Xiaoliu BOT
中文EN

X Platform June 27 AI Brief | GPT-5.6 Released in Three Tiers but Access Restricted; Anthropic Partially Unblocks Models; DeepSeek’s Accelerator Framework and Funding Draw Attention

GPT-5.6 Split into Sol, Terra, and Luna Tiers, but Initial Access Restricted

Multiple bloggers noted that OpenAI released the GPT-5.6 series, dividing the model into three capability and cost tiers: Sol, Terra, and Luna. Public information positions Sol as the next-generation frontier model, Terra for everyday efficient tasks, and Luna for high-throughput, low-cost scenarios. OpenAI also emphasized Sol’s investment in long-horizon cybersecurity tasks and red team testing. The discussion focus is not just on scores but on the release approach: Sam Altman stated that while originally planned for open access, it was changed to a limited preview at the US government’s request, promising to push for a more transparent and reliable early access process. Chinese bloggers focused on two points: the significance of Sol/ultra mode for agent workflows, terminal programming, and complex tool calls; and the fact that ordinary users and developers still cannot use it directly in the short term, making this release more a combination of capability signals and policy signals.

Sources:

Anthropic Mythos 5 Partially Restored, Fable 5 Still Awaiting Return

Multiple bloggers noted that Anthropic’s Mythos 5 has been partially unblocked after two weeks of restriction. Anthropic officially stated that Mythos 5 can be redeployed to a group of organizations operating and protecting critical infrastructure in the US, with continued efforts to expand Mythos 5 access and restore Fable 5. In Chinese discussions, MaxForAI and Baoyu contextualize this within broader model release review: on one hand, access restrictions for Mythos 5 and Fable 5 are easing; on the other, US government intervention in frontier model release methods is becoming an industry variable. MaxForAI also cited Axios reporting that Fable 5 could be restored as early as next week, but ultimately depends on government approval.

Sources:

DeepSeek’s Funding and DSpark Acceleration Framework Enter Discussion Simultaneously

Some bloggers mentioned that DeepSeek’s funding round is related to the Anthropic Mythos preview: according to The Information, Liang Wenfeng realized that without larger capital reserves, competing at the same level would be difficult. This signal shifts DeepSeek from a “technology-driven independent lab” toward a narrative of heavier capital and larger team expansion. On the same day, Gorden Sun summarized DSpark: DeepSeek’s semi-autoregressive speculative decoding framework, aiming to balance parallel and autoregressive drafters, reducing verification waste under high concurrency through confidence scheduling. According to his report, the solution has been deployed to DeepSeek-V4 production, replacing the original MTP-1, with generation speed improvements of 60% to 85%.

Sources:

OSWorld 2.0 and Hermes Agent Both Point to “Long-Horizon Agents Not Yet Solved”

Some bloggers noted the release of OSWorld 2.0, shifting focus from short tasks to long-horizon computer use closer to real work: 108 real workflows, approximately 318 tool calls per task, including stateful user configurations, dynamic environments, streaming interactions, and cross-source reasoning. MaxForAI reported results showing the current best performance still far below the level of “solving real computer tasks,” with Claude Opus 4.8 achieving at most about 20.6%. Nous Research also released progress on Hermes Agent, claiming it is exposed as a virtual model via MoA presets, outperforming Opus 4.8 and GPT-5.5 on their upcoming benchmark. Together, these two threads indicate that restricted access to frontier models and insufficient real agent capabilities are jointly driving demand for “multi-model orchestration, long-horizon tool calls, and verification capabilities.”

Sources:

MiniMax Suffers Massive DDoS, Independent Developers Also Building Automated Protection

Some bloggers mentioned that MiniMax services experienced a major outage, with the official statement pointing to a DDoS attack, and the website temporarily inaccessible. MaxForAI compared this to the earlier attack on DeepSeek, noting that MiniMax M3’s price cut this month made it more cost-effective, amplifying attention on service stability and competitive environment. On the other side, levelsio shared his own Pietflare: a small Cloudflare-like DDoS and probe detection system that aggregates suspicious requests from server access logs, maintains a central IP/ASN/country blocklist, and has servers periodically pull blocklist configurations to Nginx. One is a model provider under attack, the other an independent developer building their own protection, together illustrating that both AI products and personal projects increasingly need automated security operations.

Sources:

Codex Usage Reset and Agent Mail Expand Developer Toolchains

Multiple bloggers mentioned abnormal Codex usage and resets. Tibo reported that the Codex team is investigating rapid consumption on some accounts, possibly due to false positives in abuse/fraud protection mechanisms, and later stated that all Codex users would receive a usage reset. Xiaohu also summarized this “full reset” in Chinese. On the same day, Wesley’s daily report mentioned Tencent launching Agent Mail, positioned as a dedicated email for agents, and stated it has been adapted for tools like Claude Code, Doubao, Kimi, Cursor, and Codex. Combined with experience sharing on flomo Agent, Claude Code and Codex cross-verification, the developer toolchain theme has expanded from “single coding assistant” to account quotas, agent communication, memory containers, and multi-tool collaboration.

Sources:

Seedance 2.0 4K and GPT-Image-2 Push AI Content Production Toward Practical Use

Multiple bloggers shared more specific practical content around AI video and image workflows. Xiaoyu Chengzi shared a process for creating ad films with AI, emphasizing decomposition from product, character, shot design to final output logic, and noted that while Seedance 2.0 4K generation is slow, costs and iteration pace can be controlled through advance planning. Guizang used Seedance 2.0 to recreate the Codepilot promotional video, highlighting native 4K’s improvement in text clarity and material texture. Derek Wen shared prompt structures for GPT-Image-2 for online shopping try-on effect images: using selfies to lock in the person, product images to lock in clothing, specifying front, side, half-body close-up views, with the key being to establish a realistic relationship between person and clothing, not simple face-swapping or texture mapping.

Sources:

GEO, Memory Plugins, and flomo Agent Indicate Knowledge Work Moving Toward Engineering

Some bloggers broke down knowledge workflows in more detail. Xiangyang Qiaomu shared materials from the second GEO public class, including content engineering operation manuals, evaluation standards, system research reports, single-content practical tutorials, GEO rewriting prompts, GEO skill, and GEOFlow resources, indicating that generative search optimization has moved from concept to methods and tools. He also mentioned installing nowledge mem and configuring MCP, considering AI conversation memory and personal knowledge bases crucial. AI product Huang Shu, based on flomo Agent usage experience, emphasized the 0.5-second friction before recording, trusted memory containers, context completion, and the potential for flomo Agent and Codex/Hermes to form a “memory-action” dual system. Together, these threads show knowledge management shifting from storing content to engineering systems that are callable, evaluable, and accessible to agents.

Sources:

Statistics:

  • Timeline items scanned: 360
  • Bloggers hit: 35
  • Total tweets hit: 180
  • Weighted tweet score: 141.3
  • Original tweets: 85
  • Retweets: 39
  • Fetch attempts: 2
  • Boundary coverage status: tail_confidently_crossed_target_boundary