AI Daily: OpenAI Reopens $200 Pro Plan, Claude Sonnet 5.5 Goes Live

September 29 saw a flurry of major developments across the AI landscape — from OpenAI reopening its premium subscription tier to Anthropic updating its flagship model and Manus launching a new-generation Agent framework. Together, these moves paint a clear picture of the intensifying competition in the large language model and intelligent agent space. Here's a breakdown of the key updates.

OpenAI on the Move: Pro Subscription Returns, DevDay Preview

Members of the OpenAI Codex team revealed that the $200/month Pro subscription will be reopened — this time exclusively for new subscribers, with a revised usage calculation method. Limiting access to new users while redesigning the billing logic signals OpenAI's careful approach to allocating high-end compute resources. Pro-tier usage typically carries higher inference costs, so reopening the subscription while tightening usage caps is a deliberate move to balance supply constraints with commercial returns.

Separately, OpenAI's developer account previewed a DevDay keynote scheduled for the following day and opened a livestream link. Tibo from the Codex team tweeted, "Yesterday is in the past — today is DevDay," adding that it's "all good news," building anticipation for the developer conference.

Another noteworthy report came from The Wall Street Journal, which claimed OpenAI is withholding a new model codenamed Astra (rumored to be GPT-6.1) due to safety concerns. The report has not been officially confirmed, but if accurate, it reflects a growing trend: the release cadence of frontier models is increasingly constrained by safety evaluations.

Anthropic Updates Its Flagship: Claude Sonnet 5.5 Now Available

Anthropic announced that Claude Sonnet 5.5 is now officially live and available for use. The Sonnet series has long been Anthropic's balance point between performance and cost, targeting high-frequency everyday tasks and coding workflows. This version update comes at a moment of direct competition between OpenAI and Anthropic on coding assistants and Agent capabilities, with model update frequency clearly accelerating.

From an industry perspective, Sonnet 5.5 and OpenAI's wave of announcements landed in nearly the same window. The near-simultaneous releases from these two leading labs reflect both natural technical iteration cycles and unmistakable market positioning moves.

Manus 2.0: Cascade Framework Targets Cost Efficiency

Manus has released version 2.0 across both its web and desktop clients, with the core highlight being its new foundation — Cascade, an iteration of its token-level Agent framework. The official numbers are compelling: token consumption reduced by 23.2%, task completion time cut by 28.2%, and costs lowered by 32%.

These figures address a core pain point of current Agent products. When intelligent agents execute multi-step tasks, token consumption and latency tend to multiply rapidly, making cost control a critical barrier to real-world deployment. Cascade optimizing across all three dimensions simultaneously shows that Manus's product strategy has shifted from "can it do the task?" to "can it do it efficiently enough?"

Manus 2.0 also introduces cloud computer support, event-triggered automation, and Q-application capabilities. The addition of cloud computers and event-triggered automation signals a shift from reactive to proactive Agent execution — aligned with the industry-wide pursuit of truly autonomous agents.

What is a token, and why does it matter for Agents? A token is the basic unit of text that large language models process — roughly equivalent to half to one English word, or half a Chinese character. When an Agent executes multi-step tasks, each round of reasoning, tool calls, and result parsing consumes tokens, and context must be carried between steps, causing token consumption to grow exponentially with task complexity. A three-step Agent task involving web browsing, API calls, and report summarization can easily consume 5–10x the tokens of a single conversation turn. This is precisely what Cascade targets — by compressing context representations and eliminating redundant intermediate steps, it lowers the compute cost per run without sacrificing task quality. For enterprises looking to commercialize Agent capabilities, per-task cost directly impacts ROI, which makes cost reduction data far more persuasive than benchmark scores.

Open Source & Ecosystem: InternDecision Models, DeepSeek Plugin Directories

ModelScope has launched the InternDecision (书生决策) model series, covering three parameter scales — 0.8B, 2B, and 4B — targeting multimodal and structured decision-making tasks. The multi-tier parameter design accommodates different compute environments, with the smaller variants particularly suited for edge and lightweight deployment scenarios.

On the ecosystem tools front, the DeepSeek^Ad community highlighted three plugin directory sites focused on plugin discovery, marketplace aggregation, and curated resources. As the large model application ecosystem expands, these navigation and aggregation tools are becoming essential entry points for developers to quickly find usable plugins.

Understanding model parameter scales (0.8B, 2B, 4B) The "B" in model size notation stands for billion parameters — a core metric for a model's capacity and compute requirements. Larger parameter counts generally translate to stronger reasoning and generalization abilities, but also proportionally higher GPU memory and inference compute demands. A 0.8B model can run on consumer-grade GPUs or even some mobile devices, making it suitable for latency-sensitive or offline edge deployments. The 2B–4B range strikes a balance between capability and efficiency, fitting well on mid-sized servers or low-cost cloud instances. By offering three tiers, InternDecision allows developers to choose based on their actual compute budget and task complexity — avoiding the resource waste of a one-size-fits-all large model approach. This flexibility is a mainstream strategy in the open-source community's push toward model lightweighting.

Takeaway: Competition Accelerates, Cost Efficiency Emerges as the New Battleground

Looking across the day's announcements, several clear industry trends emerge. OpenAI and Anthropic are in direct head-to-head competition at the model and subscription level, with release cadences visibly accelerating. Manus is entering the Agent race with concrete cost-reduction data, putting operational efficiency front and center. On the open-source side, multi-scale decision models are expanding the options for practical deployment.

Beyond raw model capabilities, cost, efficiency, and safety are becoming the new frontiers of competition. Whether it's OpenAI tightening Pro usage limits and delaying a launch over safety concerns, or Manus emphasizing double-digit reductions in both token consumption and costs — it all points to the same conclusion: the large model industry is moving from "competing on parameters" into a new phase of "competing on deployment economics."