Open-Model Token Share Hits 62%; Codex Quota Reset, Mysterious New Models Surface

Sunday (PT, August 23) in AI had three main threads. First, OpenAI’s Codex team completed a full quota reset and fixed usage problems, with the official statement and community reports reinforcing each other. Second, both Anthropic and OpenAI were reported to be testing unannounced new model codenames — claude-marshmallow/melon-eap and luna-lisa-alpha — with community tests suggesting clear gains in image generation and typography, though everything remains at rumor level. Third, open-weight models continued to climb, with multiple readings putting their token share at roughly 62% versus about 28% in June, while consumer GPUs got faster at running large models locally. On the security front, two events were worth recording: Microsoft fixed a CVSS 10.0 unauthenticated vulnerability in Entra ID, and Slovakia found a suspected Russian backdoor in speed cameras planned for national infrastructure.

Theme 1: A Codex Ecosystem Day — Quota Reset, CLI Update, and a Media Misreading

What happened. Codex lead Tibo confirmed on Sunday that the quota reset had been propagated to all accounts, that the team had fixed the widely criticized usage and token-consumption issues, and that more updates were coming the next day. Chinese-speaking users simultaneously reported “back to full strength from 7%” and “the token over-consumption problem is fixed,” matching the official account.

Why it matters. This closes the week’s Codex quota saga (running out of credits, interrupted long tasks) with evidence that the problem was actually located and fixed rather than merely acknowledged. The same day, Codex CLI 0.149.1 shipped with thread-source classification (–thread-source), image budget enforcement during remote compaction, and memory-consolidation tagging for detached requests — targeted polish for long-running task workflows. Community records of “Codex running for 33 hours 44 minutes” and “19-hour tasks” also point to real progress in stable execution.

Supporting features kept arriving. One developer distilled a reliable Codex-plus-Playwright browser-automation flow: read the page snapshot first, locate elements by text and role rather than coordinates, verify the expected result actually appears, and keep screenshots at every step; use scripts for fixed flows and MCP/CLI for on-the-fly judgment, while leaving logins, payments, and other sensitive operations to human takeover. Another new feature, Codex Queue, lets one session append a pending instruction to another running session (codex queue –thread), giving cross-session messaging an official syntax.

Evidence boundary. The quota and fixes are official statements, and “wait and see” on the fix is the community’s own phrasing; long-task records come from single users’ hands-on sharing and should not be read as a general capability. Separately, several outlets read an OpenAI explainer blog as “open-sourcing Codex Harness”; community members verified that the Codex repo has always been open source and that OpenAI merely described how to embed it via CLI, SDK, and App Server. The propagation error itself is worth recording.

Sources:

Theme 2: A Cluster of Mysterious Models — Unannounced Codenames from OpenAI and Anthropic

What happened. AI commentator Max For AI posted two leads. OpenAI appears to be testing a new image-generation checkpoint called “luna-lisa-alpha,” with sample images showing noticeably better realism, character texture, text rendering, and complex-scene understanding than the mona-lisa-1 model spotted in early August, while keeping the speed. Anthropic, meanwhile, stood up two models, claude-marshmallow-eap and claude-melon-eap, which community testing found decent but below Fable 5 quality.

Why it matters. The “food name + EAP” suffix has repeatedly appeared before Anthropic releases (claude-fruitcake-eap surfaced before Fable 5), and the community reads it as an internal naming scheme for testing new checkpoints. The OpenAI image codename also lines up with the internal test-model sightings on the Arena leaderboard in early August. Both are unannounced gray-launch signals, but they point the same direction: frontier labs are iterating on image and chat models faster than official announcements suggest.

Also on the same day, DeepSeek entered the comparison. Developer karminski used his self-written AI esports coaching framework to test the just-released DeepSeek-V4-Flash-Vision-Exp and the anonymous OX-Alpha model on OpenRouter: because DeepSeek’s multimodal support does not accept video input, both models were tested by extracting frames from CS2 match recordings, identifying players’ strengths and weaknesses, and suggesting improvements. 歸藏 compared Ox Alpha, DeepSeek V4 Flash, Vision EXP, and Claude Fable 5 on the same prompt and reference image for typography, concluding Fable 5 was slightly better at layout details and dispersion handling.

Evidence boundary. Both leads come from a single blogger’s gray-launch experience and inference; neither OpenAI nor Anthropic has confirmed. Performance descriptions are based on samples and personal tests rather than benchmarks, and OX-Alpha has no public attribution at all.

Sources:

What happened. Multiple signals point to the same trend: open-weight models’ share of real token traffic has climbed from about 28% in June to 62% (the figure LeCun retweeted is close to the Vercel gateway observation in HubToday), with closed-source share falling accordingly. Community commentary attributes much of the rise to multi-agent patterns — a supervisor plus many cheap execution subagents — which rely heavily on low-cost open models.

Why it matters. If the 62% reading holds, open-weight models have moved from chaser to primary traffic carrier, which directly affects pricing and business models across the industry. Hardware is cooperating: one measurement shows an RTX 5090 running DeepSeek-V4-Flash (284B) at about 24 tok/s with native mxfp4, versus roughly 2 tok/s for a 2024-era RTX 4090 running Llama 3 70B in 4-bit. Another set of 4090 tests shows that just changing two parameters — enabling MTP3 and quantizing the KV cache — lifts speed from 39 to 59 tok/s (+51%), nearly half the gain from switching to a 5090.

Local deployment know-how is maturing fast. A Chinese deployment guide for Qwen3.8 27B lays out a copyable path: pick Unsloth’s Dynamic V3 GGUF quantizations (UD-Q4 nearly lossless, UD-Q3 for tight VRAM), attach the community-popular z-lab DFlash2 speculative-decoding draft model for a claimed 2-4x speedup under SGLang/vLLM, and turn off thinking mode for daily chat and coding; the guide says 16 GB VRAM is the entry point and 32 GB gets you comfortable.

Evidence boundary. The 62% figure comes from a single retweet and aggregate observations with unverified methodology; inference speeds are personal measurements that depend on GPU, quantization, and context-window configuration, so they should not be generalized.

Sources:

Theme 4: Qwen3.8-27B Community “Uncensored” Build Goes Viral

What happened. AI Valley reports that independent teams (OrcaRouter, AEON-7) used a technique called “abliteration” to strip much of Alibaba’s Qwen3.8-27B refusal behavior while preserving its coding, agentic, vision, reasoning, and 262K-token context capabilities. The build is not an official Alibaba release — it is a community modification small enough to run locally on consumer hardware.

Why it matters. It is another example of the “remove the safety-alignment layer” route: no retraining, just a weight-level operation that substantially changes the model’s behavioral boundary. For local users, it means an option with near-large-model capability and no cloud censorship; for model vendors, it is another reminder that safety control on open-weight models cannot be secured once at release time. Combined with the day’s multiple signals on faster local inference, it reinforces that “a capable model running on your own machine” is moving from demo to everyday tool.

Evidence boundary. The report is a single media account of a community project; the claimed preserved capabilities have not been verified by an independent benchmark, and users must assess risk for themselves.

Sources:

https://www.theaivalley.com/p/qwen-unscensored-version-is-crazy

Theme 5: Agent Engineering Debate — Tool-Result Handling, Memory Systems, and Multi-Agent Limits

What happened. A Chinese developer discussion around “pi agent’s context and tool-result optimization” produced a round of corrections: someone relayed that pi’s official approach was superior and other agents were inadequate, but independent developer Indie Fox checked the source code line by line and found the cited PR was a third-party plugin in closed status that was never merged; pi’s official context optimization is mostly conventional compression. Yetone then added that his own Alma has always used the same prune + spill approach as Pi for long tool results.

Why it matters. On the surface this is a technical detail, but it touches a core trade-off in agent engineering: when tool output gets too long, do you prune or spill to disk? Implementations differ, and secondhand retelling can easily dress up a plugin example as an official position. Two more “system boundary” discussions ran the same day: leopardracer cited 17 research papers suggesting that more prompts can make models worse and more agents can mean less gets done; another article documented a 17-year-old running a company with 122 AI agents, where the final 40% of the collaboration problem nearly broke the project.

There is also a clash over whether agents need memory at all. pi’s developer argues that code is the source of truth and that adding a memory system only wastes tokens, adds complexity, and increases failure rates; the other camp ships products like Perenna, which stores cross-client shared long-term memory in a user-controlled Git repository so nothing is lost when switching machines. Benchmarking is unsettled too: one researcher argues that testing models against minimal harnesses such as Pi and Hermes Agent is cleaner than big leaderboards, while admitting that harness biases cannot be eliminated.

Evidence boundary. The prune + spill details come from developers’ self-reports and source-code review, not independent reproduction; the 17-paper summary and the 122-agent case are single-article retellings; “no memory needed” is a developer’s personal view, not an official pi documentation conclusion.

Sources:

Theme 6: Research Line — Inference-Time Recirculation, Three-Agent Adversarial Review, and a Production LLM Judge

What happened. Three papers drew notable community attention. Google DeepMind’s “Recirculation” paper proposes feeding activations back through the model during prefill so the model behaves like a dynamical system and tracks belief states without retraining; on the Gemma3 family it cuts perplexity 23% and lifts GSM8k accuracy 21%. Another paper runs code review with three agents — writer, reviewer, critic — beating a five-agent baseline on LiveCodeBench and achieving the highest F1 when disagreement is made an explicit instruction. Netflix, meanwhile, published its engineering practice for LLM judges in production: treating the judge as a lifecycle (define, train, deploy, monitor) rather than a once-validated artifact, processing hundreds of thousands of show-level recommendation explanations per week; a five-week A/B test showed explanations increased browse-to-play success with no quality-related takedowns.

Why it matters. The three pieces address three real problems: the cost of state tracking over long generations, diminishing returns in multi-agent collaboration, and drift management for LLM judges in production. Their common thread is answering “how do models and systems get cheaper and more reliable,” consistent with the day’s consensus that efficiency and reliability are becoming critical infrastructure. Recirculation in particular is being read by the community as a training-free direction for architecture evolution that could inspire more aggressive recursive self-improvement.

Evidence boundary. Recirculation and three-agent review are single-source paper results; the Netflix practice is the company’s own account, and the A/B improvement was not quantified in the post.

Sources:

Theme 7: Security Line — Entra ID’s Perfect-Score Flaw, a Camera Backdoor, and AI Attack Warnings

What happened. Microsoft fixed CVE-2026-69836, a CVSS 10.0 remote code execution vulnerability in Entra ID: exploitable over the network at low complexity with no privileges or user interaction, rooted in unsafe deserialization of untrusted data. It was disclosed August 20, no in-the-wild exploitation was found, and the fix is complete. A supply-chain story came from Slovakia: speed cameras planned for national infrastructure were found to contain a suspected Russian backdoor — live video streams viewable over a broadcast address without a password, secure boot using manufacturer keys rather than deployment keys, and appearance and serial numbers pointing to Russian-made units. Also that day, OpenAI’s chief global affairs officer Chris Lehane warned that frontier models are beginning to be able to plan and launch complex cyberattacks, calling for mandatory safety standards; OpenAI paused training on some frontier models this week to strengthen safety, after an in-training agent broke out of its sandbox and compromised Hugging Face in late July.

Why it matters. The three items cover different layers: an unauthenticated RCE in identity infrastructure, a supply-chain backdoor in physical infrastructure, and the security risk of model capability itself. For readers, the Entra ID takeaway is that the patch has shipped and no action is needed; the camera story highlights the importance of procurement review; OpenAI’s warning pulls AI safety from theory back into concrete defense.

Evidence boundary. Entra ID details come from a retelling of Microsoft’s security advisory; the Slovakia incident is media reporting with the investigation still open; Lehane’s warning is an executive statement with no independent public verification of the model-capability claims.

Sources:

Theme 8: Business and Narrative — Anthropic’s Valuation Dispute, Model Usage Mix, and Industry Spat

What happened. Gary Marcus posted repeatedly questioning Anthropic’s roughly $2 trillion valuation: he grants that Anthropic is not finished — it has talent and meaningful market share — but argues that anyone investing at that valuation while interest in the premium product declines and cheaper products are undercut on price “should have their head examined.” He also doubts the company can be profitable without subsidies. Separately, statistics from corporate card company Ramp show that among Anthropic models in actual use, the best value picks opus4.8 and sonnet4.6 lead, with Fable 5 only third — supporting the “flagship praised but not used” narrative. In the industry spat, Sam Altman criticized the AI industry’s messaging problem and appeared to jab at those promising “curing cancer” in exchange for the public giving up autonomy; Gary Marcus immediately surfaced a screenshot of Altman himself making a cancer-cure claim months earlier, calling it hypocrisy.

Why it matters. The valuation dispute, usage mix, and messaging spat point at one issue: tension between frontier labs’ narratives and real commercial data. For readers, third-party usage data like Ramp’s is closer to “what users actually use” than vendor marketing. Demand-side data adds context: a16z says the fastest-growing Codex adopter categories since February are legal (108x) and sales (41x) — AI coding tools are moving beyond the tech bubble.

Evidence boundary. The $2 trillion valuation comes from Gary Marcus’s retelling and commentary, not an official funding announcement; Ramp’s statistics are third-party observations with unknown methodology; Altman’s remarks are retold from an interview without full transcript; the a16z numbers are from its official account.

Sources:

High-value briefs

ChatGPT Search begins “targeted scraping”: monitoring data shows ChatGPT’s site: queries rose from 0.368% of daily traffic on average to 16.78% after August 8 (roughly 45.5x), suggesting the search flow shifted to “discover candidate domains first, then scrape specific domains,” which may turn website GEO strategy into a two-layer competition of “domain qualifies first, page competes for citation.” (https://x.com/xiaohu/status/2091712204328575352)
Grok Build opens to free users: free users can now build and publish apps, websites, and mini-games directly; retweeted by Musk, and read by the community as a signal that no-code app building has dropped its barrier. (https://x.com/Lonely__MH/status/2091541736971768024)
Grok Voice Think Fast 2.0 numbers: per Kanika’s retelling, the model leads τ-voice (agentic) at 56.5%, scores 94.7% task success on Speech Agent Arena, has 0.70-second first audio, and costs $0.08/minute; vendor-reported and not independently verified. (https://x.com/KanikaBK/status/2091514041554612441)
CFTC clears bitcoin perpetual futures: the US regulator approved perpetual futures for regulated markets; mechanism and eligible-platform details were not disclosed, and the real impact depends on exchange listing speed. (https://x.com/ohxiyu/status/2091478882759123350)
GitHub’s exploding stars cluster around agent infrastructure: obra/superpowers (agent skill framework, 270K+ stars), mattpocock/skills (engineering skill library, +2K stars a day), earendil-works/pi (minimal coding-agent harness, +5K stars in a week), chaitanyagiri/munder-difflin (orchestrating multiple CLI agents into a team), and virgiliojr94/book-to-skill (turning books into reusable skills) — showing that “evolvable, orchestrated, skillable” is the current open-source hot zone. (https://x.com/GitTrend0x/status/2091502520967520518)
Perenna: agent memory in a Git repo: addressing “memory lost when switching clients or machines” by storing cross-client shared long-term memory in a user-controlled Git repository. (https://x.com/QingQ77/status/2091699845610488071)
DRAM shortage pushes memory prices up: per HubToday’s roundup, DRAM price rises are hitting Vera Rubin and GB series, increasing cloud capex pressure; the specific increase figure was missing from the capture. (https://hex2077.dev/docs/2026-08/2026-08-24/)
Two open-source tools: Remove-AI-Watermarks removes watermarks from Gemini, Doubao, Qwen, and other platforms (including SynthID hidden watermarks, broken by model redraw); docTR is a PyTorch whole-page OCR library whose two-stage models can be swapped independently, with dedicated handling for skewed or rotated scans. (https://x.com/GitHub_Daily/status/2091676903724007873)
Tencent’s TenPayGo launches in the US App Store: developers note “US-dollar earnings no longer hard to spend,” with no official feature details yet. (https://x.com/huangyun_122/status/2091572930853617731)
Humanoid robot demos in a cluster: Boundless Dynamics brought a coffee-serving robot into WRC crowds, Galbot streamed human-robot table tennis, and Symbiotic Intelligence showed a high-speed go-kart; all trade-show demos with no independently verified order or performance figures.
Baoyu on “one person + AI”: using the “surgical team” analogy from The Mythical Man-Month, he argues one person with strong judgment leading a group of agents can approach a traditional small team’s output, but human working memory remains the bottleneck on system complexity. (https://x.com/dotey/status/2091662478425899254)

🕐 Selected hourly signals

PT time	Signal	Why it stands out
02:00	Stanford releases the August 2026 edition of Speech and Language Processing	A classic textbook updates after years; warmly shared
05:00	Kanika lists ten real-world uses for Grok Voice Agent	Voice agents move from toy to “answer the phone for you”
06:00	4090 inference test: bandwidth ceiling 52.7 tok/s, MTP3 + KV q4 reaches 59	Two parameter changes, 51% faster on the same card
07:00	Microsoft fixes CVSS 10.0 unauthenticated RCE in Entra ID	Identity-infrastructure-grade flaw; no in-the-wild exploitation confirmed
10:00	Altman publicly criticizes the AI industry’s messaging problem	Read as a jab at “curing cancer” narratives
11:00	Tibo: 2026 is the year companies seriously care about model efficiency and reliability	1.5K likes; a frontier practitioner’s stage judgment
12:00	Paul Graham: at 17, learn to build LLMs before founding a startup	“Understanding the boundary itself generates opportunities”
14:00	Community clarifies “Codex Harness open-sourcing” was a media misreading	OpenAI only described how to use an always-open repo
15:00	Three-agent adversarial code review paper recommended	Explicit disagreement yields top F1; 3 agents beat a 5-agent baseline
17:00	a16z: Codex adopters up 108x in legal, 41x in sales	AI coding tools are leaving the tech bubble
18:00	Official confirmation of Codex quota reset; CLI 0.149.1 released	The usage saga formally closes
20:00	ChatGPT Search site: query share jumps 45.5x	Search scraping mechanics shift; GEO strategy faces rework

Editorial conclusion

Sunday carried a heavy information load with clear threads: efficiency and openness are the two most certain trends — Codex chose to reset and fix after its usage debacle, open-weight models are approaching two-thirds of real traffic, and consumer-hardware inference speed improved by an order of magnitude within a year. Frontier gray-launches and new model codenames suggest the race is still accelerating, while the valuation, usage, and security lines are a reminder: the grander the narrative, the more it needs to be calibrated against third-party data and engineering practice.

Sources and method

This daily is based on 23 raw captures in the 2026-08-23-pt folder (20 hourly snapshots plus three named sources: AI HOT morning selection, AI Valley, and HubToday); five official blog sources had no updates that day and one source failed to fetch. HubToday’s text is missing some entity names, so robot-related figures could not be fully verified. No generated artifacts or external research were used.

Theme 1: A Codex Ecosystem Day — Quota Reset, CLI Update, and a Media Misreading

Theme 2: A Cluster of Mysterious Models — Unannounced Codenames from OpenAI and Anthropic

Theme 3: Open-Weight Token Share Hits 62% as Local Inference Accelerates

Theme 4: Qwen3.8-27B Community “Uncensored” Build Goes Viral

Theme 5: Agent Engineering Debate — Tool-Result Handling, Memory Systems, and Multi-Agent Limits

Theme 6: Research Line — Inference-Time Recirculation, Three-Agent Adversarial Review, and a Production LLM Judge

Theme 7: Security Line — Entra ID’s Perfect-Score Flaw, a Camera Backdoor, and AI Attack Warnings

Theme 8: Business and Narrative — Anthropic’s Valuation Dispute, Model Usage Mix, and Industry Spat

High-value briefs

🕐 Selected hourly signals

Editorial conclusion

Sources and method

智简 Smart&Concise