Die Sperren fielen, die neuen Schranken standen, und die offenen Modelle holten auf

Vor der nächsten Ausgabe handeln

Auf unserem Radar

• Fable 5 restored version triggers excessive safety guardrails — users call it 'nerfed' — Despite model unchanged, safety classifier false positives trigger Opus 4.8 fallback on non-safety coding. Words like 'security,' 'vulnerable,' 'hook,' and memory-managed languages trigger it. Fable 5 transitions to usage credits after Jul 7.

• Alibaba bans Claude Code effective July 10 — US-China AI tool decoupling escalates — Tit-for-tat: Anthropic blocked Chinese access after Jun 24 distillation accusation → Alibaba retaliates with corporate ban. First major tech company to officially ban a US AI coding tool on security grounds. Employees directed to migrate to in-house Qoder.

• Gemini 3.5 Flash Computer Use now ships natively — Scores 78.4 on OSWorld, on par with GPT-5.5. Enterprise safeguards: explicit confirmation for sensitive actions, auto-stop on prompt injection. Previously a standalone model — now a built-in tool.

• MAI-Code-1-Flash GA on GitHub Copilot Business and Enterprise — Microsoft's proprietary AI coding model now available for fast, low-latency code generation. Cuts Microsoft's OpenAI dependency — measure the difference.

• Claude Science: AI workbench for scientists (beta) — 60+ curated skills (genomics, proteomics, cheminformatics). NVIDIA BioNeMo integration. Reviewer agent checks citations and calculations. 50 research projects with up to $30K credits.

• DeepSeek V4 introducing peak/off-peak API pricing (mid-July) — First major LLM provider shifting from flat-rate to demand-based pricing. Peak-hour rates doubled vs off-peak. Utility-style AI pricing is here.

• AWS $1B + Microsoft $2.5B forward-deployed AI engineering orgs — industry-wide pattern — Four major players (AWS, Microsoft, OpenAI, Anthropic) now embedding AI engineers directly with enterprise customers to co-develop agentic AI. The race to capture enterprise AI deployment is on.

• Claude winning paid consumers — 75% growth since January — Consumer data shows paid AI users increasingly choosing Anthropic's Claude over ChatGPT. 3:1 course demand advantage. When paying users switch, enterprise procurement follows.

• Google AI brain drain: 4 top researchers departed in one week — Core Gemini researchers Adler + Pritzel → Anthropic. Shazeer → OpenAI. Nobel laureate Jumper → Anthropic same week. Talent retention crisis ahead of IPO race.

• Google capped Meta's Gemini access — AI compute crunch arrives — Google told Meta it couldn't provide full Gemini capacity. Disrupted internal Meta AI projects. First signal frontier AI compute is capacity-constrained even for the largest companies.

• GLM-5.2 matched Mythos 5 on cybersecurity — open-weight, MIT license, consumer hardware — Forbes: 'The bad guys now have a model as powerful as Mythos.' Anyone can download a Mythos-class cybersecurity model with no oversight. The containment argument just collapsed.

• BIS warns AI investment bust could trigger 2008-level financial crisis — Bank for International Settlements flags AI infrastructure overinvestment, circular financing, and poorly disclosed asset pledging. $2.7T erased from AI stocks in June.

• Tesla imposes $200/week AI spending cap — Grok/xAI exempt — Starting Jul 6, Tesla employees limited to $200/week on third-party AI tools. xAI products explicitly exempt. Engineers say staff prefer Claude. First large public company to implement a quantified AI spending cap.

• Anthropic Samsung custom chip talks + OpenAI Broadcom Jalapeño shipped — Both major labs pursuing chip independence from Nvidia. Jalapeño shipped nine months from announcement; Anthropic-Samsung still in discussion phase.

• Together AI raises $800M at $8.3B for open-source AI platform — Series C led by Aramco Ventures with NVIDIA, Vista Equity Partners. Open-source AI infrastructure attracting institutional capital.

• Anthropic publishes Fable 5 Cyber Jailbreak Severity (CJS) framework — 4-axis scoring with Glasswing partners (Amazon, Microsoft, Google). HackerOne bug bounty launched. First structured jailbreak severity taxonomy.

• Palantir CEO Karp slams token-based AI pricing as 'fundamentally broken' — Stock +8% on the commentary. Signals growing enterprise pushback against per-token billing. Combined with WhatsApp Business moving to token pricing Aug 1 — the model is expanding even as backlash grows.

• Meta plans AI compute cloud business to monetize excess infrastructure — Following SpaceX's move into AI compute. Would compete with AWS, GCP, Azure. Meta simultaneously telling staff AI agents are progressing slower than expected.

• Thariq Shihipar: 'You, not the model, are now the bottleneck' — Fable 5 engineer analysis from 400K Claude Code sessions: humans own 70% of planning decisions. Domain expertise predicts coding success better than model choice.

• Armin Ronacher: Claude models degrading on custom tool schemas — Opus 4.8 and Sonnet 5 worse than older Claude models at custom tool schemas in third-party harnesses. Hypothesis: RL optimization for Claude Code's own tool is degrading general-purpose reliability.

• Gemini 3.5 Pro — July 17 launch date firming up — Multiple independent leakers converging on July 17 with brand-new pretrain and 'Nano Banana Pro.' LMSYS Arena leak shows it beating Fable 5 High on coding.

• GPT-5.6 Sol — July 7 launch rumors — Codex popups showing Sol/Terra/Luna model selector with speed slider. Timing coincides with Fable 5 credit cliff. Speculative — watch for official confirmation.

• Claude on Microsoft Foundry GA — no EU data zone — Claude models GA on Foundry with Azure billing but no European data residency. Anthropic confirms data guarantees apply to Bedrock and Vertex AI only. EU banking/healthcare blocked.

• Claude Enterprise spend controls — model-level entitlements, spend alerts, group limits — Anthropic rolled out FinOps governance: model-level entitlements (beta), spend-threshold alerts, per-team analytics, group limits. Responds to reports of companies burning annual AI budgets in months.

• Amazon Mechanical Turk stops accepting new customers July 30 — AI replaces the original ML labeling platform — After 20+ years, AWS winding down Mechanical Turk. The platform that defined crowdsourced human data labeling is dying, replaced by AI. End-of-era milestone for ML infrastructure.