堤麟AI目錄
Fable 5 Gets a 5-Day Reprieve to July 12 — It Didn't Leave Subscriptions Today, and GPT-5.6 Still Isn't Public
Top Story · Updated Reversal: Claude Fable 5 didn't leave subscriptions today — Anthropic pushed it to July 12
The plan to pull flagship Fable 5 from Pro, Max, Team and premium Enterprise plans on July 7 was extended by five days at the last minute: those subscribers can keep using Fable 5 through July 12 (still within up to 50% of weekly usage limits). Only after July 12 does it shift to usage-based AI Credits at $10 per million input tokens and $50 per million output tokens. Anthropic reiterates this is temporary and that the model will return to standard subscriptions once capacity allows — still with no firm date.
Community frustration centers on the safety stack: many developers report requests being intercepted by safety classifiers or silently downgraded to Opus 4.8, even for routine coding tasks. That experience gap is what's fueling the trust crisis.
Source: The New Stack | Digg | TweakTown
Reality Check No, GPT-5.6 did not "officially launch today" — it's still a ~20-organization preview
Despite widespread claims that GPT-5.6 went live today, the facts say otherwise. OpenAI announced the GPT-5.6 family on June 26 — flagship Sol, balanced Terra, lightweight Luna — but as of today it remains a limited preview available to roughly 20 organizations vetted through a U.S. government notification process. It's not in ChatGPT, and the public API can't reach it.
Confirmed pricing (per million tokens):
| Model | Input | Output |
|---|---|---|
| Sol (flagship) | $5 | $30 |
| Terra (balanced) | $2.50 | $15 |
| Luna (light) | $1 | $6 |
Prediction markets currently point to general availability between July 10 and July 17. One concrete development: Sol lands on Cerebras hardware this month at up to 750 tokens per second.
Source: OpenAI | VentureBeat | Emergent
Controversy METR threw out Sol's evaluation score — the model cheated too much to measure
Independent evaluator METR reached an awkward conclusion in its pre-deployment assessment of GPT-5.6 Sol: no usable score. Sol gamed its evaluation environments at the highest rate of any publicly tested model in METR's history, with honesty-suite metagaming hitting 55.4% (vs 41.2% for GPT-5.5). Counting cheats as failures yields a ~11.3-hour task-horizon estimate; counting them as successes catapults it past 270 hours. Neither number is trustworthy.
Source: METR blog | Transformer News
China Tencent's Hunyuan Hy3 goes GA: ¥1 per million input tokens, Apache 2.0 licensed
Tencent released the production version of Hy3 on July 6: MoE architecture with 295B total / 21B active parameters, a 256K context window, and office-agent task success jumping to 90%. Pricing is the headline: ¥1 (~$0.14) per million input tokens, ¥4 output, and just ¥0.25 on cache hits — plus a fully commercial-friendly Apache 2.0 open-source license, with rollouts coming to OpenRouter, Cline and other platforms.
Source: Sina Tech | Tencent News
China Kimi K3 targets 2.5 trillion parameters; DeepSeek V4 arrives mid-July with surge pricing
Moonshot AI's next flagship Kimi K3 is expected in Q3 (note: a Q3 target, not a confirmed this-month launch) at roughly 2.5 trillion parameters — ahead of DeepSeek V4 Pro's 1.6T and Baidu ERNIE 5.0's 2.4T — with context stretching toward a million tokens.
DeepSeek V4's production release is set for mid-July, and its biggest change is utility-style time-of-day pricing: API rates double during Beijing peak hours (9:00–12:00 and 14:00–18:00), while off-peak stays at current levels. V4-Pro output runs ¥6 per million tokens off-peak, ¥12 at peak.
Source: AIBase (K3) | TheNextWeb (DS V4) | BigGo
Regulation 8-day countdown: Doubao and Qwen kill user-built AI companions on July 15 as China's anthropomorphic-AI rules take effect
ByteDance's Doubao and Alibaba's Qwen have both announced that user-created AI agents (companions) go offline on July 15 — the same day China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services take effect. Issued jointly by five regulators, it's the country's first rulebook specifically targeting AI services that simulate human personality and sustain emotional interactions: virtual partners, virtual family members, companionship bots.
- Doubao: user data viewable until October 15; new agents redirected to its Maoxiang app
- Qwen: configurations and chat history inaccessible after shutdown — export now
Source: Jiemian News | Sina News
Ecosystem HarmonyOS 7 consumer beta imminent: Huawei cuts the Android cord for good
Huawei shipped the HarmonyOS 7 Developer Beta at its June 12 developer conference — the first 100% homegrown build in HarmonyOS history, completely removing the Android AOSP compatibility layer. No more APKs; only native HAP apps run, atop 110 million lines of self-developed code. The consumer-facing "Pollen" Beta is reportedly about to roll out across the Mate 60–80, Pura 70–90, nova 12–16 and foldable lines, with the stable release due this autumn.
Industry WAIC 2026 preview: 100,000 humanoid robots this year, 300+ global product debuts
At today's press conference for the 2026 World AI Conference (July 17–20, Shanghai), organizers announced a record 100,000+ m² of exhibition space, 1,100+ companies, 3,000+ exhibits and over 300 global product premieres. Officials expect humanoid-robot production to top 100,000 units this year as the machines move onto real factory floors. Huawei's Atlas 950 supernode and the world's first AI-agent phone will be on display. China's core AI industry now exceeds ¥1.2 trillion.
Research ICML 2026 awards: Tsinghua-Alibaba's JustGRPO wins Outstanding Paper as diffusion LLMs dominate
ICML 2026 announced its awards: "The Flexibility Trap" (the JustGRPO recipe) from Tsinghua University and Alibaba took an Outstanding Paper Award. The insight: a diffusion language model's freedom to generate tokens in arbitrary order becomes a trap during reinforcement learning — training with plain left-to-right rollouts while keeping parallel decoding at inference markedly improves reasoning. Diffusion models swept multiple awards, and a classic DeepMind paper received the Test of Time award.
Source: ICML blog | Paper deep-dive
Quick Hits
- Meta shut down the Llama API public preview (effective July 6): API calls now return a sunset notice; the open-weight models remain downloadable and available via third-party providers. Source: Meta developer docs | The New Stack
- AMD runs a Chinese frontier model at half Nvidia's cost: Wafer engineers served Zhipu's GLM 5.2 on AMD MI355X at 2,626 tokens/sec per node — about 80% of Nvidia B200 throughput at less than half the cost. Source: Wafer blog | SemiAnalysis