT-ONLY
A migration deadline, a governance milestone, and a $400M bet on public AI infrastructure.
- DeepSeek API Deadline — 4 Days Left — July 24 at 15:59 UTC: deepseek-chat and deepseek-reasoner return errors. No grace period. The hidden trap: deepseek-reasoner maps to V4-Flash by default, not V4-Pro. Call v4-pro explicitly for heavy reasoning. V4-Flash now defaults thinking ON. If you have not migrated, do it today. Full migration guide + code diffs →
- EU Parliament Deploys EPGenAI Hub for MEPs — September 2026 — Politico: EU Parliament rolling out multi-model AI platform (Meta, OpenAI, Anthropic, Mistral) to elected members and staff. First major legislature to deploy frontier AI to elected representatives. The institution that wrote the EU AI Act will now use the technology it regulates. Full story →
- Current AI Secures $400M — Including $100M From France — for Public AI Infrastructure — TechCrunch: nonprofit Current AI building open, public AI compute outside hyperscaler control. Largest public-interest AI infrastructure commitment ever. The counter-narrative to Oracle-Stargate-TSMC week. Full story →
Story 1 — DeepSeek Migration: Do This Today
DeepSeek-chat and deepseek-reasoner will be fully retired and inaccessible after July 24, 2026, 15:59 UTC. The hard date is non-negotiable. After that, requests using the old names will fail. The migration itself is a one-line change per API call, same endpoint, same API key. But there are two traps that will catch teams who just do a find-and-replace without reading the migration notes.
Trap 1: deepseek-reasoner maps to Flash, not Pro. If your reasoning workloads depended on deepseek-reasoner for chain-of-thought quality, a direct alias swap to V4-Flash will degrade your outputs. Call deepseek-v4-pro explicitly. Trap 2: V4-Flash defaults to thinking mode enabled, unlike old deepseek-chat which was non-thinking by default. A direct model-name swap without adjusting the thinking parameter will change response latency, token counts, and billed output. For speed-sensitive or cost-sensitive workloads running high volume, add thinking disabled explicitly. Full migration guide with code diffs →
Story 2 — EU Parliament EPGenAI Hub: Governance Meets Frontier AI
The EU Parliament plans to roll out EPGenAI Hub to MEPs and staff as early as September, letting them access Meta, OpenAI, Anthropic, and Mistral models. The significance is institutional: the EU Parliament is the body that passed the EU AI Act — the most comprehensive AI regulatory framework enacted by any major government. MEPs who will be overseeing the implementation of that regulation will now use frontier AI daily for their own legislative work. This creates a feedback loop that should, in theory, produce better-informed regulation.
The multi-model choice — Meta, OpenAI, Anthropic, Mistral — is deliberate EU policy design: no single provider dependency, and inclusion of Mistral (French, European) alongside US labs. Anthropic's inclusion is notable given that Claude has been deployed to US government agencies through Project Glasswing. The open question is data handling: the EU Parliament's GDPR obligations and the data handling terms of US-based API providers need explicit reconciliation before September deployment. Full EPGenAI Hub analysis →
Story 3 — Current AI: The $400M Counterweight to Hyperscaler Consolidation
The nonprofit Current AI, backed by $400 million in commitments from many partners including $100 million from France, is funding open, public AI infrastructure. The timing is the most important context. This announcement comes the morning after Claude Fable 5 moved to credit-only pricing ($10/$50/M), the week after Oracle cut 30,000 jobs to fund Stargate ($500B), and the same month that TSMC posted record AI chip revenue. Those three events describe a world where AI infrastructure is consolidating around entities with hyperscaler capital. Current AI is a direct bet that public-interest institutions will fund an alternative — and that $400M is enough to make that alternative meaningful. Full Current AI analysis →
This Week — Three Deadlines and One Watch
July 24 (Thursday) — DeepSeek API hard cutoff. Every integration calling deepseek-chat or deepseek-reasoner breaks at 15:59 UTC. If you have not audited your codebase, search now. The migration is one line per call but the trap (reasoner → Flash, not Pro) requires a decision, not just a find-and-replace.
July 27 (Sunday) — Kimi K3 open weights. Moonshot AI confirmed weights by July 27. When they ship, self-hosting K3 on Western infrastructure becomes possible — eliminating the China National Intelligence Law concern for regulated industries. A self-hosted K3 with 1M context at your own compute cost changes the pricing picture entirely for high-volume workloads.
August 31 — Claude Sonnet 5 intro pricing ends. Sonnet 5 at $2/$10/M (intro) steps to $3/$15/M after August 31 — matching GPT-5.6 Terra. Build and benchmark your Sonnet 5 workflows now while the intro pricing is active. This is the most underrated deadline of the next six weeks.
Watch — Opus 5 / Honeycomb. Anthropic has not acknowledged or denied the Honeycomb EAP. Credits beginning does not rule out an Opus 5 announcement before July 31. Watch anthropic.com/news. If Opus 5 ships with better token efficiency than Fable 5, the credit economics change significantly — potentially dropping the effective per-task cost below $11.80.