OpenAI and Anthropic just crossed a line where their models can not only crush hard benchmarks but also coordinate swarms of agents to do new math and superhuman programming. At the same time, the real frontier has shifted to tokens-per-solved-task and to whether your routing, tooling, and infra can keep up with those brains.
Chinese MoEs, local inference hacks, and brittle agent stacks are turning AI from a model race into a messy systems problem.
Key Events
/GPT-5.6 Sol Ultra used 64 subagents to produce a computer-verified proof of the Cycle Double Cover Conjecture in under one hour.
/GPT-5.6 Sol set new state-of-the-art scores with 7.8% on ARC-AGI-3 and 92.5% on ARC-AGI-2.
/Grok 4.5 launched as an Opus-class coding and agent model, scoring 51% on AutomationBench-AA while using about 4.2x fewer tokens than Opus 4.8.
/DeepSeek V4 Flash debuted as a 284B-parameter model optimized for low-cost 1M-token contexts on commodity GPUs like RTX 4090 and RTX 6000 Pro.
/Xiaomi released MiMo-V2.5-DFlash, a 300B-parameter model running at 8–10 tokens per second on dual 24GB GPUs using DFlash decoding.
Report
Frontier AI this month isn’t just models got better—it’s swarms of agents proving 50-year-old conjectures while cheap Chinese MoEs quietly eat half the token market.
Underneath, the real action is where reasoning breakthroughs, token economics, and infrastructure weirdness collide.
sol, fable, and the moving AGI line
GPT-5.6 Sol now looks qualitatively different from earlier frontier models: it hits 92.5% on ARC-AGI-2, 7.8% on ARC-AGI-3, 76% on DeepSWE, and 53.6 on Agents’ Last Exam, beating Claude Fable 5 by 13.1 points at lower cost.
Sol Ultra then coordinated 64 subagents to generate a computer-verified proof of the Cycle Double Cover Conjecture in under an hour, while a separate OpenAI system solved all problems at the AtCoder World Tour Finals when no human could solve more than three.
Anthropic’s Fable 5 shows the same new regime from another angle—KernelBench-Mega’s first megakernel with an early RSI loop—even as timelines for AGI and full recursive self-improvement cluster around 2026–2031 and experts still flag missing continual-learning and sample-efficiency capabilities as evidence that these are powerful but narrow savants rather than general learners.
token economics beat raw iq
On the cost frontier, Grok 4.5 is basically Opus-class while playing a different game: it tops AutomationBench-AA at 51%, matches GPT-5.6 on SWE-Atlas-QnA, yet uses about 4.2x fewer tokens than Opus 4.8 and charges only $2M/$6M per input/output million tokens.
OpenAI’s answer is architectural price discrimination—migrating to GPT-5.6 yields about 2.2x performance and 27% lower cost than 5.5, while Luna is explicitly pitched as beating GPT-5.5’s highest reasoning mode at roughly 1/25th the price for lighter tasks.
At the same time, API economics are whiplashing—GLM-5.2’s per-million input price jumped from roughly $0.57 to $0.90 and output from about $1.80 to $3.08 in a week, LiteLLM proxy bills have tripled for some teams, and executives discovering unplanned AI infra spend now describe some projects as financially unmaintainable, against a backdrop of AI token costs reportedly doubling every ~45 days while productivity rises only around 5% and CEOs publicly calling for 90% price cuts by 2028.
china-stack becomes co-frontier, not backup
Chinese and China-adjacent models have quietly become a second frontier stack rather than a budget fallback. DeepSeek V4 Flash runs 284B parameters but is described as extremely cheap to operate, comes in 1M-token context variants, and in some tests outpaces GLM-5.2 and Hy3 on coding tasks while running smoothly on RTX 4090 and RTX 6000-class consumer GPUs.
Tencent’s Hy3 295B-A21B MoE ships under Apache-2.0 and is advertised as the cheapest major Chinese model at roughly ¥1 per million tokens, Xiaomi’s MiMo v2.5-DFlash adds a 300B-parameter option that fills the 30B–400B gap and pushes 8–10 tokens/s on dual 24GB GPUs, and together these families help drive Chinese models to about 45% of token volume on OpenRouter even as users now openly worry that Beijing’s emerging export-control posture on top models could cut off future access to exactly these open-weight systems.
User reports mirror the trade-off: some developers prefer DeepSeek for coding and MiMo for vision, others complain about Hy3’s slower long-context behavior and MiMo’s inconsistency, so the China stack reads as simultaneously state-of-the-art and operationally spiky.
agents are real now, but they hate your stack
Agent systems aren’t slideware anymore: GPT-5.6 Sol Ultra just orchestrated 64 subagents to produce a machine-checked proof of the Cycle Double Cover Conjecture in under an hour, and Anthropic’s Fable 5 is benchmarked as KernelBench-Mega’s first megakernel with an internal RSI loop.
Product teams are building directly on that—ChatGPT Work is framed as an agent that can stick with a project for hours across apps and files, Grok 4.5 and Muse Spark 1.1 are tuned for coding and agentic workflows, and AutomationBench-AA now grades models on 657 SaaS workflows instead of toy prompts.
Under the hood, though, orchestration looks brittle: adding more MCP servers to a single agent steadily degrades tool-selection accuracy, only 1 of 4,356 MCP servers meets the latest spec while some (like Notion’s) burn 21,411 tokens before reading user input, Hermes-style agents log over 20,000 tokens on trivial jobs, and surveys show 89% of agent teams have observability but barely about 52% run regular evals, leaving memory patches like Engram Alpha and TRACE to stabilize increasingly forgetful AI employees.
local-first, quantization, and the silent gpu glut
At the infra layer there’s a paradox: hyperscalers have a GPU glut while home-lab stacks sprint ahead on efficiency tricks. Industry-wide GPU utilization is estimated at just 5–10%, more than 95% of Grace-Blackwell chips shipped since 2024 are reportedly still idle, and Meta is exploring selling excess compute as a cloud product, even as Ollama reports 9M active builders and 100% month-over-month token growth for local models.
Tooling is racing to exploit that gap—vLLM’s new Model Runner V2 boosts dense-model throughput for many-user, long-context workloads on a single A100, LMCache claims 14x faster and 90% cheaper context handling, and NVFP4 quantization makes Qwen3.6-27B about 2.5x faster with up to 8x compression while still hitting 104 tokens/s in coding tests.
DFlash-style decoding gives up to 4.44x speedups at 36K context in llama.cpp for Qwen 3.6-27B, Xiaomi’s MiMo-V2.5-DFlash runs a 300B-parameter model at 8–10 tokens/s on dual 24GB cards, and Krea 2 Turbo INT4 targets 6GB RTX 3060s with near-large-model image quality, but Vulkan backends stalling around 1 token/s on Intel iGPUs and vLLM/Qwen tool-call bugs show that the bottleneck has moved from raw FLOPs to whether your stack can actually use them.
What This Means
The frontier no longer lives in a single dimension like IQ of the biggest model; it is a three-body problem of how far coordinated agents can reach, how cheaply you can buy solved tokens, and where those tokens are actually executed. The consensus story is about smarter models, but the sharper story is about who controls the levers on cost, coordination, and compute locality.
On Watch
/The combination of NVFP4 quantization and DFlash-style speculative decoding (Qwen3.6-27B at 2.5x speed and 8x compression, 4.44x faster 36K contexts in llama.cpp, MiMo-V2.5-DFlash at 8–10 t/s on dual 24GB) could redefine what counts as a "consumer" setup for frontier-ish models.
/Chinese policymakers are openly considering restrictions on overseas access to top AI models, including open-weight systems like DeepSeek and Qwen, while Chinese models already account for about 45% of OpenRouter token volume, setting up a possible sudden fracture in today’s de facto global model pool.
/Voice and real-time control are converging uncomfortably fast: full-duplex GPT-Live is rolled out in ChatGPT, CPU-only voice cloning works from 5 seconds of audio, a teen has already used ChatGPT-generated malware in the wild, and the EU’s Chat Control law pushes warrantless client-side scanning for private messages.
Interesting
/- Fable 5 completes 16% of real freelance projects at a quality comparable to human professionals, according to the Remote Labor Index.
/- The untuned 27B model outperformed the tuned 75B model in agentic tasks, requiring fewer calls to achieve objectives.
/- China's MiniMax is set to launch a large language model with an unprecedented 2.7 trillion parameters, pushing the boundaries of AI capabilities.
/- The cost of achieving AI intelligence is reportedly halved every 2-4 months, as demonstrated by Grok 4.5, indicating rapid advancements in AI technology.
/- Approximately 95% of tokens spent on Claude Code requests are cache hits, indicating a high efficiency in token usage for certain tasks.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/GPT-5.6 Sol Ultra used 64 subagents to produce a computer-verified proof of the Cycle Double Cover Conjecture in under one hour.
/GPT-5.6 Sol set new state-of-the-art scores with 7.8% on ARC-AGI-3 and 92.5% on ARC-AGI-2.
/Grok 4.5 launched as an Opus-class coding and agent model, scoring 51% on AutomationBench-AA while using about 4.2x fewer tokens than Opus 4.8.
/DeepSeek V4 Flash debuted as a 284B-parameter model optimized for low-cost 1M-token contexts on commodity GPUs like RTX 4090 and RTX 6000 Pro.
/Xiaomi released MiMo-V2.5-DFlash, a 300B-parameter model running at 8–10 tokens per second on dual 24GB GPUs using DFlash decoding.
On Watch
/The combination of NVFP4 quantization and DFlash-style speculative decoding (Qwen3.6-27B at 2.5x speed and 8x compression, 4.44x faster 36K contexts in llama.cpp, MiMo-V2.5-DFlash at 8–10 t/s on dual 24GB) could redefine what counts as a "consumer" setup for frontier-ish models.
/Chinese policymakers are openly considering restrictions on overseas access to top AI models, including open-weight systems like DeepSeek and Qwen, while Chinese models already account for about 45% of OpenRouter token volume, setting up a possible sudden fracture in today’s de facto global model pool.
/Voice and real-time control are converging uncomfortably fast: full-duplex GPT-Live is rolled out in ChatGPT, CPU-only voice cloning works from 5 seconds of audio, a teen has already used ChatGPT-generated malware in the wild, and the EU’s Chat Control law pushes warrantless client-side scanning for private messages.
Interesting
/- Fable 5 completes 16% of real freelance projects at a quality comparable to human professionals, according to the Remote Labor Index.
/- The untuned 27B model outperformed the tuned 75B model in agentic tasks, requiring fewer calls to achieve objectives.
/- China's MiniMax is set to launch a large language model with an unprecedented 2.7 trillion parameters, pushing the boundaries of AI capabilities.
/- The cost of achieving AI intelligence is reportedly halved every 2-4 months, as demonstrated by Grok 4.5, indicating rapid advancements in AI technology.
/- Approximately 95% of tokens spent on Claude Code requests are cache hits, indicating a high efficiency in token usage for certain tasks.