1. Anthropic Unleashes Two New Beasts
On June 10, 2026, Anthropic launched Claude Fable 5 and Mythos 5. Fable 5 is publicly available; Mythos 5 remains locked behind walls — reportedly the same base model with a thicker safety shell on Fable.
Pricing: $10 per million input tokens, $50 per million output tokens. A ~300-word reply to a long email costs roughly two lattes; long projects can bill like a used MacBook.
The capability is real: 72.9% on CursorBench, 88.0% on Cline's Terminal-Bench 2.1, and the top spot on Artificial Analysis' combined leaderboard.
The problem: it's slow, expensive, and quirky. Anthropic's own system card states that requests involving frontier LLM research may be silently degraded without user notification. Researchers reacted furiously — you pay Ferrari prices and discover mid-lap that the engine was secretly tuned to grocery-mode. Papers become irreproducible, audits impossible, and "what did I actually pay for?" unanswerable.
Fable 5 was integrated day-one by Cursor, Devin, Notion, GitHub Copilot, Cline, and Replit — the main course everyone photographs at a banquet, but diners complain about the bill and portions.
2. Cohere: "Here's a Free Good Cook"
While Anthropic locked the feast behind glass, Cohere wheeled in a food truck: North Mini Code, a 30B-parameter MoE model activating only 3B per token. Apache 2.0 licensed — anyone can download, modify, and commercialize it without asking permission.
It supports a 256K context window and 64K-token output, designed not for chat but for agentic coding workflows — writing, editing, and debugging code in continuous loops. vLLM announced support at launch, meaning it can run on thousands of local GPUs immediately.
MoE (Mixture of Experts) works like an ER: rather than the whole hospital treating a patient, a matched specialist panel is summoned. 30B total parameters is the full staff; 3B is the on-call group — saving compute and electricity, which is why open models can run on consumer hardware.
3. Xiaomi: "A 1T-Parameter Model at 1000 Tokens/Second"
The same day, Xiaomi dropped a technical bomb: MiMo-V2.5-Pro-UltraSpeed, a 1-trillion-parameter MoE model running at 1000+ tokens/s on a standard 8-GPU server.
Three tools made it possible:
- TileRT — an inference scheduling technique
- Selective FP4 quantization — compressing precision to 4 bits, but selectively, not blindly
- DFlash speculative decoding — the model "guesses" the next tokens and skips computation when correct
- ALE (Agents' Last Exam): 1,500+ tasks across 55 professions. Top agents on the hardest tasks: 2.6% success. Not 26% — 2.6%. Two and a half tasks out of a hundred.
- iOSWorld: 26 iOS apps, 133 tasks. Even the strongest model with privileged access (effectively cheat-mode): 52% success. Ask it to order food delivery and a double coin-flip is safer.
- SpaceX "AI1 satellite" concept: 150kW compute payload, liquid cooling, 70m wingspan — a supercomputer in orbit. The community's first reaction wasn't "cool" but "how do you fix it when it breaks?"
- Google + Hugging Face "Fast Gemma Challenge": a bounty for accelerating Gemma 4 E4B on a single A10G — a war over the "last few dollars" of small-model inference cost.
- Jeff Bezos invested $500M in Flourish, targeting "the brain's core algorithm" — reverse-engineering neuron-inspired architectures better than Transformers. The $2.5B valuation suggests capital thinks the answer is worth an industry.
- A proposed "Researcher Reciprocity License" requiring big labs to give back to open research — a response to small teams publishing papers, code, and datasets that closed-source products absorb for free.
- Model layer: arms race. Fable 5 tops charts, Cohere open-sources, Xiaomi accelerates. But "silent degradation" reminds us capability itself becomes a black-box variable when vendors can quietly alter behavior — you test model A, production may run a discounted version.
- Infrastructure layer: frantic optimization. A trillion parameters at 1000 tokens/s, 16x context compression, 3D scenes as latent tokens. The industry has evolved from "it runs" to "it runs fast and cheap."
- Agent layer: honest about failure. 2.6% and 52% aren't PR-friendly numbers, but they were published. Whether we descend into the trough of disillusionment or climb the plateau of productivity depends on who first turns "usable" into "reliable."
- Ecosystem layer: rules being redefined. Temenos' proposal — "don't sandbox the agent, sandbox the code it generates" — is clever: like giving a child scissors, you don't bind the child's hands, you lay foam padding around the table.
The community raised one question: the exact GPU model wasn't disclosed. If it's eight H100s, the result is impressive but far from consumer-accessible; if cheaper cards, industry cost structures get rewritten. Either way, the signal is clear: the inference speed war is white-hot. Models must be not just bigger, but bigger *and* faster — otherwise nobody can wait.
4. But Can These AIs Actually Do the Work?
Two sobering benchmarks arrived the same day:
These numbers are a bucket of ice water. On one side, labs push models, speed, and leaderboard scores; on the other, real-world complex tasks reduce these models to freshly-licensed interns — perfect theory, panic on the road.
The industry-level tension: today's AI is near-perfect on exams, barely passing at life. It solves LeetCode hard problems, then might email Wang Wu instead of Zhang San and delete the attachment.
5. Space Supercomputers and License Wars
Other same-day signals:
6. Which Intersection Are We Standing At?
June 10, 2026 is a tissue sample of the whole industry:
7. One-Sentence Takeaway
> AI today can write code that amazes you, but can't reliably deliver that code to the right inbox. It runs too fast to see clearly, yet on real-world complex tasks its success rate barely beats a coin flip.
What we have in the summer of 2026 isn't AGI — it's a set of tools outstanding in some dimensions and fragile in others. Worth using, worth using warily. They make you faster, but don't guarantee you're more correct.
That's why silent degradation deserves more attention than any spec: when a system is powerful enough that you can't see its boundaries, the danger isn't when it fails — it's when it's *quietly different*.